The stakes are high. Nearly 80 percent of Fennoscandian forests are intensively managed, and Norway spruce is among the most productive industrial species in the region. Warming temperatures and rising carbon dioxide could, in principle, boost spruce growth in the coming decades, but only if water does not become limiting. Climate projections point in the opposite direction, forecasting more frequent droughts across the boreal zone. Norway spruce is an isohydric species, meaning it closes its stomata in response to rising vapor pressure deficit, the drying power of the air, in order to preserve leaf water potential. That strategy protects the tree’s hydraulics but comes at the cost of reduced carbon uptake, and the species’ shallow roots and generally low drought resilience on mineral soils have long worried foresters. Far less has been known about how spruce fares on peat soil, where water tables are managed by drainage ditches.

The study site offered a natural experiment. In March 2021, the stand was divided into two blocks: an untreated control and a block that was selectively harvested according to the principles of continuous cover forestry, a management approach gaining traction in the boreal region. The harvest removed 55 percent of the trees taller than 15 meters, cutting projected canopy cover from 60 percent to 22 percent and reducing leaf-area index from roughly 3 to 2 square meters per square meter. The researchers hypothesized that the thinned block, with less inter-tree competition for soil water, would weather drought better than the dense control stand, and that is largely what the data showed, though with instructive complications.

Meteorologically, the two years could hardly have been more different. The summer of 2020 was warm but drought-free by the Standardized Precipitation-Evapotranspiration Index, while June and July 2021 brought a severe drought, with monthly SPEI values as low as minus 1.87, a rainless spell lasting until late July, a steep drop in the water table, and surface soil moisture falling below the 0.2 cubic meters per cubic meter threshold the team used to define soil drought. Crucially, atmospheric drought, defined as a daily mean vapor pressure deficit above 1 kilopascal, was frequent in both years, and the hottest days saw median daytime temperatures approaching 30 degrees Celsius. The forest was thus exposed to the full spectrum of atmospheric and soil drought, separately and in combination.

The eddy-covariance tower, standing 29 meters tall on the boundary between the two blocks, revealed the ecosystem-scale signature of that stress. Net ecosystem exchange peaked at a vapor pressure deficit of about 1 kilopascal and fell toward zero as the air dried further, reaching essentially zero exchange at roughly 2 kilopascals in both blocks and both years. Light-saturated photosynthesis, ecosystem respiration, and light-use efficiency all dropped markedly in 2021 compared with 2020, and the Bowen ratio, the ratio of sensible to latent heat flux, roughly doubled during the dry summer, a classic indicator that the forest was diverting energy away from evapotranspiration. Yet the picture was not uniformly grim: drought days are typically sunnier, and the extra sunlight at least partly compensated for the reduced photosynthetic efficiency, keeping daily carbon gains comparable to those of milder days.

The most direct evidence of tree-level stress came from the sap-flow sensors installed on 16 spruce trees, eight per block. In the control stand, sap flow rose linearly with vapor pressure deficit up to about 1 kilopascal and then saturated, a sign that stomata were constricting. More striking was a progressive afternoon dip in control-tree sap flow that deepened from June through July 2021, coinciding with the daily peak in atmospheric demand, and then largely vanished in August when rains returned. Most harvest-block trees, by contrast, maintained higher sap flow across the entire vapor pressure deficit range, consistent with the idea that fewer trees sharing the same soil water reservoir face less competition. When the team upscaled tree-level sap flow to the stand using tree height relationships derived from drone imagery, the thinned block transpired roughly half as much water as the control, echoing earlier findings that thinning raises per-tree water use while lowering stand-level totals.

The drone surveys added a spatial dimension that ground sensors cannot provide. Flying a DJI Matrice 210 with thermal and multispectral cameras on four clear days, the team produced orthomosaics of canopy temperature and the Normalized Difference Vegetation Index, then segmented individual tree crowns and corrected temperatures for differences in canopy brightness. Under non-drought conditions, the two blocks were nearly indistinguishable. During peak drought, however, trees in the harvested block tended to run hotter and show lower NDVI than control trees, implying greater susceptibility to extreme drought in the open stand, possibly because thinned crowns receive more solar radiation. Rows of stressed trees also appeared along forest tracks and ditches, a spatial pattern invisible to the tower. A novel touch was the use of wet and dry reference piles of spruce branches, sprayed with the equivalent of 15 millimeters of rainfall, to convert canopy temperatures into estimates of stomatal conductance for every mapped tree.

Perhaps the study’s most valuable contribution is its honest cross-examination of the monitoring tools themselves. The sap-flow data flagged stress in the denser control stand, exactly where competition for water was fiercest, while the drone data highlighted stress in the thinned stand, where exposed crowns heated up. The eddy-covariance system, which averages fluxes over hectares, smoothed over much of this tree-to-tree variation and could not reliably detect a minority of vulnerable individuals. Each technique, the authors conclude, occupies a different niche: sap flow offers continuity and directness but covers only a handful of trees; drones offer stand-wide spatial coverage but only snapshot temporal resolution; eddy covariance offers continuous ecosystem totals but little spatial discrimination. No single method currently combines directness, continuity, coverage, and low uncertainty.

For forest management, the practical implications are tangible. The evidence that selection-harvested trees maintained higher photosynthetic uptake, sap flow, and stem growth during drought supports continuous cover forestry as a strategy for buffering spruce stands against the hotter, drier summers projected for the boreal zone. The authors also propose a deceptively simple early-warning tactic: equip the least resilient trees in a stand, those most disadvantaged by competition or exposure, with sap-flow and stem-diameter sensors, and let them act as sentinels for the onset of drought stress. As boreal summers like that of 2021 become the norm rather than the exception, such sentinel trees, watched from the ground and the air alike, may become an essential part of keeping northern forests productive and climate-friendly.

Subject of Research: Drought ecophysiology of a managed Norway spruce forest on drained boreal peatland

Article Title: Drought responses of a Norway spruce forest on drained peat soil: combining sap-flow sensors, eddy-covariance, meteorological, soil and UAV data

Article References: Alekseychik, P. K., Peltoniemi, M., Mäkipää, R., Tuominen, V., Laurila, T., Jones, H., Müller, M., Lopatin, E., Rautakoski, H., Vesala, T., & Launiainen, S. (2026). Drought responses of a Norway spruce forest on drained peat soil: combining sap-flow sensors, eddy-covariance, meteorological, soil and UAV data. Biogeosciences, 23(19), 7091-7127. https://doi.org/10.5194/bg-23-7091-2026

Image Credits: AI Generated

DOI: 10.5194/bg-23-7091-2026

Keywords: Norway spruce, drought, boreal forest, peatland, sap flow, eddy covariance, UAV remote sensing, continuous cover forestry, vapor pressure deficit, evapotranspiration, NDVI, forest management

The research, led by Matthew J. Drouillard and Anthony R. Cummings together with colleagues including Persaud Moses, Fabian Moses, and Catherine Auerbach, focused on a localized region of swidden agriculture, the traditional farming system in which small plots of forest are cut and burned, cultivated for a few seasons, and then left to regenerate as farmers move on to new ground. Rather than relying on ground surveys alone, which are slow and laborious in dense tropical terrain, the team turned to unmanned aerial vehicles, or UAVs, equipped to capture imagery of extraordinary detail. Over 255 hectares of both undisturbed forest and land disturbed by swidden activity, the researchers deployed convolutional neural networks, a form of artificial intelligence that excels at recognizing objects in images, to detect individual palms from the air.

The scale of the resulting census is remarkable. In total, the team identified 10,194 individual palms belonging to six different species. Each palm was mapped within its landscape context, allowing the researchers to compare how palm communities differed between forest that had never been cleared and forest recovering from swidden disturbance. The numbers told a clear story. Land disturbed by swidden agriculture supported a palm density of 53.3 individuals per hectare, compared with just 22.6 per hectare in undisturbed forest, an increase of 136 percent. In other words, disturbed plots held more than twice as many palms as intact forest, a result that runs directly counter to the expectation that human disturbance uniformly reduces the abundance of valued forest resources.

But abundance, the study shows, is only half the picture. When the researchers examined which species made up those palm populations, they found that swidden disturbance had dramatically simplified the palm assemblage. Using Local Moran’s I, a spatial statistics technique that identifies clusters of similar values across a landscape, the team determined that a single species, Attalea maripa, dominated both forest types, but to very different degrees. In undisturbed forest, A. maripa accounted for 77.1 percent of all palms identified. On swidden-disturbed land, its share rose to 89.0 percent. The disturbed landscape was not only richer in palms; it was also far more of a monoculture, overwhelmingly populated by one hardy, disturbance-tolerant species.

The species that lost ground in disturbed areas tell an equally important story. Astrocaryum vulgare, the second most common palm in the study area, declined in relative representation from 17.3 percent of palms in undisturbed forest to just 10.0 percent in disturbed forest. Other, less abundant palm species also saw their shares shrink. This pattern, in which total numbers rise while diversity falls, represents a genuine ecological trade-off. Swidden agriculture, the authors conclude, can increase total palm density while simultaneously simplifying the palm community and eroding the relative representation of rarer species. For Indigenous communities whose livelihoods and cultural practices depend on a variety of palm products, from food and construction materials to crafts and medicine, that erosion of diversity could carry consequences that raw density figures conceal.

The technical achievement behind these findings deserves attention in its own right. Counting individual trees across hundreds of hectares of tropical forest has always been one of ecology’s most stubborn challenges. Ground-based plots are accurate but tiny, while satellite imagery has historically been too coarse to distinguish individual crowns in closed-canopy forest. The convolutional neural network approach used here changes that calculus. CNNs are trained on labeled examples of the objects they must find, learning to recognize the distinctive spectral and textural signatures of palm crowns in aerial imagery. Once trained, they can scan vast image mosaics in a fraction of the time a human analyst would need, with consistent criteria applied across the entire dataset.

The study also provides a sobering lesson about the limits of satellite remote sensing for this kind of fine-scale ecological work. Alongside the UAV surveys, the researchers analyzed multispectral imagery from the WorldView-2 satellite, which offers a resolution of 0.5 meters, among the sharpest commercially available. Yet even at that resolution, the satellite imagery detected approximately 70 percent fewer palms than the drone-based approach. The discrepancy underscores how much ecological detail remains invisible to even advanced orbital sensors, and it suggests that studies relying on satellite data alone may substantially underestimate the abundance of individual trees, particularly in landscapes where understory and mid-story vegetation obscure the ground from above.

For the Wai Wai and other Indigenous peoples of the region, whose harvesting practices have helped shape these palm populations over generations, the findings carry practical weight. Swidden agriculture is often portrayed in policy debates as a driver of deforestation that should be curtailed or replaced with intensive permanent cultivation. This study complicates that narrative. If disturbed plots support more than double the palm density of intact forest, then the traditional cycle of clearing, cultivation, and fallow may function as a form of resource enhancement, at least for the dominant species. At the same time, the loss of relative diversity in disturbed areas suggests that unmanaged expansion of swidden activity could gradually homogenize palm communities, with uncertain consequences for the animals that depend on less common palm species and for the resilience of the ecosystem as a whole.

The authors frame their results as a contribution to what they call more sustainable land-use pathways, approaches that maintain the livelihood and cultural functions of swidden agriculture while conserving palm-community diversity. Achieving that balance will require understanding not just how many palms a landscape supports, but which species, where they cluster, and how their distributions shift across the disturbance gradient from mature forest to active gardens to old fallows. The spatial clustering analysis used in the study offers a template for this kind of work, showing exactly where dominant species concentrate and where rarer palms persist or disappear.

As tropical forests face mounting pressure from commercial agriculture, logging, and climate change, studies like this one highlight the value of combining Indigenous knowledge with modern technology. The 10,194 palms mapped across 255 hectares of Guyana are more than a dataset; they are evidence that human disturbance and ecological abundance are not always opposites, and that the future of tropical biodiversity may depend on understanding the nuanced trade-offs embedded in traditional land-use systems. The challenge for conservation, the study suggests, is not to exclude people from the forest, but to sustain the practices that enrich it while safeguarding the diversity that makes it whole.

Subject of Research: The effects of Indigenous swidden agriculture on palm population density and diversity in South-Central Guyana

Article Title: Impact of swidden agriculture on palm populations in South-Central Guyana

Article References: Drouillard, M. J., Cummings, A. R., Moses, P., Moses, F., & Auerbach, C. (2026). Impact of swidden agriculture on palm populations in South-Central Guyana. PLOS Sustainability and Transformation, 5(9), e0000275. https://doi.org/10.1371/journal.pstr.0000275

Image Credits: AI Generated

DOI: 10.1371/journal.pstr.0000275

Keywords: swidden agriculture, palms, Guyana, UAV remote sensing, convolutional neural networks, Attalea maripa, Astrocaryum vulgare, biodiversity, Indigenous land use, tropical forests, WorldView-2, spatial analysis

The stakes are considerable. Crop pathogens, pests, and environmental stressors cut annual yields by 21 to 40 percent in major staple crops, with food-deficit regions absorbing a disproportionate share of the loss and a global economic impact estimated above 220 billion US dollars per year. Ground scouting is labor-intensive, covers only a fraction of the production area, and offers no warning before stress symptoms become visible. Reflectance shifts in the 680 to 730 nanometer Red Edge band, by contrast, precede measurable chlorophyll loss, and the normalized difference red edge index has been shown to flag stressed vegetation 13 to 16 days ahead of broadband indices. Capturing those signals routinely, however, has remained out of reach for most of the world’s growers.

The research team, led by Laurensia Simanihuruk of Sepuluh Nopember Institute of Technology together with collaborators in Indonesia and Malaysia, validated their framework on a demanding real-world testbed: ten heterogeneous cocoa agroforestry typologies in Divo, Côte d’Ivoire. The underlying dataset, collected with a DJI Phantom 4 Multispectral flying at 80 meters altitude, comprises 1,272 co-registered RGB-multispectral image pairs spanning 7,632 individual frames. Multi-storey canopies, intercropped shade trees, senescent foliage, and intricate shadow patterns create exactly the kind of spectral complexity that breaks the assumptions of earlier reconstruction methods, which were mostly tuned on laboratory benchmarks with uniform illumination and homogeneous materials.

The architecture’s central insight is biophysical rather than purely computational. Vegetation reflectance follows a causal sequence: pigment absorption shapes the visible bands, the chlorophyll transition defines the Red Edge response around 730 nanometers, and cellular scattering together with water content drives the near-infrared plateau at 840 nanometers. On the Ivorian dataset, the correlation between the RGB input and the near-infrared target is a weak 0.43, while the correlation between Red Edge and near-infrared reaches 0.76, the strongest cross-band relationship in the data. A network that maps RGB directly to near-infrared therefore discards the very band that predicts its target best. TC-SA-UResNet instead routes reconstruction through a Triple Cascaded Decoder: one stream regenerates the visible bands, a second produces the Red Edge band, and a third generates near-infrared using features transferred from the Red Edge decoder through a Learnable Feature Injection module whose transfer gates start at zero and open only when they reduce the loss.

Attention mechanisms refine the cascade at two scales. A Spectral Attention module based on cross-attention lets each location in the near-infrared decoder read every location of the Red Edge decoder, which matters where a shadowed cocoa crown carries weak local evidence. Inside each decoder block, a Channel-wise Dual Attention module weights both channels and spatial locations, while a Global-Local Channel Attention module applies a separate gate per channel at each pixel, allowing fine-grained spectral discrimination. At the bottleneck, an Atrous Spatial Pyramid Pooling module captures context at dilation rates of 6, 12, and 18 pixels, and a Squeeze-and-Excitation block recalibrates the 256 feature channels before decoding begins. A ResNet-50 encoder pretrained on ImageNet supplies the hierarchical features shared by all three decoder streams.

Perhaps the most distinctive contribution is the training objective. Standard pixel losses and structural similarity terms say nothing about whether a reconstructed spectrum is physically plausible for vegetation, and a network can record low error while producing an impossible vegetation index. The authors’ Biophysical Spectral Consistency Loss adds four constraints: it matches the Pearson correlation between reconstructed Red Edge and near-infrared to that of the reference pair, compares the reconstructed NDVI field against the reference, preserves the shape of the spectral curve by matching reflectance steps between adjacent bands, and penalizes any NDVI value straying outside the physical interval from minus one to one. Correlation-guided weighting allocates the largest loss share, roughly 44 percent, to the near-infrared decoder, the band hardest to infer from visible light.

The evaluation was designed to prevent the spatial leakage that plagues many UAV studies. Rather than randomly shuffling frames, which risks near-duplicate images appearing in both training and test sets, the team split the data by plot: six plots for training, two for validation, and two entirely unseen plots for testing, ensuring the model faced canopy structures and illumination conditions it had never encountered. Against three baselines re-implemented from their original publications, a single-decoder U-Net, a dense prediction framework from Zhao and colleagues, and a two-step generative adversarial network, the proposed method achieved the highest structural similarity on the blue and near-infrared bands, the highest near-infrared coefficient of determination at 0.7898, and the smallest spectral angle of any architecture at 3.58 degrees. All 45 pairwise comparisons across bands and baselines reached statistical significance at p below 0.0001.

The ablation study reveals how each component earns its place. Swapping the vanilla encoder for ResNet-50 raised near-infrared R-squared from 0.7154 to 0.7725, and adding the biophysical constraints deliberately reshaped the optimization landscape before the ASPP-SE bottleneck and the full cascade recovered and surpassed all intermediate configurations. The authors are candid about trade-offs: near-infrared structural similarity of 0.7180 trails the blue band’s 0.9540, because near-infrared reflectance depends on internal leaf mesophyll structure that an RGB sensor simply cannot see, bounding the texture the network can recover. Yet the relative error tells a fairer story, with near-infrared at 13.7 percent against 13.4 percent for blue once each band’s error is normalized by its mean reflectance. Since vegetation indices are ratios of band differences to band sums, they depend on magnitude rather than texture, which explains why reconstructed NDVI reached a correlation of 0.9068 with the reference despite the lower structural fidelity.

The agronomic bottom line is nuanced but promising. Reconstructed NDVI achieved a mean absolute error of 0.0342, about 31 percent of one standard deviation of the measured quantity, with every reconstructed pixel falling inside the physical range, supporting vigour zoning within a flight, block ranking on a given date, and directional change detection across dates. The chlorophyll-sensitive indices NDRE and LCI fared less well, with R-squared values of 0.5379 and 0.5873, because both combine the two hardest bands and their uncertainties compound; they remain usable as screening layers to direct ground scouts toward weak zones but cannot support absolute nitrogen inference. The authors caution that the model was trained on one sensor, one crop, and a narrow clear-sky acquisition window, so transfer to other crops or imaging systems will require domain adaptation with paired observations rather than zero-shot deployment. Even with those caveats, the study marks a striking demonstration that the invisible half of the spectrum can be conjured, with measurable fidelity, from the cheap visible light every farmer’s drone already captures.

Subject of Research: Deep learning reconstruction of multispectral bands from RGB UAV imagery for cocoa agroforestry monitoring

Article Title: Triple Cascaded Spectral Attention U-ResNet for UAV RGB-to-multispectral reconstruction in cocoa agroforestry

Article References: Simanihuruk, L., Sarno, R., Sungkono, K. R., Putri, R. A., Anggraini, R. N. E., Wiratmoko, D., Larekeng, S. H., & Hitam, M. S. (2026). Triple Cascaded Spectral Attention U-ResNet for UAV RGB-to-multispectral reconstruction in cocoa agroforestry. Journal of Agriculture and Food Research, 31, Article 103333. https://doi.org/10.1016/j.jafr.2026.103333

Image Credits: AI Generated

DOI: 10.1016/j.jafr.2026.103333

Keywords: UAV remote sensing, multispectral reconstruction, deep learning, cocoa agroforestry, vegetation indices, NDVI, near-infrared, precision agriculture, smallholder farming, spectral attention, Côte d'Ivoire, crop stress detection

The research team, led by Alireza Sanaeifar of California State University together with Eve Laroche-Pinel, virologist Marc Fuchs, and Luca Brillante, worked across four commercial vineyards in Napa Valley, California, growing Cabernet Sauvignon and Cabernet Franc. Over two growing seasons, they geolocated 714 vines with satellite navigation, collected petiole samples from each, and confirmed infection status using endpoint multiplex PCR. The resulting dataset contained 399 infected and 315 non-infected vines, an approximately balanced distribution that strengthens the reliability of the classification results.

To capture the spectral fingerprints of infection, the team flew a DJI Matrice 600 Pro drone carrying a Senop HSC-2 snapshot hyperspectral camera with 29 spectral bands spanning 520 to 820 nanometers. Flights at 30 meters altitude and 5 meters per second produced imagery with a ground resolution of 2 by 2 centimeters per pixel. Because the camera records its bands sequentially, the researchers orthorectified each band individually and verified that residual misregistration stayed below roughly 5 centimeters. Reflectance was calibrated using a white reference panel placed in the field, and orthomosaics built in Agisoft Metashape allowed each PCR-tested vine to be matched precisely to its canopy pixels.

The spectral analysis revealed how subtle the disease signal really is. Infected and healthy vines showed nearly identical overall reflectance curves, with statistically significant differences concentrated in the green region between 540 and 580 nanometers, where healthy vines reflected more light, likely because infection degrades chlorophyll and alters pigment composition. Infected vines also showed localized reflectance increases around 740 to 760 nanometers, hinting at changes in internal leaf structure. Notably, the widely used NDVI vegetation index failed to separate the classes, while the Green NDVI and the Anthocyanin Reflectance Index both discriminated strongly, the latter consistent with the anthocyanin accumulation that gives red blotch disease its name.

Conventional vegetation indices, however, proved too blunt for reliable diagnosis, which is where the deep learning model enters. The researchers designed CompactSpectralViT, a lightweight patch-based vision transformer tailored to hyperspectral data. Each vine’s variable-sized canopy cube was standardized to 64 by 64 pixels across 29 bands through a three-stage pipeline of proportional resizing, adaptive center cropping, and final adjustment. The cube was then divided into 64 patches of 8 by 8 pixels, each carrying all 29 spectral values, and passed through a two-stage embedding that compressed each patch to just 48 dimensions. Four transformer encoder blocks with three attention heads each then let the model weigh relationships across the entire canopy simultaneously, capturing long-range spectral-spatial dependencies that convolutional networks process only locally.

Evaluated with 10-fold stratified cross-validation, CompactSpectralViT achieved a mean accuracy of 75.3 percent, precision of 74.5 percent, recall of 87.2 percent, an F1-score of 79.8 percent, and a ROC-AUC of 0.744. The high recall is operationally significant: in disease monitoring, a missed infected vine can seed further spread by the three-cornered alfalfa hopper, the virus’s insect vector, whereas a false alarm costs comparatively little. By tuning the classification threshold per fold, the model shifted its decision boundary toward sensitivity, accepting more false positives in exchange for catching nearly nine in ten infected vines.

The compact architecture also delivered a striking computational advantage. Benchmarking on an NVIDIA RTX 3090 showed that CompactSpectralViT required only 16.6 million multiply-accumulate operations per inference, roughly 7.6 times fewer than a convolutional baseline and a hybrid CNN-transformer comparator, and used just 138 megabytes of GPU memory versus 376 and 440 megabytes for the alternatives. It also outperformed both deep learning baselines in accuracy, recall, and ROC-AUC, suggesting that direct patch-level tokenization lets self-attention operate across the whole canopy from the first layer rather than relying on late-stage reasoning bolted onto convolutional features.

Interpretability analysis added a second layer of insight. SHAP-based feature importance and permutation importance, two independent attribution methods, agreed on nine of their ten most informative wavelengths. When the model was retrained on only the ten SHAP-selected bands, accuracy held at 75.6 percent, essentially matching the full 29-band system. The selected wavelengths clustered in three physiologically meaningful regions: the green band from 520 to 570 nanometers, tied to chlorophyll stress; the orange-red region from 610 to 620 nanometers, linked to anthocyanin accumulation; and the near-infrared from 780 to 820 nanometers, sensitive to leaf structure and water content. The authors caution that these findings guide sensor design but do not prove that a true multispectral camera would perform identically, since bandwidth and noise characteristics differ.

Against traditional machine learning on the same ten-wavelength input, the gap was decisive. A support vector machine reached only 62.9 percent accuracy and a random forest 65.1 percent, compared with 75.6 percent for the transformer, which also showed lower variability across folds. Part of this advantage reflects input representation, since the classical models received flattened feature vectors that discard spatial structure, but the result still underscores that self-attention over spectral-spatial patches extracts disease signatures that conventional classifiers miss. The accuracy itself is comparable to earlier ground-based hyperspectral studies of the virus, but the new work achieves it while imaging entire vineyard blocks by drone rather than one vine at a time, and across four sites and two cultivars.

The study’s authors are candid about limitations. Because sites and seasons were pooled across validation folds, transferability to unseen vineyards remains untested, and the infected class mixed symptomatic and asymptomatic vines, so the reported accuracy reflects detection under mixed symptom expression rather than purely asymptomatic diagnosis. Still, the work demonstrates that lightweight transformer architectures can perform credibly with limited labeled data under real field conditions, and it charts a practical path forward: automated vine segmentation, longitudinal monitoring of individual vines, integration with precision roguing programs, and eventually multi-disease frameworks that distinguish viral, fungal, and abiotic stresses from the air. For an industry where infection costs can reach tens of thousands of dollars per hectare over a vineyard’s lifetime, a drone and a compact neural network may soon become standard tools in the fight against an invisible pathogen.

Subject of Research: UAV hyperspectral imaging and a lightweight vision transformer for detecting grapevine red blotch virus in vineyards

Article Title: CompactSpectralViT: A lightweight vision transformer for grapevine red blotch virus detection from UAV hyperspectral imagery

Article References: Sanaeifar, A., Laroche-Pinel, E., Fuchs, M., & Brillante, L. (2026). CompactSpectralViT: A lightweight vision transformer for grapevine red blotch virus detection from UAV hyperspectral imagery. Smart Agricultural Technology, 15, Article 102599. https://doi.org/10.1016/j.atech.2026.102599

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102599

Keywords: grapevine red blotch virus, hyperspectral imaging, vision transformer, UAV remote sensing, precision viticulture, deep learning, plant disease detection, Napa Valley, PCR validation, feature importance, SHAP, multispectral sensors

The software, called SIFMap, was developed by J. Buffat, A. Elibol, H. Scharr, S. Choza-Farías, S. Salattna and J. Bendig and described in the journal SoftwareX. It is a Python toolbox purpose-built to process data from a novel high-resolution imaging sensor known as SIFcam, which is carried aboard uncrewed aerial vehicles. The package automates the entire journey from raw image pairs captured in flight to finished, stitched maps of fluorescence at 760 nanometers, a wavelength where the fluorescence signal can be cleanly separated from reflected sunlight. The code is freely available under the GPL 3.0 license, with documentation, example datasets and two complete flight datasets published alongside the paper.

The trick behind measuring SIF from the air lies in a clever piece of physics called Fraunhofer Line Discrimination, or FLD. Deep in the solar spectrum there are narrow dark bands, Fraunhofer lines, where absorption in the sun’s own atmosphere dims incoming sunlight. Within one of these dark lines, the light reaching a plant is weak, so any fluorescence the plant emits stands out clearly against it. SIFcam exploits this by carrying two cameras fitted with ultra-narrowband interference filters centered at 757.9 and 760.7 nanometers, straddling the oxygen absorption feature in the far red. By comparing the two channels, the software can compute fluorescence using the most basic formulation of the FLD method, combining the reflectance in both bands with an estimated irradiance to isolate the faint fluorescent flux.

What makes SIFcam unusual is that it behaves like a multispectral camera array, capturing two synchronized views of the same scene in a single snapshot. Flying at a typical altitude of 25 to 30 meters over crops such as wheat, the system produces images with spatial resolution on the order of centimeters, fine enough to distinguish patterns within and between individual plants. But that snapshot design creates a serious computational challenge: because the fluorescence calculation requires combining individual pixels from both channels, the two images must be aligned with subpixel accuracy. Vibration, wind and differences in integration time between the channels mean that even a rigidly mounted camera pair produces slightly shifting geometry from one exposure to the next.

SIFMap solves this with a homography-based alignment strategy. When a camera looks down at a scene that is effectively flat, as is the case for a closed crop canopy viewed from more than 25 meters up, the relationship between two views can be described by a single planar homography matrix, a mathematical transformation borrowed from the geometry of multiple views. The toolbox estimates this homography independently for every image pair using a Random Sample Consensus, or RANSAC, procedure applied to Scale Invariant Feature Transform, or SIFT, keypoints, the workhorse of modern image matching. This per-pair estimation absorbs the distortions introduced by flight dynamics and ensures the pixel-level precision the fluorescence retrieval demands.

Once individual images are internally aligned, the software must figure out how hundreds of overlapping frames fit together into one map. A typical SIFcam flight consists of 300 to 1000 image pairs with 80 percent forward overlap and 70 percent sidelap. SIFMap tackles this in two stages. First, a registration module identifies which images actually overlap by comparing feature descriptors across all image pairs in an all-against-all nearest-neighbour search, then verifies the candidates with projective matching and RANSAC, producing a graph of connected images from which isolated or corrupted frames are automatically removed. Second, a global alignment module places every image into a common coordinate frame. It picks a reference image using a minimum spanning tree of the matching graph, accumulates pairwise motions along the shortest paths, and then refines all transformations simultaneously by minimizing a nonlinear symmetric transfer error with least-squares optimization, complete with analytically computed Jacobian matrices, sparsity analysis and iterative outlier removal.

Transparency is the heart of the project. Commercial photogrammetry packages such as Agisoft Metashape and Pix4Dmapper can process SIFcam data, but their core algorithms are proprietary, making it difficult for scientists to understand how parameter choices affect the radiometry of the final product. The open-source alternative WebODM offers more control but relies on a full Structure-from-Motion pipeline that is computationally heavy, and it does not recognize SIFcam as a multispectral array, processing only one spectral band at a time. In benchmark tests on a 32-core machine, SIFMap stitched a 342-pair dataset in roughly 487 seconds of mapping time and a 931-pair dataset in about 912 seconds, while WebODM needed more than two hours for either dataset. SIFMap matched commercial software on speed while using less memory and, crucially, exposing every step to scrutiny, including pixel-level statistics that reveal how many images contribute to each map pixel.

The architecture is deliberately modular. The toolbox is organized into three main modules, SIFcam, match and align, supported by a data module that defines the underlying structures. Each stage of the pipeline, from radiometric preprocessing through band alignment, registration, global alignment and visualization, is a distinct unit that can be inspected, adjusted or replaced. Preprocessing converts raw digital numbers into calibrated at-sensor radiance using dark-field and flat-field corrections, then into reflectance products at both wavelengths by interpolating against in-flight measurements of Lambertian reflectance panels. The visualization stage offers several aggregation strategies for overlapping images, including a distance-weighted scheme that smooths out lighting changes caused by shifting sun-observer geometry. Because the design is generalized, the same registration and alignment machinery can be applied to other image types, and the developers note that maps could even be generated from ordinary RGB imagery.

For users, the package ships with a small 24-pair test dataset and two complete flights, together with configuration files that parameterize every stage. The most influential settings concern the registration and alignment steps: a RANSAC residual threshold that governs how strict the feature matching is, a minimum correspondence count that decides when two images count as connected, and outlier-removal parameters that control how aggressively noisy feature matches are pruned between successive optimization rounds. The default configuration has proven reliable across comparable SIFcam datasets, but the documentation walks users through the cases where dataset-dependent tuning is warranted. Support is provided by the developers at Forschungszentrum Jülich, where the work was carried out with funding from the German Federal Ministry of Education and Research.

The implications reach well beyond one sensor. Remote sensing of chlorophyll fluorescence has matured over five decades from ground instruments to satellite products that track ecosystem photosynthesis across the globe, yet a persistent gap has remained at the field scale, where researchers need resolution fine enough to see how stress spreads through a crop canopy. Existing airborne and ground-based SIF systems face trade-offs in resolution, cost and deployment flexibility, and low-cost UAV sensor setups have been held back by the lack of adapted processing software. By delivering a fast, transparent, end-to-end mapping pipeline, SIFMap lowers that barrier, letting plant scientists, breeders and agronomists turn routine drone flights into quantitative maps of photosynthetic performance. The current version is restricted to flights above 25 meters over relatively flat, crop-like canopies, because the homography approach assumes planar scenes, but the team is already outlining solutions for closer, more structurally complex targets in future releases, along with additional retrieval and mosaicking methods.

Subject of Research: Open-source software for mapping solar-induced chlorophyll fluorescence from UAV-borne snapshot imaging

Article Title: SIFMap: A python toolbox for creating solar-induced fluorescence maps from UAV-borne snapshot imaging data

Article References: SIFMap: A python toolbox for creating solar-induced fluorescence maps from UAV-borne snapshot imaging data. (n.d.). https://doi.org/10.1016/j.softx.2026.103065

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103065

Keywords: solar-induced fluorescence, SIFMap, UAV remote sensing, chlorophyll fluorescence, SIFcam, Fraunhofer Line Discrimination, image mosaicking, Python, plant phenotyping, photosynthesis, open-source software, precision agriculture

The problem the researchers set out to solve is deceptively simple to describe and notoriously hard to fix. Sugarcane is one of the world’s most important industrial crops, feeding sugar, bioethanol, and bioenergy industries in Brazil, India, China, and Thailand. Conventional fertilization treats entire fields as uniform blocks, applying the same rate everywhere regardless of soil fertility, moisture, topography, or crop vigor. The consequence is a double failure: over-fertilized zones waste money and leak nitrogen into groundwater and the atmosphere, while under-fertilized patches starve silently, dragging down overall yield. In Thailand, where rising fertilizer costs and labor shortages have squeezed growers, the economic and environmental stakes of getting this wrong have never been higher.

The technical heart of the new framework is a two-phase cyber-physical architecture. In the pre-application phase, a DJI Phantom 4 Multispectral drone flies automated missions at 40 meters altitude, capturing imagery in five spectral bands from blue to near-infrared, alongside thermal infrared readings of canopy temperature. With 80 percent forward and 75 percent side image overlap, the flights achieve a ground sampling distance of 3.2 centimeters per pixel, and RTK-GNSS positioning keeps orthomosaic alignment error below roughly four centimeters. The imagery is stitched into orthomosaics, segmented into 5-by-5-meter management zones, and converted into vegetation indices: NDVI, GNDVI, and SAVI, plus a green chlorophyll index. Field teams add ground-truth soil moisture from time-domain reflectometry probes, interpolated across the grid using inverse distance weighting.

Those features feed an XGBoost model, a gradient-boosted decision-tree ensemble chosen for its predictive power, robustness against overfitting, and, crucially, its suitability for embedded hardware. Trained on 180 management-zone samples drawn from six UAV surveys during the fertilization window, with targets ranging from 50 to 150 kilograms of nitrogen per hectare based on leaf diagnostics and Thai Department of Agriculture guidelines, the model achieved a testing R-squared of 0.918, with a root-mean-square error of 15.6 kilograms of nitrogen per hectare. In head-to-head benchmarking against random forest, artificial neural network, and support vector machine models, XGBoost delivered the highest accuracy and the lowest error, with differences that reached statistical significance. Its residuals clustered tightly around zero, with roughly 72.5 percent of predictions falling within ten kilograms of the reference rate, and the largest deviations confined to transitional crop zones where canopy conditions changed abruptly.

What distinguishes this work from a growing pile of AI-in-agriculture papers is the insistence on transparency. Rather than accepting the model as a black box, the team integrated SHAP, SHapley Additive exPlanations, a technique rooted in cooperative game theory that quantifies exactly how much each input feature pushed any given prediction up or down. The analysis revealed that NDVI alone accounted for 28.4 percent of the model’s total influence, with canopy temperature contributing another 23.1 percent, meaning crop vigor and physiological stress together drove more than half of every fertilizer decision. Higher NDVI values reduced recommended nitrogen, while elevated canopy temperatures, a signature of stressed, less transpiring plants, pushed recommendations upward. SHAP dependence analysis even uncovered threshold-like behavior: below an NDVI of roughly 0.52, recommendations rose sharply, while above 0.72 they plateaued, a nonlinear pattern that linear agronomic rules would miss entirely.

The explainability is not merely academic. Because every prescription can be traced back to specific, agronomically interpretable variables, farmers and agronomists can audit why a particular zone received 154 kilograms of nitrogen while its neighbor received 82. The researchers report that the SHAP layer improved detection of abnormal predictions, enabled spatial inconsistency diagnosis, and increased operator confidence, addressing one of the most persistent barriers to AI adoption in agriculture: distrust of opaque recommendations. In a low-vigor zone, for example, low NDVI and GNDVI combined with high canopy temperature produced positive SHAP contributions that raised the nitrogen rate, while stronger chlorophyll indices partially offset the increase, a narrative any trained agronomist can follow.

Execution happens at the field edge, not in the cloud. The trained model and a GPS-referenced prescription map are deployed on an NVIDIA Jetson Nano, a credit-card-sized computer running in its standard 10-watt mode, paired with an ESP32 microcontroller that drives the fertilizer metering hardware. During application, the Jetson Nano looks up the current GNSS position, retrieves the zone’s target rate, and sends commands to the ESP32, which converts the rate into an auger rotational speed using a laboratory-calibrated linear relationship, Q = 0.0575ω − 0.35, that achieved an R-squared of 0.999 across speeds from 20 to 100 rpm. Pulse-width modulation at 0.1 percent duty-cycle resolution regulates a DC motor driving the auger, delivering between 0.8 and 5.4 kilograms of fertilizer per minute. The measured latencies are striking: 82 milliseconds for AI inference, 28 milliseconds for communication, and 92 milliseconds for the embedded control loop, fast enough that at a 4-meter-per-second travel speed the system can update fertilizer rates every 32.8 centimeters of travel.

The agronomic payoff was tested in a randomized complete block design with three treatments and four replications. Conventional farmer practice applied an average of 148.6 kilograms of nitrogen per hectare; the AI-driven variable-rate system applied 116.2, a 21.8 percent reduction, while cutting over-application zones from 31.5 percent of the field to 8.7 percent. Nitrogen-use efficiency climbed from 43.1 to 54.9 percent, estimated runoff nitrogen losses fell by 42.3 percent, and fertilizer costs dropped by roughly 21.9 percent per hectare. Yields rose from 82.4 to 91.9 tonnes per hectare, and the benefits extended beyond raw tonnage: the standard deviation of plant height fell by 69.1 percent, stem-diameter variability by 63.9 percent, and chlorophyll-index variability by 65.9 percent, producing a more uniform crop that synchronizes better with mechanized harvesting and sugar processing. Low-vigor zones, which received up to 4.8 percent more fertilizer than conventional practice, recovered most dramatically, gaining 15.8 percent in yield.

The authors are candid about the study’s limits. Validation covered a single commercial field and one growing season, and the test set came from the same field and season as the training data, meaning the evaluation represents internal rather than independent spatial or temporal validation. Environmental benefits such as reduced nitrate leaching and greenhouse-gas emissions were modeled rather than directly measured, the applicator relied on calibrated feedforward control without closed-loop flow sensors, and dynamic response behavior and field distribution uniformity remain uncharacterized. The team’s stated contribution is deliberately positioned at the system level: not a new sensor, algorithm, or GIS method, but the first field-validated pathway linking UAV sensing, explainable machine learning, edge computing, and physical variable-rate actuation in commercial sugarcane.

Even with those caveats, the implications ripple outward. Fertilizer production is among the most carbon-intensive industrial processes on Earth, and nitrogen that leaves fields as nitrate or nitrous oxide imposes costs far beyond the farm gate. A framework that trims inputs by more than a fifth while raising yields, and that explains every decision in terms a farmer can verify, offers a template that could extend to maize, wheat, rice, and other row crops. The researchers point toward multi-site, multi-season validation, multi-nutrient management, and autonomous spatiotemporal decision-making as next steps. If those follow-up trials hold, the sight of a drone prescribing fertilizer zone by zone, with an AI that shows its work, may soon be as ordinary in sugarcane country as the harvest itself.

Subject of Research: Explainable AI-driven UAV remote sensing for variable-rate nitrogen fertilization in precision sugarcane nutrient management

Article Title: Field-validated explainable AI-based UAV remote sensing for variable-rate fertilization in precision sugarcane nutrient management

Article References: Sangpradit, K., & Samseemoung, G. (2026). Field-validated explainable AI-based UAV remote sensing for variable-rate fertilization in precision sugarcane nutrient management. Smart Agricultural Technology, 15, Article 102585. https://doi.org/10.1016/j.atech.2026.102585

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102585

Keywords: precision agriculture, UAV remote sensing, explainable AI, XGBoost, SHAP, variable-rate fertilization, sugarcane, nitrogen-use efficiency, edge computing, multispectral imaging, Jetson Nano, sustainable farming

Tiller density, defined as the number of tillers per square meter, is set during a narrow window of crop development that runs roughly from the three-leaf stage to jointing, corresponding to the Feekes 3 through 6 growth stages. During this period, genetics, weather, soil conditions, and management decisions all interact to shape how many shoots survive. Because tiller density directly influences spike number at harvest, kernels per spike, and kernel weight, and is closely tied to the crop’s nitrogen status, knowing it early gives breeders and farmers a powerful lever. Breeders can use the information to select genotypes with optimal tillering behavior, while farmers can fine-tune nitrogen applications before the crop locks in its yield potential.

The problem has always been scale. Manual counting is accurate but punishingly slow, and point-scale sampling misses the spatial variability that defines real fields. Ground-based tools such as handheld spectrometers and terrestrial LiDAR offer richer data but remain labor-intensive across large or multi-site trials. Unmanned aerial vehicles have emerged as the natural middle ground, capturing high-resolution multispectral imagery over breeding nurseries and commercial fields in a fraction of the time. Previous UAV studies of wheat tillering, however, leaned heavily on a small set of hand-crafted vegetation indices, leaving the full information content of drone imagery largely untapped.

The research team, led by scientists at South Dakota State University, set out to close that gap with a framework that fuses two complementary families of image features. The first is spectral: five reflectance bands from a MicaSense Altum-PT multispectral sensor, transformed into forty-eight vegetation indices that capture canopy greenness, chlorophyll content, pigment composition, and stress responses. The second is textural: forty Gray-Level Co-occurrence Matrix statistics—mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation—computed for each band, which describe the spatial structure, patchiness, and uniformity of the canopy in ways that average reflectance alone cannot.

The field campaign was ambitious in its breadth. The team collected 721 ground-truth samples across five South Dakota wheat fields between 2023 and 2024, spanning two university breeding sites and three commercial farmer fields in Brookings, Hayes, and Winner. At each sampling point, PVC quadrats of roughly 0.15 square meters were placed in the plots, tillers were counted by hand at the Feekes 4 or 5 stage, and a DJI Matrice 210 RTK drone flew overhead at 45 meters with eighty percent image overlap, producing orthomosaics with a ground sampling distance of 1.35 centimeters per pixel. Reflectance panels and a downwelling light sensor handled radiometric calibration, while survey-grade ground control points delivered centimeter-level geometric accuracy. The measured tiller densities ranged from 156 to nearly 1,356 tillers per square meter, with a mean of 628.9 and a coefficient of variation of 35.7 percent—a wide dynamic range that any predictive model would need to master.

With 88 predictors per sample in hand, the researchers benchmarked four regression architectures: Random Forest Regression, a fully connected Deep Neural Network, a one-dimensional Convolutional Neural Network, and an attention-enhanced CNN dubbed Atten-CNN. The attention module was a multi-head self-attention block with four heads and a key dimension of 32, inserted after the convolutional feature extractor. Rather than treating all inputs equally, the attention mechanism dynamically reweights the 88 spectral and texture features, amplifying the informative ones and suppressing noise—a capability that standard CNNs, which excel at local patterns but struggle with long-range dependencies, lack. Hyperparameters were tuned with five-fold cross-validation, and the models were trained with the Adam optimizer under a step-decay learning rate schedule.

The results were decisive. On the held-out test set, Atten-CNN achieved a coefficient of determination of 0.81 with a relative root mean square error of 15.97 percent, outperforming the baseline CNN (R² of 0.77), Random Forest (0.71), and the DNN (0.70). The convolutional architectures also showed tighter, less biased residuals across the tiller density gradient, while Random Forest and DNN exhibited a stronger tendency to overestimate low densities and underestimate high ones. To probe spatial generalizability, the team ran a leave-one-location-out validation in which the model was trained on three sites and tested on a completely unseen fourth. Performance in commercial fields was striking—R² values of 0.95, 0.88, and 0.86 in Brookings, Hayes, and Winner, with relative errors as low as 7 percent—but dropped to 0.34 when extrapolating to the genetically diverse Aurora breeding nursery, a signal that breeding plots with their extreme phenotypic variability carry spectral-textural signatures that production-field training data do not fully cover.

Interpretability came from SHAP analysis, which quantifies each feature’s contribution to individual predictions. The red-edge sensitive indices CRI_2 and NDRE emerged as the two most influential predictors for the convolutional models, consistent with the physiology of tillering: denser stands produce more green leaf area and a stronger canopy chlorophyll signal early in the season. Supporting players included the MERIS Terrestrial Chlorophyll Index, the TCARI/OSAVI ratio, and the Datt Index. Notably, the attention mechanism elevated band-wise GLCM Correlation in the green, red, and blue bands to the top of the importance ranking, immediately behind the red-edge indices—evidence that the model was exploiting within-canopy spatial patterns, pixel co-occurrence structure, and canopy patchiness as indirect proxies for tiller distribution and stand uniformity.

The study is candid about its limits. All models underestimated tiller density at the high end of the range, a classic saturation effect in optical remote sensing where reflectance signals lose sensitivity as canopies close. The authors propose three remedies: adding three-dimensional structural features from LiDAR or Structure-from-Motion point clouds, incorporating thermal imagery that responds to transpiration-driven cooling in dense stands, and applying weighted loss functions that penalize errors in the upper density quartile more heavily. The random train-test split also reflects within-distribution performance rather than true temporal transferability, and the framework’s validity is currently bounded by the soils, management systems, and density ranges of the eastern and central Great Plains. Future work will pursue multi-year datasets, end-to-end models that learn directly from raw imagery, and transfer learning across regions and wheat classes.

Even with those caveats, the implications are substantial. A drone flight lasting minutes can now deliver plot-level tiller density estimates that once required days of fieldwork, at an accuracy sufficient to guide early-season nitrogen management and to rank thousands of breeding genotypes for tillering behavior. Because the framework is built on extracted features rather than raw image patches, it remains data-efficient in a domain where labeled samples are scarce, and it can be readily adapted to other traits and crops. As attention-based architectures continue to prove their worth in agricultural remote sensing—from yield prediction to disease detection—this study marks another step toward digital phenotyping at scale, where the field itself becomes a continuously monitored, data-rich experiment.

Subject of Research: UAV-based estimation of wheat tiller density using spectral and texture features with attention-based deep learning

Article Title: High-throughput estimation of wheat tiller density using UAV-derived spectral-textural features and attention-based deep learning

Article References: Kaushal, S., Maimaitijiang, M., Subedi, S., Thapa, S., Janjua, U. U. R., Koupal, D. J., Singh, M., Kaur, K., Kumar, P., Sitaula, P. R., Billah, M. M., Halder, J., Irshad, M. A., & Sehgal, S. K. (2026). High-throughput estimation of wheat tiller density using UAV-derived spectral-textural features and attention-based deep learning. Smart Agricultural Technology, 15, Article 102521. https://doi.org/10.1016/j.atech.2026.102521

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102521

Keywords: wheat, tiller density, UAV remote sensing, deep learning, attention mechanism, high-throughput phenotyping, multispectral imagery, vegetation indices, GLCM texture features, SHAP interpretability, precision agriculture, crop breeding

Temperate savanna woodlands form a broad ecotone between closed forests and open grasslands, and the ones studied here span two vast sandy regions of Inner Mongolia: the Hunsandak Sandland, covering roughly 52,000 square kilometers, and the Horqin Sandland, extending over about 35,100 square kilometers. These landscapes support exceptionally rich biodiversity and act as reservoirs of genetic resources, yet the woody vegetation that anchors them has received far less attention than the grasses and herbs around it. Because trees and shrubs in these systems control carbon storage, community stability, and the spatial distribution of biodiversity, knowing exactly where each individual tree stands, and how large its crown is, is fundamental to nearly every ecological question that can be asked about the region.

The challenge is that savanna trees are maddeningly heterogeneous. Natural savanna woodlands stretch across large environmental gradients, and the woody vegetation responds with pronounced spatial variation, producing wildly diverse phenotypes at the level of single trees. Classical image-processing techniques such as local maximum filtering, marker-controlled watershed segmentation, template matching, region growing, and edge detection can work under simple conditions, but they falter in natural forests with complex canopy structures. They depend on manual parameter tuning, extract features only crudely, and generalize poorly. Deep convolutional neural networks offered a way forward, but the standard workhorse for instance segmentation, Mask R-CNN, carries a heavy price: its ResNet-50 backbone and Feature Pyramid Network demand substantial computational resources, and the pyramid module can lose information when propagating features to high resolutions, undermining the detection of small or sparsely distributed crowns.

To build DeepTree, the research team first assembled an unusually rich dataset. Between July 20 and 30, 2020, a DJI Phantom 4 RTK quadcopter flew nine sample plots across six banners of Inner Mongolia under clear, cloudless skies, capturing thousands of photographs at roughly 100 meters altitude with 75 percent forward and 70 percent side overlap. The imagery was processed into nine digital orthophotos with pixel resolutions between 2.63 and 3.06 centimeters, then resampled to a uniform 10 centimeters and sliced into 1,000-by-1,000-pixel tiles, each covering one hectare, yielding 663 tiles in total. From eight of the plots, the team manually annotated 8,724 individual tree crowns using Labelme, exporting the labels in COCO format. The ninth plot was deliberately held back as an independent cross-site test, ensuring that no spatially overlapping regions leaked between training and evaluation data.

The architectural surgery at the heart of DeepTree is elegant in its simplicity. The researchers replaced ResNet-50 with MobileNet V3, a backbone that uses just 13 layers and depthwise separable convolutions instead of 50 conventional layers, dramatically cutting computational complexity. They swapped the Feature Pyramid Network for NAS-FPN, a feature-fusion module discovered through neural architecture search, which enhances multi-scale representation and improves detection of small targets while using only 32 channels. The framework then branches into two variants: MNMS R-CNN (S), which pairs this lightweight backbone and neck with a standard region-of-interest head for maximum speed, and MNMS R-CNN (L), which adds an enhanced SCNet-based head with Squeeze-and-Excitation, Global Context, and Feature Relay components to prioritize detection completeness and segmentation accuracy.

The performance gains are remarkable. MNMS R-CNN (S) slashed computational cost from 248.0 GFLOPs to just 5.6 GFLOPs and shrank the parameter count from 42.9 million to 3.8 million, while boosting inference speed from 10.2 to 19.2 frames per second, an 88.2 percent improvement, all on 1.5 gigabytes of GPU memory instead of 6.2. The larger variant, MNMS R-CNN (L), required only 9.0 GFLOPs and 8.1 million parameters yet achieved the best overall accuracy in the study: bounding-box average precision rose from 34.9 to 46.5 percent, a relative improvement of about 33 percent, while recall jumped from 62.4 to 81.2 percent and the F1 score climbed to 81.1 percent. Comparisons against YOLO11n-seg and the high-capacity HTC framework confirmed that neither compact size nor raw capacity alone delivers the right balance; MNMS R-CNN (L) outperformed both on the key detection metrics while remaining far leaner than HTC’s 76.93 million parameters and 401.1 GFLOPs.

Perhaps the most ecologically telling results emerged when the models were tested across different canopy densities. In high-density scenes, MNMS R-CNN (L) lifted recall from Mask R-CNN’s 45.5 percent to 84.9 percent, cutting missed crowns from 72 to just 20 in a representative case. Even in sparse, low-density stands, where isolated small crowns blend into the grassy background, it nearly doubled recall, from 28.2 to 46.5 percent. Stratified analysis revealed a persistent pattern: omission errors concentrated among smaller crowns, particularly in low-density scenes where weak spectral and textural contrast against grass undermined detection. The team also found that performance degraded as imagery was downsampled, with F1 falling from 81.1 percent at 10 centimeters to 74.1 percent at 0.5 meters and 55.4 percent at 1.0 meter, underscoring how much fine spatial detail matters for resolving individual crowns.

Beyond simply finding trees, DeepTree extracts the structural traits that ecologists actually need. For crown width, the coefficient of determination reached 0.98 with a root mean square error of just 0.30 meters; for crown area, R squared was again 0.98 with an error of 2.66 square meters. The model recovered crown-width estimates spanning 1.41 to 16.05 meters and crown areas from 1.33 to 203.71 square meters, closely matching reference ranges that Mask R-CNN systematically truncated. An independent field validation using 166 measured elm trees, matched to drone detections by geographic coordinates, yielded R squared values of 0.82 for crown width and 0.78 for tree height, with the modest gap attributable to the one-year interval between imagery acquisition and field measurement, GPS positioning uncertainty, and differences in how crown width is defined on the ground versus in a segmentation mask.

Applied across all nine plots, the model mapped tree densities ranging from about 30 to 102.7 trees per hectare and canopy areas from roughly 709 to 3,117 square meters per hectare, revealing stark spatial heterogeneity in woody cover. In the held-out ninth plot, never seen during training, the model still achieved 75.8 percent precision and 76.0 percent recall, demonstrating genuine cross-site transferability, the property that separates a local demonstration from a scalable monitoring tool. The authors acknowledge limitations: RGB imagery alone struggles when crowns, shrubs, and grass share similar colors, and future work should fuse in LiDAR and spectral data, expand training libraries to more species and seasons, and test the framework across broader environmental gradients. Even so, DeepTree points toward a future where drone-based biodiversity assessment, carbon accounting, and long-term monitoring of the world’s fragile forest-grassland ecotones can run on modest hardware, one tree at a time.

Subject of Research: Lightweight deep learning instance segmentation for individual tree crown delineation and structural trait extraction from UAV imagery in temperate savanna woodlands

Article Title: DeepTree: Deep learning-based lightweight instance segmentation for individual tree crown delineation and structural trait extraction

Article References: Duan, T., Yang, B., Li, X., Hou, D., Hu, P., Li, X., Cong, W., & Wang, F. (2026). DeepTree: Deep learning-based lightweight instance segmentation for individual tree crown delineation and structural trait extraction. Smart Agricultural Technology, 15, Article 102552. https://doi.org/10.1016/j.atech.2026.102552

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102552

Keywords: deep learning, UAV remote sensing, instance segmentation, tree crown delineation, Mask R-CNN, MobileNet V3, NAS-FPN, savanna woodlands, Inner Mongolia, ecological monitoring, phenotype extraction, lightweight models

The study, published in the journal Environmental Challenges, tackled a deceptively simple question: is a given grassland on peat soil shallow-drained or deep-drained? The answer matters enormously for climate accounting. Under guidelines from the Intergovernmental Panel on Climate Change, drainage status is defined by the mean annual water table depth. If the water table sits, on average, less than 30 centimetres below the surface, the site is classed as shallow-drained; at 30 centimetres or deeper, it is deep-drained. Each class carries a different emission factor, so national greenhouse gas inventories depend on knowing which category applies where. Ireland’s inventory currently estimates around 339,130 hectares of grassland peat soils, of which roughly 141,000 hectares are believed to be deeply drained, but the spatial detail behind those figures remains coarse.

Traditional monitoring relies on dipwells, perforated pipes sunk into the soil and fitted with pressure loggers that record water levels every 15 minutes. These instruments are accurate but expensive to install and maintain, and they only describe the ground immediately around each sensor. Extrapolating from a handful of points across an entire farm, let alone a whole country, invites error. Remote sensing promised a way out, yet satellite imagery is frequently blocked by Ireland’s stubborn cloud cover and its resolution, typically around 10 metres, is too coarse to resolve the drainage ditches and microtopography that govern water movement in peat.

The researchers, led by Charmaine Cruz of Dublin City University together with colleagues from Teagasc and other Irish institutions, turned to unoccupied aerial vehicles flying below the clouds. Their platform was a DJI Matrice 350 RTK carrying two sophisticated instruments: a MicaSense Altum-PT sensor capturing five multispectral bands plus thermal imagery sensitive to temperature differences of just 0.05 degrees Celsius, and a DJI Zenmuse L2 LiDAR system capable of building centimetre-scale three-dimensional models of the ground surface. Flights at altitudes of 100 to 120 metres, with 80 percent overlap between images, were conducted twice at each site, once in summer and once in winter, to capture the seasonal extremes between which the mean annual water table lies.

Two contrasting commercial farms in County Offaly served as the proving grounds, roughly 10 kilometres apart. The Clara site, about 97 hectares of intensively grazed dairy grassland, borders a natural raised bog. The Tumbeagh site, around 27 hectares of ungrazed, unimproved grassland, contains heterogeneous fen peat with cutover margins and borders a state-owned industrial bog previously used for peat extraction. Both farms carry a legacy of drainage ditches of varying depth and condition. Peat depth, probed on a 15 by 15 metre grid with fibreglass rods and verified with soil corers, ranged from zero to over five metres at both sites, revealing dramatic lateral transitions from mineral ground to deep organic soil.

On the ground, the team installed 12 dipwells at Tumbeagh and 16 at Clara, each fitted with a submersible pressure transducer logging water levels every 15 minutes, and supplemented these with 10 older dipwells per site. Tipping-bucket rain gauges recorded rainfall at the same temporal resolution. The summer surveys followed dry spells, with just 7.4 and 10.0 millimetres of rain in the preceding fortnight, while winter surveys followed 65.4 and 82.8 millimetres. As expected, water tables plunged in summer, at some mineral-soil dipwells reaching more than two metres down, and rose in winter, occasionally breaking the surface.

From the drone data the researchers generated 15 candidate explanatory variables: the five calibrated reflectance bands, vegetation indices such as NDVI, EVI2, NDWI and MSAVI2, a thermal index, a LiDAR-derived digital terrain model with slope and topographic wetness index, distance to mapped open drains, and interpolated peat depth. A random forest regression model, implemented in Python’s scikit-learn library, learned the relationship between these proxies and the dipwell measurements, then predicted water table depth across every pixel of each farm. Model performance, assessed through the out-of-bag score, ranged from 0.66 to 0.84, with the strongest performance at the Clara site in summer.

The variable importance rankings delivered the study’s most striking insight. In summer, peat depth dominated predictions at both sites, with importance scores of roughly 0.30 at Tumbeagh and 0.23 at Clara, far ahead of any other variable. Deeper peat consistently coincided with shallower summer water tables. In winter, however, the hierarchy shifted: topography took the lead at Tumbeagh while surface temperature became the top predictor at Clara, with peat depth slipping to second or third place. This seasonal flip means a model trained on one season’s imagery cannot simply be reused for another, underscoring the value of the two-survey approach the team adopted.

When the summer and winter predictions were averaged and classified against the IPCC threshold, the verdict was sobering. At Tumbeagh, every field fell into the deep-drained category, with 91 percent of the peat area showing mean annual water tables between 30 and 60 centimetres. At Clara, 94 percent was deep-drained, though a small 2-hectare zone near restored peatland and redundant drains qualified as shallow. The classification matched an independent year-round hydrological assessment of the same farms, suggesting that two well-timed drone flights can substitute for continuous monitoring where resources are limited. Notably, the model assigned drainage status even to fields without dipwells, extending coverage beyond what ground measurements alone could achieve.

The implications stretch from carbon farming schemes to European Union reporting obligations. Field-scale drainage maps could help regulators prioritise rewetting projects, guide farmers toward lower-intensity management on the wettest ground, and feed into future Tier 3 national inventory methods that demand spatially explicit emission estimates. The approach has limits: battery life, line-of-sight rules and altitude restrictions confine drone surveys to a few square kilometres, so national mapping would still require satellite imagery, likely with separate models for fens, raised bogs and blanket bogs whose distinct vegetation confounds transferable predictions. The authors also point toward hyperspectral sensors, ground-penetrating radar and deep learning as the next frontier. For now, the message is clear: the hidden plumbing of Ireland’s peat grasslands, long inferred from sparse points on a map, can finally be seen in full, one centimetre-scale pixel at a time.

Subject of Research: Mapping water table depth and drainage status of grassland peat soils using UAV remote sensing and machine learning

Article Title: Determination of drainage status at varied grassland peat soil sites using a combination of multi-sensor UAV technology and on-the-ground measurements

Article References: Cruz, C., Fenton, O., Bari, M. I., Broderick, E., Daly, E., Donoghue, M., Fealy, R. M., Green, S., McCarthy, E., O'Sullivan, L., Shnel, A., Tuohy, P., & Connolly, J. (2026). Determination of drainage status at varied grassland peat soil sites using a combination of multi-sensor UAV technology and on-the-ground measurements. Environmental Challenges, 25, Article 101667. https://doi.org/10.1016/j.envc.2026.101667

Image Credits: AI Generated

DOI: 10.1016/j.envc.2026.101667

Keywords: peat soils, water table depth, UAV remote sensing, LiDAR, random forest, greenhouse gas emissions, drainage status, Ireland, grassland, machine learning, carbon emissions, rewetting

The review, published in Smart Agricultural Technology, synthesises 199 peer-reviewed studies published between 2014 and 2026 that applied unmanned aerial vehicle remote sensing to rice phenotyping, the systematic measurement of plant traits. Led by Xiaoao Yang of Guangdong Province and colleagues including Spyros Fountas of Greece’s Agricultural University of Athens, the team screened 2,327 records from Scopus and Web of Science, ultimately distilling a corpus they call Core-199. Their goal was to build a rice-specific framework connecting platforms, sensors, data-processing methods and agronomic tasks, from nitrogen diagnosis to yield prediction, in a field that has grown explosively but unevenly.

The case for drones rests on a simple observational gap. Manual field surveys are labour-intensive and subjective, demanding an estimated 200 to 500 man-hours per hectare, while destructive sampling rules out repeated monitoring of the same plants. Satellites, meanwhile, revisit fields at best every five days and deliver pixels of 10 to 30 metres, far coarser than the 0.01 to 0.1 metre detail needed to evaluate individual breeding plots. UAVs slot neatly between these extremes, offering centimetre-scale imagery on schedules chosen by the operator, carrying anything from cheap RGB cameras to hyperspectral imagers and laser scanners.

Platform choice matters. Multirotor drones dominate rice phenotyping because they hover stably, need little takeoff space and suit the small, fragmented paddies typical of Asian rice farming. Fixed-wing aircraft cover more ground faster but cannot hover and demand launch infrastructure, so they remain rare in plot-level work. The review also situates UAVs within a three-tier observation system: satellites for broad regional mapping, drones for sub-decimetre plot detail, and handheld or tractor-mounted proximal sensors for calibration and real-time decisions. Economic analyses cited in the review suggest the tiers are complementary, with break-even areas for satellite-based nitrogen management ranging from about 2.5 to 13 hectares depending on imagery resolution.

Each sensor family contributes something distinct. RGB cameras, cheap and sharp, excel at structural traits: plant height from digital surface models, canopy cover, panicle counting and lodging assessment. Multispectral cameras capture the red-edge and near-infrared bands that power vegetation indices such as NDVI, supporting routine retrieval of leaf area index, chlorophyll, nitrogen status and yield. Hyperspectral imagers, with hundreds of narrow bands, detect subtle biochemical signals and early stress but bring high costs and data redundancy. LiDAR actively probes the three-dimensional canopy, and thermal infrared cameras, used in only two of the 199 studies, reveal canopy temperature linked to water status and heat tolerance, a capability the authors flag as a high-priority research frontier as flowering-stage heat stress intensifies.

The application chapters reveal both striking progress and stubborn caveats. Nitrogen monitoring is the most mature task, with reported coefficients of determination ranging from about 0.49 to 0.94 depending on sensor, trait and model, and one machine-learning precision-nitrogen strategy raising yields by 7 to 15 percent and economic returns by 4 to 16 percent. Chlorophyll retrieval has reached R-squared values as high as 0.97 in multistage hyperspectral frameworks. Yet the review repeatedly warns that these numbers cannot be compared across studies, because target traits, units, growth stages, sensors and validation designs differ so widely, and the authors decline to claim any general superiority for multisource fusion over simpler single-sensor approaches.

Lodging and disease monitoring showcase the field’s move toward real-time, on-board intelligence. Semantic segmentation networks now delineate lodged rice at pixel level with mean intersection-over-union above 90 percent, and one edge-computing workflow on an Nvidia Jetson Xavier NX processed imagery at nearly 14,418 square metres per second, covering roughly 10 square kilometres in an 80-minute flight. On the disease front, thermal and optical fusion allowed researchers to identify infection a remarkable 72 hours before visible lesions appeared, with an FPGA implementation consuming just 0.076 watts per classification. A lightweight false-smut detector built on YOLOv12n cut parameters by 25 percent while maintaining a mean average precision of 80.7 percent.

Yield prediction has evolved from single-date vegetation indices into multi-temporal, multisource and even process-coupled models that assimilate drone-derived nitrogen into crop simulation frameworks such as CERES-Rice. Organ-level phenotyping offers an alternative route: detecting and counting panicles from aerial imagery, with one framework classifying yield levels at 83.63 percent accuracy and another reporting yield-estimation errors between 1.4 and 11.7 percent across test plots. The review also cautions that some eye-catching near-perfect R-squared values in the literature describe proxy traits such as panicle counts or plant height rather than direct grain-yield prediction, and one segmentation-based study with R-squared of 0.98 carried relative errors of 21 to 31 percent.

Perhaps the most sobering statistic concerns genetics. Only six of the 199 reviewed studies linked UAV-derived traits to genetic association analysis, three using genome-wide association studies and three using QTL mapping. Those that did recovered known genes such as sd1, Ghd7.1 and TAC1, and one drought study across 240 accessions identified 111 significant loci, but no drone-derived QTL has yet been independently validated by another research group. The authors frame this as the principal evidence gap between high-throughput phenotyping and breeding impact.

Cross-regional generalisation emerges as the field’s central technical challenge. Environmental background interference, especially the standing water, sun glint and mixed pixels of flooded paddies, and site or cultivar bias are the best-documented causes of model failure when models move between regions, years or seasons. The review proposes a staged roadmap: standardised multi-environment public datasets with rich metadata as the foundation, interpretable hybrid models combining radiative-transfer physics with machine learning in the transition, and eventually closed-loop systems connecting drones, cloud processing and field decisions. Until then, the authors conclude, the technology’s promise depends less on fancier algorithms than on standardised data, honest external validation and independent confirmation of genetic findings, the unglamorous infrastructure that will decide whether drone phenotyping graduates from academic demonstration to routine agricultural practice.

Subject of Research: UAV remote sensing for high-throughput rice phenotyping

Article Title: Advances in UAV remote sensing for high-throughput rice phenotyping: a systematic review

Article References: Yang, X., Zhou, Z., Huang, H., Wei, X., Kong, X., Fountas, S., & Tang, Y. (2026). Advances in UAV remote sensing for high-throughput rice phenotyping: a systematic review. Smart Agricultural Technology, 15, Article 102561. https://doi.org/10.1016/j.atech.2026.102561

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102561

Keywords: UAV remote sensing, rice phenotyping, precision agriculture, high-throughput phenotyping, hyperspectral imaging, LiDAR, yield prediction, nitrogen monitoring, disease detection, lodging monitoring, GWAS, edge computing