The computational study used sports footage drawn from the International Olympic Committee’s official historical image library, an open sports-action video collection on Kaggle and YouTube channels associated with major sporting events. The source pool contained 123,900 candidate frames, including transitional and static frames. After resampling and removal of unnecessary frames, a high-density motion sequence was assembled for model training and evaluation. The 78 action subtypes included vault, floor exercise, rings and parallel bars in gymnastics; snatch, clean and jerk and their stages in weightlifting; different jumping postures and entry angles in diving; and starting, acceleration and mid-run phases in sprinting.

Seven annotators with computer-vision backgrounds manually outlined athletes frame by frame in LabelMe 4.6.0, using polygonal contours and an average of 2.37 minutes per frame. No pretrained automatic segmentation or annotation-assistance tools were used. Three independent consistency tests on a randomly selected sample of 1.47 thousand frames produced Fleiss’ kappa values of 0.87, 0.89 and 0.88, with a mean of 0.88. The reference boundary was the visible outer contour of the athlete’s overall body, including the trunk, limbs and head; internal areas hidden by occlusion or equipment were not inferred.

For preprocessing, non-constant-frame-rate streams were resampled to 29.97 frames per second. Segments shorter than 0.5 seconds and static sequences with pixel variance below 12.7 across five consecutive frames were removed. Images were resized to 513 by 513 pixels with bicubic interpolation, and central cropping retained 80% of the frame. Channel-wise Z-score standardization was applied, followed by augmentation consisting of horizontal flipping with 50% probability, rotation of plus or minus 15 degrees, brightness and contrast adjustment from 0.85 to 1.15, and Gaussian noise with a standard deviation of 0.01.

The model used an Xception-65 backbone with ImageNet-pretrained weights. Its encoder combined depthwise separable and dilated convolutions, while the decoder fused high-resolution shallow features with deeper semantic features. The atrous spatial pyramid pooling configuration used dilation rates including 3, 7 and 13 to capture context at different scales. A fully connected conditional random field then refined the initial mask through five mean-field iterations, using pixel-position and RGB-color similarities. The network itself performed binary foreground-versus-background segmentation with one output channel; the 78 action labels were used to characterize and evaluate the dataset rather than included in the loss function.

Training used the AdamW optimizer, an initial learning rate of 0.0031, cosine-annealing decay to 1 × 10−6, a batch size of 12 and 83 epochs. Experiments ran in TensorFlow 2.13.0 with the Keras application programming interface on an NVIDIA Tesla V100 GPU with 32 GB of memory. The optimized pipeline’s reported mIoU was 7.14 percentage points above the native DeepLabv3+ baseline, equivalent to an 8.80% relative increase. Forward inference time decreased from approximately 2.03 to 1.42 milliseconds per frame across 17,470 test frames, corresponding to a reported 1.43-fold speedup. This timing comparison was separate from the segmentation-accuracy measurement.

Performance varied by action category. The sprint start produced the highest reported category-level results: 89.31% IoU, 94.17% boundary accuracy and 92.87% recall. The authors associated this performance with the athlete’s extended posture, clear limb contours and strong contrast against the background. Weightlifting clean and jerk reached 88.73% IoU, 93.41% accuracy and 91.19% recall. Gymnastics vault reached 84.19% IoU, while diving take-off reached 83.07% IoU, 88.13% accuracy and 86.43% recall. The study attributed the lower gymnastics result to rapid limb folding during aerial rotation and the lower diving result to gray similarity around splashes and edge truncation at water entry.

Comparisons with other architectures highlighted differences in measured speed and accuracy. U-Net had a reported inference delay of 14.31 milliseconds but obtained 72.03% IoU for diving. Mask R-CNN reached 87.13% segmentation accuracy for sprinting and had a 48.73-millisecond delay. The conditional random field’s Gibbs energy decreased during its five refinement iterations, reaching reported values of 0.43 × 10^4 for gymnastics, 0.31 × 10^4 for weightlifting and 0.19 × 10^4 for sprinting. These findings describe performance on the study’s frames and experimental setup rather than real-time judging in competition settings.

The study identifies limitations in the breadth and balance of its dataset. Samples were unevenly distributed among action categories, and coverage of partial occlusion, poor lighting and intense reflections remained limited. Although results across training, validation and test splits were reported as relatively stable, the evidence does not establish that any particular optimizer or learning-rate schedule caused that stability. The findings support further evaluation of this boundary-extraction approach as a visual input for action recognition, posture analysis, motion tracking and related sports-analysis tasks, but the paper does not demonstrate clinical benefits, systematic real-time deployment or superior performance in a broad range of unseen sports.

Subject of Research: Pixel-level athlete boundary extraction from gymnastics, weightlifting, diving and sprinting video frames

Article Title: Extraction of sports action boundaries using the DeepLabv3+ algorithm

Article References: Guo, R. (2026). Extraction of sports action boundaries using the DeepLabv3+ algorithm. Discover Artificial Intelligence, 6(1), Article 1245. https://doi.org/10.1007/s44163-026-02257-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02257-1

Keywords: DeepLabv3+, sports video analysis, athlete segmentation, boundary extraction, computer vision, gymnastics, weightlifting, diving, sprinting, conditional random field, Extraction, sports

The motivation is straightforward. Histopathological examination of hematoxylin and eosin-stained slides remains the diagnostic gold standard for breast cancer, but it depends heavily on expert interpretation, and that interpretation is affected by workload, fatigue, and interobserver variability. As laboratories digitize their slides into whole-slide images, or WSIs, the opportunity for computational assistance grows, but so does the technical challenge. A single WSI is a gigapixel-scale image, far too large to feed directly into a neural network, so any practical system must break the slide into manageable pieces, analyze them, and then reassemble the results in a form a pathologist can inspect.

HistoBreastCAD addresses this in stages. First, a semantic segmentation model identifies tissue regions of interest, separating diagnostically relevant tissue from background. Second, a patch-based classifier assigns each analyzed region to one of four diagnostic categories: normal tissue, benign lesion, ductal carcinoma in situ, or invasive carcinoma. Third, the patch-level predictions are mapped back onto the original slide as color-coded probability overlays, so that model outputs can be reviewed in their spatial context. Finally, a web-based interface built on a React frontend and a FastAPI backend lets users pan and zoom across the gigapixel slide with OpenSeadragon, view the automated overlays, and draw their own manual annotations in a separate layer.

The segmentation module was trained and compared using three established architectures: U-Net with a ResNet-34 encoder, DeepLabV3 with a ResNet-50 backbone, and DeepLabV3+ with a ResNet-50 backbone. DeepLabV3+ combines atrous convolutions with atrous spatial pyramid pooling to capture contextual information at multiple receptive-field scales, while its decoder progressively recovers spatial detail. The models were trained on 256-by-256-pixel tiles extracted from ten annotated whole-slide images from the public BACH dataset, producing 248,501 tiles at native resolution and 16,421 tiles at the next downsampled pyramid level. Training used a combined Dice and focal loss with the AdamW optimizer, mixed precision, and early stopping.

Crucially, the evaluation was designed to prevent data leakage. Because multiple patches derive from the same source slide, the team grouped cross-validation folds by whole-slide image rather than by tile: in each of five folds, two complete slides were held out for validation while the remaining eight were used for training, and each slide appeared in validation exactly once. Under this protocol, the final DeepLabV3+ configuration at native resolution achieved an out-of-fold micro-Dice of 0.8390, an intersection-over-union of 0.7226, precision of 0.8278, and recall of 0.8505. DeepLabV3 performed closely behind, while U-Net showed markedly greater instability across folds, with fold-level micro-Dice standard deviations exceeding 0.28 at the downsampled level.

A subtle methodological finding emerged from the choice of checkpoint-selection metric. Tile-averaged Dice, the team found, can be misleadingly inflated by empty-background tiles that score perfectly while carrying no diagnostic information, potentially rewarding checkpoints with very low recall on actual tissue. When the researchers retrained DeepLabV3+ using pooled pixel-level micro-Dice directly for checkpoint selection, the aggregate Dice stayed essentially unchanged, but recall rose from 0.7828 to 0.8505 while precision fell from 0.9039 to 0.8278. In other words, the better metric did not manufacture a higher score; it shifted the operating point toward sensitivity and eliminated a silent failure mode.

For classification, the team evaluated four convolutional architectures on 400 labeled BACH microscopy images, 100 per diagnostic class, again with source-image-grouped five-fold cross-validation so that all 48 patches from a given source image stayed within the same partition. EfficientNetB0, which uses compound scaling of network depth, width, and input resolution, was the strongest individual model with a macro-F1 of 0.8142. The final system, however, uses a frozen equal-weight soft-voting ensemble of EfficientNetB0 and a ResNet18 variant with a global-average-pooling attention branch. The ensemble averages the two models’ softmax probability distributions with weights of 0.5 each, introducing no additional trainable parameters, and achieved an accuracy of 0.8283 and a macro-F1 of 0.8282, outperforming EfficientNetB0 alone in all five validation folds with a paired t-test p-value of 0.0134.

The most sobering results came from an exploratory reader study on the 100 microscopy images of the held-out BACH test set, which lack public reference labels. Two pathologists independently assigned four-class diagnoses, with 95 images receiving valid paired annotations. Interobserver agreement was moderate, with observed agreement of 66.32 percent and a Cohen’s kappa of 0.5105. Model-to-reader concordance was lower: EfficientNetB0 reached kappa values of 0.3242 against one reader and 0.2183 against the other, and the ensemble did not improve on this despite its benchmark advantage. Notably, the models assigned roughly 35 percent of the test images to the normal category, while each reader used that label for only 10 cases, a distributional mismatch that likely contributed to the disagreement.

The authors are careful about what these numbers mean. The cross-validated classification results quantify performance on held-out BACH source images, not end-to-end diagnostic accuracy of the complete pipeline, and the reader study measured concordance with individual expert opinions rather than correctness against an adjudicated reference standard. The system also relies on a single public benchmark, ten annotated slides for segmentation, and established CNN architectures rather than the large-scale pathology foundation models now emerging in the field. External multi-institutional validation, prospective multi-reader studies with consensus diagnoses, and formal explainability assessment are all listed as prerequisites before any claim of clinical generalizability.

Even with those caveats, the study offers a useful template for how computational pathology tools should be evaluated. By enforcing source-level partitioning, auditing tile metadata, confining augmentation and class balancing to training data, and separating benchmark performance from reader agreement, HistoBreastCAD demonstrates that an integrated segmentation-classification-visualization workflow is technically feasible with modest hardware, including consumer-grade GPUs, and that the discipline of leakage-safe evaluation can reveal exactly where a system’s strengths end and its open questions begin.

Subject of Research: Deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images

Article Title: HistoBreastCAD: an integrated deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images

Article References: Celi-Vega, O., Jiménez-Gaona, Y., Corral-Dominguez, P., Gutiérrez, Á., & Socola-Barzallo, Z. (2026). HistoBreastCAD: an integrated deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images. Frontiers in Bioinformatics, 6, Article 1870283. https://doi.org/10.3389/fbinf.2026.1870283

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1870283

Keywords: breast cancer, histopathology, deep learning, whole-slide imaging, semantic segmentation, computer-aided diagnosis, ensemble classification, EfficientNet, DeepLabV3+, BACH dataset, digital pathology, interobserver variability

The core insight behind the work is that most existing AI classifiers analyze the entire ultrasound image at once, forcing the network to learn from a mixture of lesion signals and irrelevant background tissue. That background bias inflates the risk of false negatives, the most dangerous kind of error in cancer screening because it means a malignant tumor is missed. The researchers instead built a segmentation-guided pipeline: stage one produces a pixel-level mask that isolates the suspicious region, and stage two classifies only that focused region of interest as normal, benign, or malignant. By feeding the classifier lesion-centric information rather than whole images, the framework concentrates its computational attention where it matters diagnostically.

In the segmentation stage, the team compared four architectures under identical training conditions: the classic U-Net, an attention-enhanced U-Net++, a standard DeepLabV3+ with a ResNet50 encoder, and their own modified DeepLabV3+. The modification is subtle but powerful. DeepLabV3+ already uses atrous, or dilated, convolution, which inserts gaps between filter elements so the network can perceive larger portions of the image without losing spatial resolution. Its Atrous Spatial Pyramid Pooling module applies several dilation rates, in this case 12, 24, and 36, so lesions of different sizes are captured at multiple scales. The researchers then inserted a Convolutional Block Attention Module, or CBAM, into the decoder, precisely where high-level semantic features are fused with low-level spatial detail.

CBAM works in two sequential steps. Channel attention determines which feature maps carry the most diagnostically meaningful information, using global average and max pooling followed by a lightweight multi-layer perceptron. Spatial attention then determines where in the image that information resides, applying a seven-by-seven convolution over pooled feature maps to highlight lesion-relevant regions and suppress background artifacts. The refined attention maps are multiplied back onto the original features, so the network effectively learns to look at the tumor and ignore the noise. The authors chose CBAM over transformer-based attention because it adds negligible cost: the segmentation model carries roughly 40 million parameters, of which the CBAM modules contribute only about 0.03 million.

The numbers from the comparison are striking. The proposed DeepLabV3+ with CBAM achieved a mean Dice score of 90.14 percent, a median Dice of 97.19 percent, and a mean Intersection-over-Union of 86.3 percent on the BUSI breast ultrasound dataset, outperforming U-Net, attention U-Net++, and standard DeepLabV3+. A Wilcoxon signed-rank test on 263 paired validation samples confirmed the improvement was statistically significant, with a p-value of 0.0225 for Dice and an extraordinarily small 6.84 times ten to the minus twentieth for IoU. The high median Dice is particularly encouraging, indicating that the model segments well not just on average but across the typical case, with low variability from image to image.

Stage two takes those generated masks and classifies them using a modified DenseNet121, a deliberately lightweight network of about 7.26 million parameters. DenseNet’s dense connectivity, in which each layer receives the outputs of all preceding layers within a block, encourages feature reuse and strong gradient flow, which suits the fine morphological detail of ultrasound. The team placed a CBAM module after each of the four dense blocks, at roughly 256, 512, 1024, and 1024 channels, creating a hierarchy of attention refinements from shallow texture cues to deep semantic patterns. Initialized with ImageNet-pretrained weights and fully fine-tuned, the classifier achieved F1-scores of 0.9945, 0.9873, and 0.9947 for normal, benign, and malignant cases respectively, with a best validation accuracy of 99.24 percent.

Most importantly for clinical screening, the model achieved 100 percent recall on the malignant class, meaning no cancerous lesion in the test set was misclassified. An ablation study confirmed that every component earns its place: removing CBAM from either stage degraded performance, and removing segmentation guidance entirely dropped accuracy to 91.50 percent and recall to 89.60 percent, a loss of nearly eight percentage points. That makes the segmentation-guided design the single largest contributor to the framework’s reliability, followed by attention in the classifier and then attention in the segmenter.

To test whether the system generalizes beyond a single dataset, the researchers ran cross-dataset validation on UDIAT, an independent ultrasound collection acquired in Sabadell, Spain, with a different scanner. Performance dipped as expected, with a mean Dice of 81.84 percent and classification accuracy of 93.18 percent, but remained strong given the shift in acquisition conditions and lesion characteristics. The authors are candid about limitations: both datasets are small, the external validation came from a single center with one scanner, and DeepLabV3+ demands more computation than simpler models. They plan to explore lighter backbones such as MobileNetV3, integrate explainable AI methods like Grad-CAM, and compare performance against radiologists on larger multi-center datasets.

The authors emphasize that in practice the framework would serve as an assistant, with the predicted masks and classifications reviewed by a radiologist rather than used for autonomous decisions. Even so, the study demonstrates a compelling recipe for medical AI: localize first, classify second, and use cheap, targeted attention rather than heavyweight architectures to do it. With false negatives being the costliest error in cancer screening, a pipeline that cuts them while adding almost no computational overhead could bring segmentation-guided diagnosis closer to real clinical workflows.

Subject of Research: A two-stage attention-enhanced deep learning framework for segmentation-guided classification of breast cancer in ultrasound images

Article Title: Attention enhanced DeepLabV3+ framework for breast cancer classification

Article References: Singh, P., Kumar, S., & Gupta, M. (2026). Attention enhanced DeepLabV3+ framework for breast cancer classification. Discover Artificial Intelligence, 6(1), Article 1436. https://doi.org/10.1007/s44163-026-02435-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02435-1

Keywords: breast cancer, deep learning, ultrasound imaging, DeepLabV3+, CBAM, DenseNet121, medical image segmentation, computer-aided diagnosis, attention mechanism, BUSI dataset, false negatives, machine learning

The research team built a two-stage architecture that mirrors how a forensic expert actually works. The first stage performs segmentation: a DeepLabv3+ network, a convolutional model widely used for dense pixel-level prediction, was trained to delineate six distinct injury regions on head CT scans. DeepLabv3+ is particularly suited to this task because of its atrous spatial pyramid pooling module, or ASPP, which applies convolutions at multiple dilation rates in parallel. This allows the network to capture both fine local detail, such as the edges of a small contusion, and broader contextual information, such as the spatial relationship between a hemorrhage and the surrounding skull and brain tissue. In this study, the segmentation network achieved a mean Dice coefficient of 0.87, a standard overlap metric indicating substantial agreement between the model’s injury maps and expert annotations.

The second stage of the cascade handles classification. A ResNet18 backbone, a relatively lightweight residual convolutional network, was combined with the same ASPP module to categorize each case as either an acceleration injury or a deceleration injury. Crucially, the classifier did not operate on raw images alone. Instead, it received segmentation-guided information from the first stage, meaning the features extracted by the segmentation network were fed forward to inform the classification decision. This design choice proved decisive. In an ablation study, in which components of the system are systematically removed to measure their contribution, the macro-F1 score, a balanced measure of precision and recall across classes, improved from 0.71 to 0.82 when segmentation-guided information was included. The cascaded model as a whole attained a mean area under the receiver operating characteristic curve, or AUC, of 0.94 for injury mechanism classification, a figure that places its discriminative power in the range typically associated with expert-level medical imaging systems.

The physics underlying the classification problem explains why it is so difficult. In an acceleration injury, the head is stationary and is set into motion by an impacting force, as when a fist or a blunt object strikes a stationary skull. In a deceleration injury, the head is already moving and is abruptly stopped, as when a falling person’s head strikes the ground or a vehicle occupant’s head hits the windshield. Both scenarios produce contusions, hemorrhages, and fractures, but the distributions of these lesions differ in subtle ways rooted in inertial biomechanics. Coup injuries, arising at the site of impact, and contrecoup injuries, arising on the opposite side of the brain as it rebounds against the inner table of the skull, follow patterns that depend on the direction and magnitude of the forces involved. Distinguishing these patterns reliably has traditionally required expert judgment, and even experts can disagree when the injury presentation is ambiguous.

What makes the new system scientifically interesting is not merely its accuracy but the way its errors and successes align with established biomechanical principles. The authors validated the model externally across frontal, temporal, and occipital regions of the head, and found that its performance patterns were consistent with what biomechanics would predict. This kind of consistency check matters because deep learning models are notorious for exploiting spurious correlations, learning to recognize hospital-specific imaging artifacts or demographic confounds rather than genuine pathology. A model whose regional performance tracks known mechanisms of force transmission is more likely to have learned something real about injury physics, which is essential if the tool is ever to carry weight in a courtroom.

The study also demonstrated that the system could reliably distinguish injured brains from normal images, a prerequisite for any forensic screening application. This binary detection task may seem trivial compared with mechanism classification, but in forensic practice it is far from it. Postmortem changes, varying image quality, and the enormous anatomical variability of human brains all complicate automated analysis. The fact that the cascaded architecture maintained high accuracy on this task while simultaneously performing the harder mechanism classification suggests that the segmentation stage is doing genuine work, localizing pathology in a way that generalizes across cases rather than memorizing individual examples.

The work sits within a rapidly expanding literature on artificial intelligence in forensic imaging. Previous studies have used deep learning to detect skull fractures from curved maximum intensity projections, to identify critical findings in head CT scans, and to classify fractures on the basis of their location and number in order to discriminate between falls and blows. Each of these efforts has attacked a piece of the forensic puzzle. The present study is notable for targeting the mechanism question itself, which is arguably the most legally sensitive piece of all. It also joins a growing family of cascaded and joint segmentation-classification architectures, in which one network’s output refines another’s input, a strategy that has proven effective in medical imaging tasks ranging from cell classification to cancer diagnosis.

The practical implications for forensic practice could be substantial. Determining the manner of injury currently depends on the availability of experienced examiners, and in jurisdictions with heavy caseloads or limited forensic resources, such expertise may be scarce. An automated system that provides an objective, quantifiable assessment of injury mechanism could serve as a second opinion, flagging cases where the imaging evidence contradicts the stated history of events. The authors emphasize that the system offers a route toward objective injury mechanism determination, a phrase that carries particular weight in a field where subjective judgment has historically been the norm. Because all data underlying the findings are freely accessible to other researchers, independent groups can scrutinize and extend the work, which is exactly the kind of transparency forensic science needs.

There are, of course, caveats that temper enthusiasm. The model was trained and validated on data from a single institution, the Academy of Forensic Science in Shanghai, and although external validation across anatomical regions was performed, broader multi-center validation across different scanners, populations, and injury contexts will be needed before deployment. Legal medicine also demands more than a probability score; courts require explainable reasoning, and a 0.94 AUC does not by itself explain why a particular injury pattern indicates deceleration rather than acceleration. Future work will likely need to couple such models with interpretability techniques and with the biomechanical modeling traditions, including finite element analysis, that have long informed the field. The study was approved by the Scientific and Ethical Committee of the Academy of Forensic Science, conducted with informed consent from the families involved, and supported by Chinese national and municipal research programs, reflecting the institutional investment now flowing into computational forensics.

Nevertheless, the study marks a genuine step forward in the marriage of artificial intelligence and legal medicine. By combining a segmentation network that sees where the injury is with a classification network that reasons about how it happened, Liu and colleagues have built a system that approximates, in silicon, the diagnostic chain that forensic experts perform in their heads. The quantitative gains, from a macro-F1 of 0.71 to 0.82 with segmentation guidance, and a mean AUC of 0.94 overall, are not just incremental numbers; they demonstrate that structural information about injury location and extent is the key to unlocking mechanism inference. As such cascaded architectures mature and accumulate validation across centers and populations, the prospect of a forensic imaging tool that can tell investigators not only what happened to a brain, but how it happened, moves from aspiration toward reality, with implications for accident reconstruction, criminal investigation, and the pursuit of justice in cases where the only witness is the injured tissue itself.

Subject of Research: Deep learning classification of acceleration and deceleration craniocerebral injuries for forensic injury mechanism determination

Article Title: Classification of acceleration and deceleration craniocerebral injuries by using cascaded deep learning models

Article References: Liu, Y.-W., Tian, Z.-L., Liu, Y.-Y., Fu, E.-H., Dong, H.-W., Wan, L., Zhang, J.-H., Zou, D.-H., & Liu, N.-G. (2026). Classification of acceleration and deceleration craniocerebral injuries by using cascaded deep learning models. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-03971-2

Image Credits: AI Generated

DOI: 10.1007/s00414-026-03971-2

Keywords: craniocerebral injury, forensic medicine, deep learning, DeepLabv3+, ResNet18, image segmentation, computed tomography, injury mechanism, acceleration injury, deceleration injury, biomechanics, legal medicine

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