Across the four cohorts and two model initializations, AtrousLoRA produced Dice scores between 1.14 and 4.50 percentage points higher than standard LoRA. These findings apply to the evaluated datasets, prompt conditions and initialization settings rather than to unrestricted automatic segmentation. Accurate aortic segmentation involves delineating vessels with different calibers and irregular boundaries.

The researchers noted that full encoder fine-tuning of SAM updates many parameters, while conventional LoRA does not aggregate image context across different receptive fields within its task-specific update. VesselSAM was developed to accommodate both considerations by modifying the query and value projections in each image-encoder block of the base ViT-B vision transformer. At the technical level, the researchers inserted a multi-branch Atrous Attention Module into each rank-4 residual path associated with the selected query and value projections.

This architecture allowed the low-rank adaptation to process image context at multiple receptive-field scales while retaining a relatively small trainable parameter set. VesselSAM was initialized from MedSAM for the binary segmentation task, and the comparative evaluation considered both SAM and MedSAM initialization. The Aortic Vessel Tree cohorts were Dongyang, Rider and KiTS.

Tight bounding boxes derived from the ground-truth annotations were supplied at validation and test time, making the reported scores specific to a prompted segmentation setting rather than a fully automatic detection task. VesselSAM attained a slice-averaged Dice similarity coefficient of 93.50% on the Dongyang cohort, 93.25% on Rider and 93.02% on KiTS. The corresponding result for the Type-B Aortic Dissection dataset was 93.26%.

The consistent performance across these four evaluated cohorts shows that the reported numerical results were not confined to a single named dataset, although the evidence does not establish how the method would perform in other clinical settings or without ground-truth-derived prompts. The model updated 6.8 million parameters, equivalent to 7.2% of its total parameters, underscoring the parameter-efficient nature of the adaptation. The study’s central finding is therefore twofold: AtrousLoRA outperformed standard LoRA under the specified evaluation, and VesselSAM achieved high overlap scores for binary aortic vessel segmentation without full encoder fine-tuning.

Those results support the approach as a prompt-dependent adaptation of pretrained segmentation models, not as evidence of clinical benefit or clinical deployment. Overall, the research presents AtrousLoRA as a way to incorporate multiscale spatial processing into low-rank adaptation for aortic vessel segmentation. The reported gains over standard LoRA were measured across four cohorts and two initializations, with the strongest absolute DSC reaching 93.50%.

Important limits remain: the system relied on tight ground-truth-derived box prompts, and the supplied evidence does not demonstrate performance for prompt-free segmentation, unseen populations, or routine clinical use.

Subject of Research: Prompted parameter-efficient adaptation of SAM for aortic vessel segmentation in CT angiography

Article Title: VesselSAM: a parameter-efficient adaptation of SAM to aortic vessel segmentation with AtrousLoRA

Article References: Zhao, X., Zhuang, P., Zhang, Z., Iltaf, A., Ahmed, R. M., & Li, B. (2026). VesselSAM: a parameter-efficient adaptation of SAM to aortic vessel segmentation with AtrousLoRA. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02911-1

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02911-1

Keywords: VesselSAM, AtrousLoRA, aortic segmentation, CT angiography, Segment Anything Model, MedSAM, low-rank adaptation, Dice coefficient, vision transformer, medical imaging, parameter-efficient, adaptation

The new framework, called CA-CLIP, builds on vision-language models that align images and text in a shared embedding space. Category names, attribute descriptions, and symptom descriptions can all participate in recognition, allowing textual semantics to guide visual classification. Earlier plant disease studies using such models focused mainly on general image recognition, few-shot learning, or single-disease scenarios, and global image-text alignment was often sufficient for those tasks. But when several diseases coexist on one leaf, global alignment alone cannot capture the local, class-specific evidence needed to separate overlapping symptoms. CA-CLIP addresses this gap with a class-aware local alignment mechanism that matches each disease category to its own local visual evidence rather than relying on a single global image-text score.

Constructing reliable training material for coexisting diseases posed its own challenge. Annotating real images of multiple diseases on the same leaf is costly, and the resulting datasets are often small and imbalanced. The team therefore built a benchmark called CC-Leaf, a resource for coexisting leaf diseases, by combining real disease images with controlled image synthesis. The pipeline used the Segment Anything Model, a general-purpose segmentation tool, to extract leaf and lesion regions with minimal manual effort. Because lesion boundaries can be blurry and colors can shift between healthy and diseased tissue, the researchers applied manual corrections and color harmonization to keep the synthesized images visually plausible. The result is a dataset covering five crops and eleven disease categories, with images containing one, two, or three coexisting diseases per leaf.

The synthesis workflow proceeded in several stages. First, real leaf and lesion regions were segmented and cleaned. Next, lesion patches were grafted onto healthy leaf backgrounds at controlled locations and densities, with blending and color adjustments to avoid abrupt edges. Finally, the team verified that the synthesized images preserved the characteristic appearance of each disease and that co-occurring combinations remained phenotypically reasonable. This controlled construction allowed the researchers to vary disease density, lesion size, and co-occurrence patterns systematically, producing a dataset that is larger and more balanced than what real-world annotation alone could provide.

On the recognition side, CA-CLIP combines a vision encoder with a text encoder that processes category names and symptom descriptions. The key innovation is the class-aware local alignment module: for each disease category, the model identifies local image regions whose features align with that category’s textual semantics, producing a per-class matching score. These scores are then aggregated to decide which diseases are present. This design lets the model separate evidence for different categories even when their symptoms overlap spatially. The researchers also introduced a co-occurrence-aware training objective that encourages the model to recognize each disease independently rather than treating co-occurring diseases as a single combined pattern.

Evaluations on the CC-Leaf benchmark showed that CA-CLIP outperformed representative baselines across multiple metrics. In multi-label recognition scenarios with two or three coexisting diseases, the framework achieved higher accuracy and better per-class recall than models relying on global image-text alignment alone. The class-aware local alignment mechanism proved especially valuable when symptoms of different categories overlapped, a situation where global approaches tended to confuse categories or miss the less prominent disease. The co-occurrence-aware objective further improved recognition of each disease independently, reducing the tendency to treat co-occurring diseases as one combined pattern.

The study also examined how the framework behaves under different data conditions. When trained with limited annotated data, CA-CLIP retained strong performance, suggesting that the textual semantics and local alignment mechanism help compensate for scarce supervision. The researchers further tested robustness to variations in lesion density and co-occurrence patterns, finding that performance degraded gracefully as conditions became harder. These analyses indicate that the framework is not merely fitting the specific construction of the benchmark but is learning transferable diagnostic evidence.

Beyond raw accuracy, the framework offers practical advantages for agricultural deployment. Because textual descriptions participate in recognition, the model’s per-class matching scores provide a form of interpretability: the regions that drive each disease decision can be identified and inspected. This is valuable in the field, where agronomists need to understand why a model flags a particular pathogen. The framework also operates on ordinary leaf photographs, requiring no special imaging equipment, which lowers the barrier to adoption in real-world settings.

The researchers acknowledge limitations. The benchmark was constructed with controlled synthesis, and while the images are visually plausible, real-world conditions introduce additional variability such as leaf deformation, complex backgrounds, and uncontrolled lighting. The team notes that future work could extend the framework to field-collected images and to a broader range of crops and pathogens. They also point out that the current design focuses on leaf diseases and does not yet address other plant organs or disease types.

Nevertheless, the study represents a meaningful advance for automated plant pathology. By pairing a carefully constructed benchmark for coexisting leaf diseases with a vision-language framework built for symptom-mixed recognition, the researchers have addressed a scenario that earlier single-disease approaches could not handle reliably. The combination of local, class-aware alignment and co-occurrence-aware training offers a template for future diagnostic systems that must separate overlapping evidence rather than treat each image as a single pattern. As automated plant pathology moves from laboratory benchmarks to field deployment, frameworks like CA-CLIP may prove essential for the messy, multi-pathogen reality of agricultural practice.

Subject of Research: CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework

Article Title: CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework

Article References: Liu, S., Wu, J., Zhao, W., Liu, J., Chen, J., & Cui, Y. (2026). CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework. Smart Agricultural Technology, 15, Article 102617. https://doi.org/10.1016/j.atech.2026.102617

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102617

Keywords: plant pathology, vision-language models, coexisting diseases, multi-label recognition, Segment Anything Model, image synthesis, local alignment, agricultural AI, leaf diseases, benchmark construction, interpretability, deep learning

The problem the researchers set out to solve is one that has frustrated the medical AI community since SAM’s debut. Lesions—whether skin melanomas, gastrointestinal polyps, thyroid nodules, intracranial hemorrhages, or focal cortical dysplasia—are notoriously amorphous. They blur into surrounding tissue, vary wildly in size and shape between patients, and often lack the crisp edges that SAM’s natural-image training rewards. Earlier evaluations, including a large-scale benchmark of 1,050,000 two-dimensional medical images with more than 6 million masks, confirmed that SAM’s zero-shot transfer to medical tasks lags behind specialized state-of-the-art methods, and that its reliance on point, box, text, or mask prompts introduces instability and sensitivity that clinicians cannot tolerate in practice.

SCMamba-SAM attacks the problem on three fronts. First, it keeps SAM’s powerful pre-trained ViT-B image encoder—twelve layers of multi-head self-attention—but inserts Low-Rank Adaptation (LoRA) layers into each self-attention block. LoRA freezes the original Query, Key, and Value weight matrices and learns only small low-rank updates to the Query and Value projections through two trainable linear layers with a reduced projection dimension. This parameter-efficient strategy lets the encoder quietly re-tune itself to the statistics of medical imagery without the cost and overfitting risk of full fine-tuning, which would require updating all of SAM’s parameters on scarce labeled data.

Second, the team bolted on a novel module they call the Encoding Mamba Adapter (EMA). Mamba, a recent architecture built on structured state space models, can model long-range dependencies across a sequence with approximately linear complexity in sequence length—an advantage over the quadratic cost of Transformer self-attention when data are limited and overfitting looms. The EMA is a serial pair of Spatial-Mamba and channel-Mamba blocks: the first captures long-range interdependencies among spatial tokens, while the second, transposed so that channels become the sequence, builds global relationships among feature channels. Each block consists of a squeeze-MLP that compresses the input, a Mamba layer that propagates information through a selective state space, and an expand-MLP that restores the original dimensions. Notably, the authors replaced every linear layer in the original Mamba design with multi-layer perceptrons, a modification their ablation experiments show yields more complex feature extraction. Four EMAs are plugged into the third, sixth, ninth, and twelfth transformer blocks of the encoder, and their adapted outputs are fused with the original features by addition.

Third, SCMamba-SAM discards SAM’s prompt encoder altogether. Instead of asking a radiologist or dermatologist to click points or draw boxes, the framework employs an enhanced mask decoder with deep supervision. Adapted features are progressively up-sampled through four deconvolution layers, each combining up-sampling, ReLU activation, and convolution, to restore spatial resolution and fuse multi-scale information. During training, intermediate decoder outputs at each stage are routed through segmentation heads and supervised against ground-truth annotations using a weighted sum of Focal and Dice losses, so the model learns to attend to lesion regions at every scale rather than only at the final output. At inference time, the pipeline is fully automatic: an image goes in, a lesion mask comes out.

The empirical case for the design is substantial. Across six public lesion segmentation datasets spanning dermatoscopy (ISIC2017 and ISIC2018), endoscopy (Kvasir-SEG), ultrasound (TN-SCUI2020), computed tomography (CT-ICH2020), and magnetic resonance imaging (MR-FCD2023), SCMamba-SAM outperformed CNN-based, Transformer-based, GAN-based, and SAM-based competitors. On ISIC2017 it achieved an IoU of 0.777 and a Dice coefficient of 0.861, edging out every baseline including fine-tuned SAM variants and the adaptation framework MSA. On ISIC2018 it reached an IoU of 0.900 and a Dice of 0.945. On Kvasir-SEG it beat the second-ranked SAM-Med2D by 6 percent in IoU, 3 percent in Dice, and 7 percent in Recall, and it succeeded where fine-tuned SAM failed on small polyps. On thyroid ultrasound it surpassed MedSAM by 10 percent in IoU, and on the particularly difficult focal cortical dysplasia MRI dataset it achieved an IoU of 0.462, six points above the runner-up, correctly segmenting MRI-negative cases that stumped other models.

Ablation studies dissected exactly where the gains come from. Starting from pre-trained SAM alone—IoU 0.550, Dice 0.676—adding LoRA fine-tuning lifted performance to 0.765 IoU and 0.849 Dice, confirming that domain adaptation of the frozen encoder is the single largest contributor. Deep supervision nudged Recall upward from 0.816 to 0.824, and the full EMA module pushed all four metrics to their peaks. Position mattered too: adapters placed at the middle and late transformer layers contributed most, while a shallow-layer adapter actually hurt performance. Replacing the Mamba layers with conventional self-attention degraded Dice by nearly five points, evidence that the state-space formulation, not just extra capacity, drives the improvement. Hyperparameter sweeps identified a LoRA rank of 128 as the sweet spot—larger ranks bought only marginal gains at substantial parameter cost—and settled on scaling factors of 8 for token number and 3 for token dimension inside the EMA.

The framework is not without trade-offs. SCMamba-SAM’s parameter count of 395 million and computational load of 1,564 GFLOPs far exceed those of the lighter CNN and Transformer baselines it outperforms, a cost the authors acknowledge and flag as a target for future lightweight architectures. The current model also processes single-modality images only; for tasks like focal cortical dysplasia, where T1, FLAIR, and PET sequences carry complementary information, multi-modal fusion could further sharpen segmentation. The training setup—PyTorch 2.1.0 on an NVIDIA A800 GPU, 200 epochs, Adam optimizer at a learning rate of 0.0001, and 256-by-256 input resizing—reflects a research-scale pipeline rather than a deployed clinical system.

Even so, the significance of the work lies in its demonstration that foundation models trained on the natural world can be repurposed for medicine with surgical, rather than wholesale, modification. By combining parameter-efficient LoRA tuning, Mamba-based adapters that inject long-range context without the overfitting penalty of full Transformers, and a prompt-free decoder with deep supervision, SCMamba-SAM turns a general-purpose segmentation engine into an automated lesion-delineation tool. For clinicians, the payoff could be faster, more reproducible quantification of lesions—information that directly informs treatment decisions and surgical planning. For the field, it offers a template: the path to medical AI may run not through training ever-larger models from scratch on scarce clinical data, but through clever, efficient adaptation of the massive models the natural-image world has already built.

Subject of Research: Adapting the Segment Anything Model with Mamba-based modules for automated medical lesion segmentation

Article Title: SCMamba-SAM: transferring knowledge from natural images to medical tasks

Article References: Tian, Y., Yuan, Y., Wang, L., Yang, X., Wu, Y., & Xu, X. (2026). SCMamba-SAM: transferring knowledge from natural images to medical tasks. Frontiers in Neurology, 17, Article 1938737. https://doi.org/10.3389/fneur.2026.1938737

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1938737

Keywords: Segment Anything Model, medical image segmentation, Mamba, state space models, LoRA fine-tuning, deep learning, lesion segmentation, computer-aided diagnosis, Frontiers in Neurology, domain adaptation, Vision Transformer, deep supervision

The study, published in Frontiers in Robotics and AI by researchers at the University of Belgrade’s School of Electrical Engineering and the University of Kragujevac, addresses a persistent gap in circular economy practice. Global municipal solid waste is projected to reach roughly 3.40 billion tons by 2050, up from 2.01 billion tons in 2018, and efficient sorting is a key enabler of recycling, reuse, and recovery—the strategies that keep material value circulating rather than consigning it to landfill. Yet while the waste industry has automated considerably, a stubborn fraction of material at the end of sorting lines still demands manual labor, and existing robotic solutions remain poorly matched to the realities of small-batch production: limited waste volumes, frequent material changes, and the need for rapid reconfiguration.

At the heart of the new system is a perception architecture that departs from conventional deep-learning pipelines. Most robotic sorters rely on supervised models such as convolutional neural networks or YOLO-style detectors, which require large annotated datasets and retraining whenever new object types appear—a serious liability in dynamic industrial settings where waste composition shifts constantly. The researchers instead combined the Segment Anything Model (SAM), a foundation model capable of class-agnostic segmentation without task-specific training, with a lightweight MobileNetV2 classifier for material recognition. The segmentation stage isolates individual items from the conveyor and its surroundings; the classifier then assigns each segmented object to a waste category. Because SAM generalizes across object appearances, adapting to a new packaging type reduces data preparation to sorting cropped images into class folders rather than laborious instance-segmentation annotation.

One practical wrinkle in foundation-model segmentation is redundancy: SAM frequently generates multiple proposals for different parts of the same physical object, such as a bottle’s label or cap. The team’s post-processing pipeline merges these fragments, treating objects almost fully enclosed within a larger bounding box as subregions and merging neighboring masks that share more than 20 percent of their joint contour. Pixel counting across the merged parts then determines the dominant material class. The result is a single bounding box and mask per physical item—the clean, stable representation that downstream localization and grasp planning require.

The perception stack is deliberately hierarchical. Its first level handles core localization and classification for real-time operation. An optional second level adds geometric reasoning: point clouds processed with Open3D are fed to a PointNet network that assigns each object to a grasp-relevant primitive—cylindrical, box-like, or irregular—enabling robust grasp-strategy selection under partial occlusion and independence from surface texture and printing, which vary wildly across packaging brands. A third optional module performs damage assessment using Dynamic Graph Convolutional Neural Networks (DGCNN), refining discrete class outputs into a continuous damage score through a calibrated fuzzy estimator. Damage information is not merely diagnostic: heavily deformed items may need adapted grasping strategies, while sharp or broken objects should be preferentially routed to the robot to keep human operators away from hazardous edges.

The hardware reflects the same collaborative philosophy. The workstation’s centerpiece is a MAiRA Pro M, a seven-degree-of-freedom collaborative robot equipped with torque sensors in every joint and a six-axis force-torque sensor at the end-effector—sensing that underpins safe operation in shared workspaces. Vision comes from an Intel RealSense D455 stereo depth camera, which provides depth resolution up to 1,280 by 800 pixels at 30 frames per second and enables three-dimensional reconstruction of item shape and volume. For grasping, the team selected a modular four-finger soft pneumatic gripper from the SoftGripping construction kit, whose silicone fingers conform to irregular geometries and handle deformable items such as cans, Tetra Pak cartons, and plastic bottles without crushing them. The four-finger configuration was chosen over a three-finger alternative because it proved more robust under uncertainties in object position, orientation, and mass distribution—precisely the conditions a moving conveyor generates.

Software integration runs on ROS2, with motion planning handled by MoveIt and conveyor synchronization achieved through the industrial OPC UA protocol, which feeds encoder data to the control system so it can continuously update the predicted rendezvous point for dynamic grasping. A data-driven task scheduler sits at the top of the hierarchy, allocating work between robot and human operator based on five perception outputs: waste classification, object position, damage level, safety level, and conveyor speed. Objects meeting predefined safety and manipulation constraints go to the robot; items requiring human dexterity or exceeding the robot’s safe handling envelope are delegated to the operator. Notably, occluded, ambiguous, or out-of-distribution objects are labeled as unknown, allowing the system to hand them to a human rather than gamble on a low-confidence grasp—an explicit embodiment of the Industry 5.0 principle that automation should complement rather than replace human workers.

Grasping on a moving conveyor without halting the belt is orchestrated by a six-phase state machine. The robot first moves toward an estimated intersection point where its end-effector trajectory will meet the object’s path, calculated from the object’s initial location and conveyor velocity. It then aligns, synchronizing gripper velocity with the object’s motion along the conveyor line; grasps while actively tracking; lifts vertically to clear the belt; breaks synchronization and moves above the destination bin determined by the initial classification; and finally releases the object before returning to its home position. The approach echoes recent planning-oriented frameworks such as RaapWaste, which decouple task-level scheduling from motion-level planning, but extends them by folding human availability and object safety into the allocation decision.

Laboratory validation focused on three representative packaging categories—aluminium cans, plastic bottles, and Tetra Pak containers—with items manually spaced on the conveyor to emulate the output of an upstream industrial singulation stage. The SAM plus MobileNetV2 pipeline achieved an accuracy of 0.93, precision of 0.93, recall of 0.92, and F1-score of 0.92. The DGCNN damage module reached validation accuracies of 83.5 percent for cans, 88.2 percent for bottles, and 78.7 percent for Tetra Pak, for an overall 83.9 percent across 740 validation samples. On the manipulation side, 200 grasp attempts yielded a 92.0 percent grasp success rate; of the 184 successful grasps, 178 completed the full pick-and-place sequence, giving an end-to-end sorting success rate of 89.0 percent and a mean cycle time of 10.47 seconds—roughly 5.73 picks per minute under controlled conditions. Analysis of the 16 failures pointed to three culprits: unstable initial grasps on severely deformed objects in seven cases, object loss during transfer in five, and gripper-to-object misalignment in four.

The authors are candid about the study’s boundaries. Validation was confined to controlled laboratory conditions with three waste categories, mechanically preconditioned inputs, and a single robotic platform, gripper, and camera; the complete perception pipeline averaged 238.1 milliseconds per frame on an NVIDIA TITAN RTX—about 4.2 frames per second—though the modular architecture allows swapping in faster variants such as FastSAM or EfficientSAM without redesigning the rest of the system. Collaborative scheduling efficiency and operator workload were not quantitatively evaluated. Even so, the work stands as a proof of feasibility for a different vision of recycling automation: one where perception, semantic understanding, and human-robot task sharing are integrated in a reconfigurable workstation sized to the factories that actually need it. As packaging designs and material streams continue to evolve, that flexibility may prove to be the technology’s most valuable output.

Subject of Research: A modular collaborative robotic workstation using SAM-based perception for small-batch packaging waste sorting

Article Title: Development of a small batch flexible station for packaging waste sorting in circular economy applications

Article References: Lukić, B., Petrović, M., Knežević, N., Gordić, Z., Vukićević, A. M., & Jovanović, K. (2026). Development of a small batch flexible station for packaging waste sorting in circular economy applications. Frontiers in Robotics and AI, 13, Article 1870171. https://doi.org/10.3389/frobt.2026.1870171

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1870171

Keywords: circular economy, collaborative robots, waste sorting, Segment Anything Model, deep learning, packaging waste, human-robot collaboration, ROS2, soft gripper, recycling automation, Industry 5.0, computer vision

The research, published in the journal Earth Surface Dynamics, tackles a surprisingly stubborn problem in geomorphology. Grain-size analysis is fundamental to understanding natural hazards, hydrologic conditions, and ecosystems, but traditional field methods are costly, labor-intensive, and slow. A trained observer recording sizes along a stream can typically measure only a few hundred pebbles, yet robust statistical analysis of grain-size distributions often requires far more measurements than that. Photo-based alternatives, from image texture analysis to machine-learning classifiers, have tried to close the gap, but each comes with compromises that limit their usefulness in the messy, shadow-strewn environment of a real mountain river.

OrthoSAM builds on the Segment Anything Model, or SAM, a foundation model released by Meta AI researchers in 2023 that was trained on more than one billion masks across eleven million images. SAM’s remarkable capability is zero-shot inference: it can segment objects in entirely new image domains, from cell microscopy to pebble photographs, without any additional training. Its architecture consists of three decoupled components, an image encoder, a prompt encoder, and a mask decoder, which together allow the model to generate precise segmentation masks from simple input prompts such as points or polygons. For scientific imagery, this flexibility is a genuine advantage over custom-trained neural networks, which tend to perform well only within the narrow conditions they were trained on.

But SAM was never designed for what river scientists need. The model rescales every input image to 1024 by 1024 pixels to fit its transformer architecture, which means a 24-megapixel photograph is effectively compressed down to roughly 0.7 megapixels. For a sprawling orthomosaic containing thousands of densely packed pebbles, that compression is catastrophic: small stones shrink into invisibility and larger objects lose mask quality. The standard automated scheme also applies a grid of 32 by 32 equally spaced input points, optimized for images containing a limited number of objects, not the hundreds to thousands of grains found on a single riverbed image. Hardware constraints compound the difficulty, since each input point generates three candidate masks that must all be held in GPU memory.

The Potsdam team, led by Vito Chan together with Aljoscha Rheinwalt and Bodo Bookhagen, engineered a workflow of three interlocking components to overcome these limits. First, a tiling scheme divides large orthomosaics into 1024 by 1024 pixel patches with a definable overlap, so that SAM processes each patch at full effective resolution. Masks that touch a tile border are discarded to avoid artificially over-segmented pebbles along the seams, and a filtering box ensures only one mask is kept per object. Second, an improved input point generator determines how densely the prompt grid must be spaced so that every object, even the smallest pebble of interest, receives at least one input point, while a centroid-based refinement step reduces duplicate and overlapping masks. Third, a multi-scale resampling scheme runs the segmentation at several resolutions and merges the results, allowing boulders too large to fit within a single tile to be captured in a coarser pass and stitched back into the final labeled map.

Validating such a system posed its own challenge, because reliable ground-truth datasets with thousands of individually delineated pebbles simply do not exist. The researchers therefore built a synthetic pebble generator, producing images of 10,000 by 10,000 pixels filled with up to 5,000 non-overlapping solid circles of random sizes, rendered in black and white, in color, with Gaussian noise, and with simulated shadows cast by hemispherical domes under a directional light source. Because the true positions and sizes of every circle are known, the team could rigorously quantify detection quality, mask accuracy, and the fidelity of the resulting size distributions.

The results were striking. On 1,872 synthetic black-and-white pebbles, OrthoSAM achieved a precision of 1.0, a recall of 0.87, and a mean Intersection over Union, the standard measure of mask accuracy, of 0.98. On 27,528 colored pebbles with shadows, the workflow reached a precision of 0.98, a recall of 0.94, and a mean IoU of 0.91. A two-sample Kolmogorov-Smirnov test confirmed that the predicted grain-size distributions were statistically indistinguishable from the ground truth, with p-values exceeding 0.25 in every synthetic image. The team also identified a crucial detection limit: pebbles with a diameter below 30 pixels are not reliably detected, a threshold that directly informs how field crews should plan camera distance and image resolution.

To prove the concept on real terrain, the researchers applied OrthoSAM to three orthomosaics from the Ravi River in the western Himalaya, images assembled from hundreds of photographs taken with a Sony camera and processed to a spatial resolution of 0.2 millimeters per pixel. The workflow delineated 6,087 pebbles across the three scenes, with manual verification of every predicted mask yielding a precision of 0.93 and a recall of 0.94. For each pebble, the software computes a rich table of measurements, including projected area, the lengths of the a- and b-axes, perimeter, color statistics, and a normalized isoperimetric ratio that serves as a proxy for roundness, all exported in a format ready for further geomorphic analysis.

The study is candid about the method’s remaining weaknesses. Shadows proved to be a subtle adversary: pebbles completely covered by shadow were segmented as accurately as those in uniform light, but partially shadowed pebbles, where stark contrast divides a single object into lit and dark halves, were responsible for most of the observed drop in precision and mask quality. Strong image noise also degrades accuracy, prompting the authors to recommend keeping camera ISO settings moderate during data collection. Because SAM performs only instance segmentation without any classification component, non-pebble objects such as marker panels, wooden branches, and structure-from-motion artifacts can also be delineated and must be filtered out before statistics are computed. The team calls for a community-based reference dataset for granular material segmentation, analogous to the benchmark datasets that transformed lidar research.

The implications reach well beyond one Himalayan river. Grain-size distributions govern how sediment moves through drainage basins, how flood hazards develop, and how aquatic ecosystems function, and the ability to sample more than ten thousand objects from a single virtual outcrop promises a new scale of observation for Earth scientists. OrthoSAM’s Python-based pipeline is openly available on GitHub and archived on Zenodo, and the Ravi River orthomosaics are published under an open license, meaning any research group with a capable GPU can begin counting stones. A 10,000 by 10,000 pixel synthetic image currently takes about four hours to process on a 16-gigabyte NVIDIA Quadro RTX 5000, so wider adoption will depend on computing resources, but the direction is clear. The humble pebble count, a field technique essentially unchanged since 1954, is being reborn as an automated, reproducible, and scalable digital science.

Subject of Research: Automated delineation of river pebbles and grain-size analysis from high-resolution orthophotos using an extended Segment Anything Model

Article Title: OrthoSAM: multi-scale extension of the Segment Anything Model for river pebble delineation from large orthophotos

Article References: Chan, V., Rheinwalt, A., & Bookhagen, B. (2026). OrthoSAM: multi-scale extension of the Segment Anything Model for river pebble delineation from large orthophotos. Earth Surface Dynamics, 14(3), 391-416. https://doi.org/10.5194/esurf-14-391-2026

Image Credits: AI Generated

DOI: 10.5194/esurf-14-391-2026

Keywords: OrthoSAM, Segment Anything Model, grain-size analysis, geomorphology, river pebbles, orthomosaics, deep learning, image segmentation, remote sensing, Himalaya, Ravi River, sediment transport

The stakes are higher than they might appear. Grain morphometry underpins a remarkable breadth of Earth science. The sizes and shapes of clasts record how sediments are produced on hillslopes, how rivers hydraulically sort and transport their loads, and how deposits in the stratigraphic archive were laid down. Grain size data have even been used to reconstruct transport conditions on Mars, where the shapes of ancient pebbles hint at how far water once flowed. But the traditional ways of collecting these data are laborious and biased. Manual Wolman counts and line sampling yield relatively few observations and depend heavily on the observer. Point-cloud methods struggle with grain occlusion and require geometric assumptions. Earlier image-based tools either need field calibration or systematic manual correction, and they tend to over- or underestimate grain sizes in predictable but troublesome ways.

Machine learning promised to change this, and to some extent it has. Texture-based neural networks can predict summary statistics of a grain size distribution directly from image texture, while segmentation-based approaches delineate every grain individually, opening the door to datasets of tens of thousands of measurements. The catch, as Mair and colleagues emphasize, is that these segmentation models have behaved like narrow specialists. Trained on homogeneous, task-specific datasets, they lose accuracy the moment they encounter a new camera, a new lighting regime or a new sediment type. Each new setting has effectively demanded a new model and a freshly curated training set, which has blunted much of the promised efficiency.

The Bern team’s solution was to reach for a different class of model altogether. In recent years, transformer-based foundation models such as Meta’s Segment Anything Model, or SAM, trained on eleven million images and more than a billion annotation masks, have shown an extraordinary ability to generalize to unseen data. But SAM in its basic form is inefficient at densely segmenting large numbers of similar objects, which is precisely what a gravel bar photograph demands. The breakthrough came with Cellpose-SAM, a next-generation model developed for segmenting cells in biomedical imagery. It grafts SAM’s transformer image encoder onto the flow-based segmentation machinery of the original Cellpose framework, which tracks predicted vector flows to separate touching objects. The hybrid keeps SAM’s generalization while excelling at dense, prompt-free segmentation of many instances of one object class, and its architecture has grown to more than 304 million trainable parameters, roughly fifty times larger than the U-Net backbone it replaces.

Retraining such a model for geoscience required data, and lots of carefully annotated ones. The researchers assembled the ImageGrains 2.0 dataset, combining 81 previously labelled tiles with 162 new ones for a total of 243 image tiles containing more than 29,000 manually annotated grain masks. The imagery spans fluvial gravels photographed from uncrewed aerial vehicles and handheld cameras in Switzerland, Spain, Peru, New Zealand and Namibia; painted and unpainted pebbles used to track grain mobility; angular proglacial deposits; near-vertical outcrops of lithified conglomerates; X-ray computed tomography slices of glacial till; and micro-CT images of bioclastic marine sand. Crucially, the team deliberately included awkward content, such as vegetation, sieves, hands, scale objects, water bodies and harsh shadows, to test whether the model could distinguish genuine grains from everything else in a messy field photograph.

The performance gains are substantial. Benchmarked against the manual annotations using average precision at an intersection-over-union threshold of 0.5, the fine-tuned Cellpose-SAM model achieved a median score of 0.72 across all image tiles, roughly 20 percent higher than the second-best method, a Cellpose 2 model trained on the same data. The advantage held in both training and unseen test splits, and it proved largely independent of how many images each sediment type contributed to training. The model scored highest of all on X-ray CT tiles, which made up only about a tenth of the dataset. Even more telling was the generalization test: when the researchers withheld two entire subsets from training, the model still matched or beat specialist models that had been fine-tuned on those very images, evidence that the transfer-learning strategy preserves the broad visual competence of its SAM heritage.

Accurate masks translate directly into accurate measurements, and here the numbers are striking. The model recovered 93 percent of ground-truth grains, and mean differences in grain size were around 3 percent or less for both the long and short axes of ellipse fits to each grain. The resulting grain size distributions were statistically indistinguishable from the ground truth in 88 percent of cases for the a-axis and 82 percent for the b-axis, compared with only 54 and 57 percent for the second-best model. Shape metrics fared equally well, with average deviations below 2 percent for roundness, roughness, elongation and orientation. The team also identified a practical rule of thumb: perfect statistical agreement between predicted and true grain size distributions was only achieved when segmentation scores exceeded roughly 0.68 average precision, a threshold that varied with how well sorted the sediment was.

The framework extends into three dimensions as well. Applying the model to a stack of 400 X-ray CT slices from a drill core of glacio-fluvial sediment in southern Germany, the researchers segmented 4,647 well-defined coarse grains in full volumetric detail, something 2D photographs can never provide. Because CT scans capture complete grains rather than partially occluded surfaces, they sidestep the geometric extrapolations that plague LiDAR and photogrammetry-based approaches. The authors note that rigorous 3D benchmarking still awaits a manually annotated volumetric ground truth, whose creation has so far been impractical, but the demonstration suggests CT-based sediment analysis, from clast fabric to depositional history, is now within reach of automated workflows.

Accessibility was a design priority. ImageGrains 2.0 ships as an installable Python package with an interactive graphical user interface for annotating images and correcting model predictions, so researchers need not write code at all. Training runs in under ninety minutes on a high-end GPU, and although transformer models are demanding, the authors show that fine-tuning remains feasible on a mid-range desktop graphics card, with inference requiring as little as 3 gigabytes of memory. Because the model builds on already-trained foundation weights, its marginal energy footprint is modest compared with training such architectures from scratch. Adapting the model to a genuinely new sediment type or imaging style can require as few as seven additional annotated tiles, a dramatic lowering of the barrier that once made every new site a modelling project of its own.

The deeper message of the study concerns where progress will come from next. The authors argue that future gains in grain measurement will depend less on new architectures than on larger, more variable, publicly shared annotated datasets, ideally with annotations from multiple experts to establish consensus-level benchmarks. They explicitly frame their work as a step toward a foundation model for granular particles in geoscientific imagery, one that could serve any discipline needing dense segmentation and morphometric analysis, from sedimentology to planetary science. With the dataset, model weights and code all openly available, the Bern team has effectively invited the community to build on their groundwork, turning what was once a tedious field chore into a problem that a well-fed neural network can now handle in minutes.

Subject of Research: Deep-learning-based automated segmentation and morphometric measurement of sediment grains in geoscientific imagery

Article Title: ImageGrains 2.0: Improved precision and generalization for grain segmentation

Article References: Mair, D., Witz, G., Do Prado, A., Garefalakis, P., Wild, A., Ville, F., Schuster, B., Horn, M., Österle, J., Fabbri, S. C., Litty, C., Achleitner, S., Leistner, S., Hiller, C., & Schlunegger, F. (2026). ImageGrains 2.0: Improved precision and generalization for grain segmentation. Earth Surface Dynamics, 14(4), 527-551. https://doi.org/10.5194/esurf-14-527-2026

Image Credits: AI Generated

DOI: 10.5194/esurf-14-527-2026

Keywords: ImageGrains 2.0, grain segmentation, Cellpose-SAM, deep learning, sedimentology, grain size distribution, X-ray computed tomography, transfer learning, Segment Anything Model, fluvial sediment, geomorphology, open-source software

The research tackles a stubborn bottleneck in outdoor recreation management: the concept of carrying capacity has evolved well beyond the search for a single maximum number of visitors. Modern practice, codified in the Interagency Visitor Use Management Framework, asks managers to define desired conditions, select measurable indicators, set thresholds, and link monitoring to adaptive actions. Yet implementation routinely stalls at the measurement step. Manual counts are expensive, automated counters suffer from group undercounting and false triggers, and neither approach tells managers where vehicles can plausibly park, when parking fills, or how use concentrates along a linear corridor that may stretch for many kilometers. The Payette River, with its fragmented jurisdiction spanning the USDA Forest Service, the Bureau of Land Management, state transportation agencies, and local law enforcement, exemplifies the problem.

The authors built two complementary workflows. The first quantifies what they call parking opportunity, the spatial extent of surfaces where a vehicle could realistically be parked, using high-resolution aerial imagery from the National Agriculture Imagery Program at 0.6-meter resolution. The analysis area was defined by intersecting 500-meter buffers around major highways and the river itself, then constrained to terrain with slopes of five percent or less using a one-meter digital elevation model from the USGS 3DEP program. This slope filter was applied twice, first to screen candidate training labels and again to mask out implausible predictions on steep hillsides during inference, ensuring the model never reports parkable surface on a cliff face.

Label development for the segmentation workflow relied on the Segment Anything Model, or SAM, in its largest ViT-H configuration. SAM generated 12,942 candidate object proposals from the tiled aerial imagery, but manual review revealed that most of these were trees, a striking illustration of why general-purpose foundation models still require human curation when applied to infrastructure mapping. After screening, 817 polygons survived with meaningful labels, of which only 16 represented paved parking and 50 represented dirt or gravel parking. This severe scarcity of positive examples shaped the modeling strategy: the team trained a DeepLabV3+ network with a ResNet-34 encoder using a weighted multi-class cross-entropy loss that penalized parking misclassification heavily, and applied on-the-fly augmentations including flips, rotations, affine transforms, and brightness perturbations to squeeze more effective diversity from the sparse labels.

The final segmentation model achieved a mean intersection-over-union of 0.64 across classes, with 0.60 for paved parking and 0.56 for dirt and gravel parking, substantially better than an unweighted, non-augmented baseline. Post-hoc checks found that fewer than five percent of model-derived paved parking areas fell within four meters of roadway centerlines, suggesting that confusion between roads and parking lots, the error mode with the greatest practical consequence, was marginal in this setting. The resulting maps revealed that parking opportunity is highly uneven along the corridor, giving managers a replicable, spatially explicit supply-side baseline for the first time. The authors are careful to note that mapped opportunity is not the same as authorized or safe parking, and that field verification of ownership, legality, and stall geometry remains essential before any investment decisions.

The second workflow addresses the demand side, using YOLO26 object detection applied to a large corpus of images from motion-triggered cameras positioned at parking areas and river viewpoints. Rather than labeling images at random, the team employed an active learning strategy: a COCO-pretrained baseline detector produced preliminary detections, and an uncertainty- and diversity-aware selector chose an initial batch of 500 images for human annotation, prioritizing ambiguous predictions, hard negatives likely to cause false positives, and predicted-empty images to stabilize specificity. A second loop of 1,000 river-only images specialized the detector for the notoriously difficult visual conditions of water scenes, where glare, specular highlights, partial occlusion, and clustered watercraft confound naive models. Two annotators achieved inter-annotator agreement above 0.80 on Krippendorff’s alpha, following strict rules such as labeling only the visible portion of occluded vehicles and refusing to label reflections or wake patterns as boats.

Watercraft remained so rare in the imagery that the workflow added a targeted mining phase. The team calibrated a confidence threshold by sweeping candidate values against validation data, then ran large-scale inference across the full river corpus, ranking images by a mining score that favored confident, meaningful detections. High-ranked candidates were annotated and oversampled during a final fine-tuning stage, with mined examples strictly restricted to the training pool to avoid evaluation leakage. The payoff was dramatic: watercraft recall increased substantially in the final iteration alongside precision of 0.843, lifting mean average precision at the 0.50 threshold to 0.554. Vehicle detection, by contrast, was strong throughout, reaching precision of 0.856, recall of 0.733, and mAP50 of 0.865 in the final model, supporting the use of vehicle counts as a reliable comparative indicator of access pressure.

Aggregated results paint a vivid picture of corridor dynamics. Parking occupancy proxies showed sharp within-day pulses aligned with arrival and departure cycles, with the highest demand at Banks and Beehive Bend concentrated on weekends and midday, while sites such as Alder Creek, Danskin, and Deadwood showed much smaller peaks. On the river, watercraft detections clustered into two pronounced windows, mid-morning and mid-afternoon, with the Banks-to-Beehive stretch showing the most intense activity, consistent with typical launch-to-takeout travel times. Calibration against 67 hours of manual hand-tally counts yielded only modest absolute agreement, with a correlation of 0.35 and an R-squared of 0.12, but crucially the temporal ordering of peak-use windows and the relative ranking of sites remained stable between calibrated and uncalibrated series. The authors therefore recommend treating the on-river metric as a relative activity indicator rather than an absolute census of boats.

The study’s broader significance lies in its honesty about what machine learning can and cannot deliver for public land management. The authors explicitly frame their outputs as physical infrastructure and demand indicators, necessary inputs to capacity frameworks but not capacity thresholds themselves, since acceptable levels of crowding and ecological impact require social and ecological standards that no camera can supply. They also confront the ethics of camera deployment in public spaces, recommending placement that avoids capturing faces, resolution limited to detection needs, masking of sensitive regions, signage, and aggressive data retention policies that discard raw imagery once detections are extracted. Post-hoc benchmarks showed DeepLabV3+ outperforming U-Net and FPN architectures for segmentation, while the compact YOLO26n variant, at just 2.5 million parameters versus 32.1 million for an FCOS comparison, delivered comparable vehicle detection, an operational advantage for resource-constrained agencies.

For a field where monitoring capacity has long been the weakest link in adaptive management, the Payette River case offers a template that could extend to trail networks, scenic byways, and other linear recreation corridors wherever aerial imagery and fixed cameras are feasible. The workflow’s active learning strategy, in which a modest amount of targeted annotation informed by uncertainty and hard-example mining yields outsized gains on rare but management-critical targets, transfers naturally to other sparse-event monitoring problems. The authors caution that generalization beyond a single corridor remains a working hypothesis, and that seasonal domain shift, changing vegetation, and camera mounting constraints will demand local recalibration. But the core message is clear and timely: with careful human-in-the-loop design, transparent error accounting, and explicit acknowledgment of uncertainty, computer vision can finally give recreation managers the continuous, spatially explicit evidence they need to diagnose bottlenecks before they become conflicts.

Subject of Research: Computer vision workflows for measuring parking capacity and recreation demand along a linear river recreation corridor

Article Title: Computer vision workflows for physical carrying capacity analysis and monitoring in linear recreation corridors

Article References: Smith, J. W., & Lamborn, C. C. (2026). Computer vision workflows for physical carrying capacity analysis and monitoring in linear recreation corridors. Machine Learning with Applications, 26, Article 101012. https://doi.org/10.1016/j.mlwa.2026.101012

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101012

Keywords: computer vision, carrying capacity, outdoor recreation, semantic segmentation, object detection, YOLO, Segment Anything Model, DeepLabV3+, active learning, Payette River, visitor monitoring, machine learning

The stakes are considerable. Upper respiratory and gastrointestinal infections remain among the most common causes of illness in school-aged children worldwide. According to figures cited by the research team, more than 90 percent of children aged three to six catch at least one respiratory infection each year, and roughly half experience a diarrhoea-related illness. Attendance at kindergarten or primary school amplifies this burden, because classrooms are environments of constant physical contact, shared surfaces, and imperfect hygiene. Proper hand hygiene is regarded as the most cost-effective and straightforward measure available to reduce these infections and the school absenteeism that comes with them. Crucially, healthy habits formed in early childhood tend to persist into adulthood, which is why schools are such an attractive setting for intervention.

The World Health Organization’s 2009 guidelines on hand hygiene in health care already promote multimodal strategies that combine education, supplies, and real-time feedback. A systematic review of eighteen cluster-randomized controlled trials concluded that such programmes can reduce respiratory-tract infections and sickness-related absence, although the methodological quality of the underlying studies varied. Previous work has also highlighted the pedagogical power of fluorescent markers viewed under ultraviolet light. When children apply a glowing gel and then see, under a black light, exactly which parts of their hands they missed, the invisible becomes vividly concrete. Studies in paediatric waiting rooms and primary schools have shown that this kind of visual concretization can improve washing technique, and the evidence generally suggests that combining interactive education with fluorescent visualisation works better than either element alone.

What has been missing, the Catalan researchers argue, is objectivity and scale. Traditional UV-based interventions rely on manual or semi-automated scoring of photographs, which is time-intensive and vulnerable to observer variability. To address this, the team—based at the Preventive Medicine Unit and Infection Control Group of Joan XXIII University Hospital and collaborating primary care researchers—developed a semi-automated, AI-assisted image analysis pipeline, and, to their knowledge, no published school-based intervention study had previously integrated such a tool into UV-based hand hygiene feedback and outcome assessment.

The study was designed as a controlled before-after repeated measures study conducted within the Sentinel Schools Network of Catalonia in the Camp de Tarragona region during the 2023 to 2025 academic period. Two primary schools were matched in advance by low-socioeconomic urban context, management type, and school size, then pragmatically assigned to intervention and control groups to minimise cross-contamination between arms. Sixty-four pupils at the intervention school and sixty-two at the control school took part. Hand hygiene was assessed at baseline and again at one, three, and six months after the intervention. At each assessment, participants applied roughly one millilitre of fluorescent alcohol-based hand rub, rubbed their hands, and placed them inside a custom-built UV device that provided a uniform dark background and controlled illumination. Two photographs per child captured the dorsal and palmar surfaces under standardised conditions, and the percentage of adequately covered hand surface was calculated.

The intervention itself was deliberately brief. Pupils at the intervention school received a thirty-minute educational session covering the seven-step technique for alcohol-based hand rub and the role of hand hygiene in infection prevention, delivered through a short video, slides, and classroom materials, followed by a practical activity with real-time visual feedback. No reinforcement activities took place between follow-up visits. Control-school pupils received the same educational session only after the study ended; during the study they were exposed to the UV device during assessments but received no direct feedback on the areas they had missed.

The AI pipeline worked in three stages. First, a YOLOv8 object detection model localised each hand and classified its view and laterality—right or left, dorsal or palmar. The team labelled 5,197 images of hands and augmented them to 57,163 through transformations such as mirroring, rotation, translation, scaling, and illumination changes, training on 51,446 images and validating internally on 5,717. Evaluated on 2,381 new labelled images after 247 training epochs, the model achieved an overall accuracy of 94 percent across the four hand classes. Second, a pretrained Segment Anything Model from Meta segmented each hand from its background without additional training, failing on only 48 hands at first attempt, 38 of which were recovered with minimal manual adjustment. Third, colour isolation using thresholds in hue, saturation and value space identified the fluorescent regions and generated binary coverage masks. This final stage proved the weak link: HSV-based fluorescence detection showed low agreement with ground-truth annotations and required human supervision, a limitation the authors attribute to image-to-image illumination variability that could be mitigated by redesigning the device or adopting a more robust automated approach.

The effectiveness results told a nuanced story. At baseline, adjusted estimates of clean hand surface coverage were similar in both schools, at 0.888 in the control school and 0.910 in the intervention school. At one month, both groups improved comparably, reaching 0.937 and 0.943 respectively—a rise the authors suggest may reflect assessment awareness, the implicit prompting of the UV procedure itself, or limited contamination between school communities. From three months onward, however, the trajectories diverged. The control school regressed partially toward baseline, falling to 0.913, while the intervention school held and then built on its gains, reaching 0.944 at three months and 0.958 at six months. The adjusted between-school differences became statistically significant at three months, with a difference of 0.031, and widened at six months to 0.041. Within the intervention school, all follow-up contrasts against baseline were statistically significant, with the largest gain at six months; within the control school, only the one-month change was statistically supported. The analysis used a beta mixed-effects regression model with a logit link, a random intercept for each pupil to account for repeated measures, and adjustments for multiple testing. Girls performed modestly better than boys, and the model revealed meaningful between-pupil heterogeneity in baseline performance.

The authors are careful about what these numbers can and cannot show. Because the study involved only a single matched pair of schools, the between-school contrasts are treated as exploratory rather than confirmatory cluster-level causal estimates. Residual confounding from unmeasured school-level characteristics cannot be excluded, and baseline differences in age, grade composition, household size, and parental education between the two schools were present, although additional statistical adjustment for age and grade did not meaningfully change the results. The incomplete end-to-end automation of the image analysis is a second acknowledged limitation. Still, the researchers argue that embedding prevention and health promotion interventions in school settings can generate actionable evidence for public health and education practice, and that if the approach is replicated across a larger number of clusters and more diverse contexts, it could be scaled up and integrated into routine school health programmes—particularly if measurement and feedback can be delivered with minimal staff burden.

The broader significance lies in the marriage of two ideas. One is pedagogical: making the invisible visible, so that a child can see the glowing residue on the backs of the hands or between the fingers that a quick rinse leaves behind. The other is technological: replacing subjective human scoring with a standardised, reproducible computer-vision workflow that limits observer bias and could, in principle, support large-scale monitoring across many schools. Neither idea is new on its own, but their combination in a real school intervention is a first, and the six-month durability of the effect—modest in absolute terms but statistically consistent—adds to growing evidence that structured instruction plus UV-based feedback can sustain improvements where feedback alone cannot. For a measure as cheap and universal as handwashing, even small, persistent gains in technique among children could translate into fewer infections, fewer missed school days, and habits that last a lifetime.

Subject of Research: A school-based hand hygiene intervention using UV visual feedback and AI-assisted image analysis in primary school children

Article Title: Evaluation of a hand hygiene intervention in two primary schools using UV-based feedback and AI-assisted image analysis

Article References: Bordas, A., Colom-Cadena, A., Aceiton, J., Martínez-Torres, S., García-Pino, A., Escaramis, G., Muntada, E., Rey-Reñones, C., Gens-Barberà, M., Casabona, J., Basora, J., & Martín-Luján, F. (2026). Evaluation of a hand hygiene intervention in two primary schools using UV-based feedback and AI-assisted image analysis. Public Health in Practice, 12, Article 100856. https://doi.org/10.1016/j.puhip.2026.100856

Image Credits: AI Generated

DOI: 10.1016/j.puhip.2026.100856

Keywords: hand hygiene, primary schools, ultraviolet feedback, artificial intelligence, image analysis, YOLOv8, Segment Anything Model, infection prevention, school health promotion, children, public health, Catalonia

The work, published open access in the Journal of Ambient Intelligence and Humanized Computing, responds to a long-standing problem in environmental communication. Conventional charts are excellent for analysis, but research on psychological distance shows that large-scale, slow-moving phenomena such as climate change often fail to register as personally urgent when they are presented as abstract numbers. Eco-visualization, a practice first articulated in museum contexts and later explored in energy-feedback research, attempts to close that gap by blending data representation with aesthetic, narrative, and experiential elements. The catch has always been skill: designing these hybrid images demands both creative sensibility and technical competence in data processing, segmentation, and color mapping, a combination that excludes many talented designers.

The team’s answer is a structured, six-step Human-AI co-creative pipeline. The process begins with generative image creation: the user describes a dataset topic, and a text-to-image model proposes visually meaningful scenarios, such as a flooded coastal city for sea-level data or a bleached coral reef for ocean acidification. The user can accept a suggestion, request variations, or start over, iteratively steering the AI toward a contextually appropriate scene. Next comes dataset configuration, in which the user uploads structured quantitative data in CSV or TSV format; the system parses the file, displays columns as interactive chips, and pre-computes minimum and maximum values that will later define the color scale.

The third step is where the machine-learning engineering becomes visible. The system integrates Meta AI’s Segment Anything Model, or SAM, to automatically analyze the generated image and produce pixel-level masks for each identifiable object or region. Three Vision Transformer backbones of increasing size, ViT-B, ViT-L, and ViT-H, are available, but the interface deliberately hides that choice: ViT-B runs by default, and larger variants sit behind an optional advanced control. If the automatic masks miss something, a lasso tool lets users draw custom regions, and a mask editor supports merging segments through Boolean union operations, undoing recent changes, or resetting to the original output. The design philosophy is that users should evaluate the masks that matter to their visualization, not choose model architectures.

Color is treated with scientific seriousness. The system encourages palettes validated for perceptual uniformity and accessibility, such as the scientific color maps developed by geoscientist Fabio Crameri and colleagues, whose 2020 Nature Communications paper documented how arbitrary gradients can mislead viewers and distort quantitative interpretation. Users can also build custom palettes from start and end colors, and the AI suggests alternatives derived from the dominant colors of the generated image so the visualization blends naturally into its scene. During overlay generation, the chosen palette is mapped linearly onto the dataset’s value range, and the gradient is applied only inside the masked region using alpha blending, with independent hue, saturation, and lightness coefficients giving designers fine-grained control over how strongly data colors replace the original pixels. A separate edge-smoothing parameter softens boundaries between modified and untouched areas.

To understand whether this pipeline actually works for its intended audience, the researchers ran a remote user study with 13 participants, twelve of them from design backgrounds, chosen precisely because they combine visual expertise with limited experience in data-driven toolchains. Participants received no training, only a brief written introduction and a guided task: generate a mountain landscape, load monthly atmospheric CO2 data, merge masks if desired, colorize only the lake region with a scientific palette, and export the result. The evaluation combined the NASA Task Load Index for cognitive load, the short User Experience Questionnaire for pragmatic and hedonic quality, open-ended feedback, and a demographic section, all collected anonymously in compliance with GDPR.

The results were strikingly divided. Cognitive load, measured on a 0 to 100 scale, averaged 59 with a standard deviation of 16.1, squarely in the high range, and the spread suggests the mental burden varied considerably from person to person. Pragmatic quality, the UEQ-S dimension capturing usability and task effectiveness, scored a mean of minus 0.96, below the minus 0.8 threshold that marks a problematic rating. Items probing whether the system felt easy and clear scored poorly, and participants reported that the browser froze during segmentation, a consequence of running the model locally on CPU. One participant described the HSL adjustments as monotonous and struggled to grasp what the parameters actually changed, hinting that the interface failed to communicate the rationale behind its controls.

Yet hedonic quality told the opposite story, averaging 1.33, a clearly positive evaluation. Participants rated the system as highly inventive and on the leading edge, with scores of 6.23 and 5.69 respectively on those items, indicating that the core concept of embedding environmental data inside emotionally resonant imagery resonated strongly even as the interaction stumbled. The overall UEQ-S score of 0.18 landed in the neutral range, with the system’s conceptual appeal only partially compensating for its usability problems. The authors argue this gap exposes a fundamental tension in Human-AI co-creation: granting users expressive control increases creative freedom but inflates cognitive demand when the purpose of each parameter is not immediately obvious, pointing toward design strategies like progressive disclosure and contextual explanation.

The team is candid about the study’s limits. The sample was small, deliberately composed of designers, and constrained to a prescribed task with a shared dataset, so the findings speak to first-use learnability rather than open-ended creative practice. CPU-based segmentation introduced latency that degraded perceived responsiveness, and the AI’s role, while framed as collaborative, remains fundamentally reactive: it responds to prompts but does not learn from users across sessions or proactively shape the creative process, placing the current system closer to AI-assisted design than to fully bidirectional co-creation. Future work will move segmentation to GPU-accelerated or cloud inference, add contextual guidance for parameter controls, and run ecologically valid studies in which participants choose their own datasets, audiences, and visual narratives, ideally involving environmental scientists, educators, and communication professionals alongside designers.

Even with those caveats, the contribution is concrete and timely. At a moment when generative models are flooding the internet with images, this research asks a sharper question: can those models help make invisible environmental processes visible without sacrificing scientific integrity? By pairing perceptually reliable color science with iterative human steering, the pipeline sketches a workflow in which data and imagination reinforce rather than compete with each other. The mountain lake that darkens as carbon dioxide climbs is not a replacement for the chart behind it, but it is a complement that speaks to a different part of the mind, and the study’s mixed verdict, captivating concept, demanding interface, maps the road that co-creative tools must travel before that vision becomes routine practice in climate communication.

Subject of Research: A human-AI co-creative pipeline for designing data-driven eco-visualizations

Article Title: Co-creating eco-visualizations: a human–AI collaborative pipeline for data-driven visual design

Article References: Ceccarini, C., Bottari, B., Liçaj, A., Matteucci, E., & Delnevo, G. (2026). Co-creating eco-visualizations: a human–AI collaborative pipeline for data-driven visual design. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10.1007/s12652-026-05128-w

Image Credits: AI Generated

DOI: 10.1007/s12652-026-05128-w

Keywords: eco-visualization, human-AI co-creation, data visualization, generative AI, image segmentation, Segment Anything Model, environmental communication, cognitive load, user experience, scientific color palettes, climate data, co-creative systems

The Segment Anything Model, or SAM, was introduced by Meta AI researchers in 2023 as a foundation model for image segmentation, trained on more than a billion masks across eleven million images. Its promise was audacious: the ability to segment, or precisely outline, virtually any object in any image without task-specific training. In medical imaging, remote sensing, pathology, and even planetary geology, researchers have rushed to harness SAM’s general-purpose vision capabilities. But SAM has a well-documented Achilles heel. When confronted with images dominated by dense, repetitive, fine-grained textures—the branching vessels of a retina, the tangled architecture of a surgical field, or in this case the grooves, channels, and tread patterns of a shoe sole—the model’s performance drops sharply. Footwear outsoles are exactly this kind of challenge. Their tread designs combine geometric regularity with manufacturing variation, wear patterns, and material contrasts that confound a model trained mostly on natural scenes and everyday objects. Shaghayegh Chavoshian, Ali Barzegar Khanghah, and Atena Roshan Fekr, based at the KITE Research Institute of the Toronto Rehabilitation Institute and the Institute of Biomedical Engineering at the University of Toronto, set out to close that gap.

Their strategy was transfer learning, a technique in which a model pretrained on a vast general dataset is retrained, or fine-tuned, on a smaller but highly specific dataset so that its general visual knowledge is redirected toward a narrow task. Fine-tuning foundation models has already proven fruitful in medical image segmentation, where adapted SAM variants have been used for anatomical structures, tumor delineation, and surgical video analysis. The Toronto team reasoned that the same principle should apply to footwear science, where manual annotation of outsole images is notoriously slow and expensive. Segmenting an outsole by hand requires a trained expert to trace, pixel by pixel or polygon by polygon, the boundaries between tread features and the spaces between them—a process that consumes substantial time per image and becomes prohibitive when applied across the many shoes a serious slip-resistance study requires. Automating this bottleneck would open the door to analyzing footwear at a scale previously impractical, potentially transforming how safety standards are written and how shoes are evaluated for consumers, workers, and older adults. The motivation is reinforced by a long line of prior work showing that outsole features such as tread groove geometry, sole hardness, and material wear all influence the friction available at the shoe-floor interface, which makes accurate, scalable measurement of outsole geometry a genuinely valuable scientific target rather than a mere convenience.

To fine-tune the model, the researchers assembled a dataset of forty footwear outsoles, each manually annotated using a graphical annotation tool that allows precise polygonal outlining of image features. Forty shoes may sound modest next to SAM’s eleven-million-image pretraining corpus, but that is precisely the point of transfer learning: the general model already understands edges, shapes, and textures; it needs only a comparatively small volume of domain-specific examples to learn what matters in a shoe sole. The ground truth against which the model was judged came not from mechanical friction devices alone but from human-centered data, in which footwear had been classified as having either low or high slip resistance based on real human testing. Some of that human-centered footwear data was drawn from open access material available through the Rate My Treads website, a resource that aggregates winter footwear performance information. This choice of labels matters. Mechanical slip testers measure friction under controlled conditions, but human slip resistance emerges from an interplay of gait biomechanics, loading rates, and perception that benchtop devices only approximate. By anchoring the labels to how shoes actually perform on people’s feet, the researchers kept the machine learning pipeline aligned with the outcome that ultimately counts—whether a person slips.

The results of the fine-tuning were substantial. Compared with the original, out-of-the-box SAM, the adapted model reduced segmentation loss by 8.11 percent and lifted the intersection over union, the standard overlap metric between predicted and true segmentation masks, to 70.45 percent. Perhaps more strikingly, pixel accuracy climbed from 56.90 to 78.10 percent, and the F1 score, which balances precision and recall, rose from 53.70 to 66.30 percent. Those numbers tell a clear technical story: without task-specific adaptation, SAM could barely delineate outsole features better than chance on some measures, but after fine-tuning it captured roughly two-thirds to three-quarters of the relevant structure. The researchers also examined which image quality factors influenced performance, finding that resolution, contrast, and intensity all significantly affected segmentation quality. This is consistent with a growing literature showing that deep segmentation networks are sensitive to the spectral and spatial characteristics of their inputs; a model fine-tuned on images of one resolution or contrast profile may degrade when fed imagery that differs. For anyone hoping to deploy such systems in the field—photographing shoes in a store, a clinic, or a workplace—the finding underscores that image capture protocols must be standardized for reliable results.

The segmentation outputs were then put to work. From each predicted outsole mask, the pipeline estimated the outsole-ground contact areas—the regions of the sole that would actually press against a floor during walking. These contact features were fed into a downstream machine learning classifier tasked with predicting the slip resistance category of the shoe, low or high, using the human-derived ground truth labels. With an 80/20 train-test split, the classification model achieved 70 percent accuracy and a 64 percent F1 score. While these figures are a distance from clinical certainty, they represent a meaningful proof of concept: the geometry of a shoe’s contact patch, extracted automatically by an adapted foundation model, carries enough information to forecast, with better-than-chance reliability, how the shoe will behave under a human foot on a slippery surface. Earlier work by the same group and collaborators had shown that convolutional neural networks and machine learning models could predict slip resistance from engineered tread features, including studies of winter footwear on glycerol-contaminated surfaces; the new study shortens that pipeline by replacing manual feature extraction with learned segmentation.

The broader context makes the advance more than an academic exercise. Winter footwear rated for slip resistance is still evaluated largely through mechanical tests and, increasingly, through human-centered trials such as those conducted with the maximum achievable incline method, in which participants walk up progressively steeper icy slopes until they slip. Portable slip simulators, cart-type friction measurement devices, and computational models of shoe-floor friction have each expanded the toolkit, but these approaches remain expensive, time-consuming, and difficult to scale to the flood of new footwear models reaching the market each year. Rating programs that rely on such testing simply cannot keep pace. If a vision model can screen outsole designs computationally, flagging promising candidates for human testing and discouraging poor performers before they reach consumers, the entire evaluation ecosystem could accelerate. The authors point to implications for two populations in particular: older adults, for whom a fall can trigger a cascade of fractures, hospitalization, loss of independence, and mortality, and workers in occupations—from construction and healthcare to marine and winter industries—who depend on protective footwear to stay safe in hazardous environments.

The study also contributes to a rapidly evolving conversation about when and how to adapt foundation models. Across computer vision research, SAM has been shown to struggle in concealed scenes, camouflaged object detection, and numerous specialized domains, prompting a wave of adaptation techniques ranging from lightweight adapter layers and parameter-space reconstruction to knowledge distillation, in which a compact student model learns to mimic a large teacher. The Toronto study adds footwear science to the list of domains where modest fine-tuning unlocks outsized gains, and its finding that image quality variables materially shape performance echoes parallel observations in brain tissue segmentation and satellite imagery classification. The researchers acknowledge the inherent constraints of their dataset size—forty annotated outsoles leaves the classifier with limited statistical power—and the 70 percent classification accuracy should be read as an early benchmark rather than a deployable performance ceiling. Scaling the annotated corpus, standardizing image acquisition, and refining the contact-area features that bridge segmentation and prediction are the obvious next steps.

Funded by the Digital Research Alliance of Canada and reviewed by the University Health Network Research Ethics Board, the work forms part of a broader research program at KITE that spans mechanical and human-centered slip testing, gait analysis with multimodal transformers, and slip detection during real human walking trials, alongside earlier investigations of how the edge of a footwear sole influences measured slip resistance. Together these strands sketch a future in which the slip resistance of a shoe could be assessed rapidly, cheaply, and at scale—by photographing its sole, letting a fine-tuned foundation model trace its contact geometry, and letting a classifier render a verdict before the shoe ever meets an icy sidewalk. For the millions of people who navigate winter sidewalks, wet kitchens, oily factory floors, and slick hospital corridors every day, that future could translate into fewer falls, fewer fractures, and a far more transparent footwear market in which slip resistance is not a marketing slogan but a measurable, machine-verified property.

Subject of Research: Fine-tuning the Segment Anything Model for automated footwear outsole segmentation to predict slip resistance.

Article Title: Transfer Learning on Segment Anything Model for Footwear Outsole Segmentation to Predict Footwear Slip Resistance

Article References: Chavoshian, S., Khanghah, A. B., & Fekr, A. R. (2026). Transfer Learning on Segment Anything Model for Footwear Outsole Segmentation to Predict Footwear Slip Resistance. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04331-2

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04331-2

Keywords: Segment Anything Model, transfer learning, footwear outsole segmentation, slip resistance, machine learning, computer vision, fall prevention, biomedical engineering, slip and fall injuries, image segmentation, footwear safety, deep learning