Search Results (939)

Plant diseases, being a subject of interdisciplinary research, significantly reduce crop yield, quality, and economic returns, while the misidentification of pathogens often leads to ineffective treatments and may harm beneficial organisms and ecosystems. This work develops an approach for robust visual classification of plant diseases under limited and heterogeneous data based on multi-scale fractal texture descriptors integrated into a convolutional neural network. The proposed method employs wavelet transform modulus maxima to extract two complementary fractal characteristics, local fractal dimension and singularity spectrum width, from leaf images at several spatial scales. These descriptors form multi-channel fractal maps fed into a fractal attention module (FAM) inserted after the third stage of a ResNet-50 architecture. The FAM learns to emphasize spatial regions where fractal properties are most discriminative, while a parallel branch encodes global fractal statistics into an auxiliary vector combined with backbone features at the final classification layer. Experiments are conducted on a large heterogeneous collection of 11 public plant disease datasets under 5-shot, 50-shot, and full-scale training regimes. The fractal-augmented model raises classification accuracy from 57.06% to 67.73% on 5 shots and from 80.81% to 86.11% on 50 shots, red outperforming the plain ResNet-50 in these settings, converges within 1–2 epochs versus 25–40, and shows markedly better resilience to color distortions, random occlusions, and grayscale conversion in most cases. The generated attention maps provide spatially explicit explanations of the model’s decisions, increasing transparency for practical use. The proposed approach demonstrates that fractal analysis, embedded as a modulating signal inside a deep network, can serve as an efficient and interpretable inductive bias, which is particularly valuable under data scarcity and noisy agricultural imagery. Full article

Early-season crop yield prediction is critical for food security assessment and timely agricultural decision-making, yet large-scale remote-sensing applications are often constrained by a mismatch between pixel-level observations and county-level yield labels. This mismatch limits the use of fine-grained spatial heterogeneity, especially under partial-season observations. Single-scale approaches are also limited in capturing complementary pixel-level and county-level information, restricting representation of yield formation processes. To address this, we propose TandemNet, a multi-scale multiple-instance learning framework for early-season prediction of japonica rice yield in Northeast China. TandemNet treats each county–year as a bag of rice pixels and adopts a dual-branch architecture to jointly learn pixel-level growth trajectories and county-level statistical responses. A phenology-conditioned cross-attention module fuses the two scales under varying growing-season windows. Using Sentinel-1, Sentinel-2, and MODIS data from 2018 to 2023, leave-one-year-out validation shows that TandemNet outperforms Random Forest, XGBoost, LSTM, and Transformer baselines across most phenological stages. It achieves reliable prediction at the tillering stage, approximately 2–3 months before harvest, with an $R^2$ of 0.69 and an RMSE of 655.98 kg/ha. Ablation and attention analyses further indicate a stage-dependent shift from local heterogeneity to county-level consistency. These results demonstrate that modeling cross-scale interactions improves the timeliness, accuracy, and interpretability of rice yield prediction. Full article

Accurate and transferable soybean mapping is essential for agricultural monitoring and area verification, yet conventional Transformer models can be limited in modeling complex nonlinear phenological relationships and maintaining training stability. This study proposes an improved Transformer-KAN model for multi-temporal Sentinel-2 data. Temporal positional encoding, multi-head self-attention with relative positional bias, and a Pre-LayerNorm residual structure are introduced to strengthen phenological sequence modeling, while FastKAN replaces the conventional MLP-based feed-forward network to enhance nonlinear feature representation. The model was trained using 2023 samples from Hailun City and directly evaluated in Bozhou, McLean, and Cass without retraining or fine-tuning. Cross-year transferability was further evaluated by applying the Hailun-trained model to data from Bozhou and McLean from 2021 to 2025. The proposed model achieved overall accuracies of 0.975, 0.983, 0.958, and 0.966 in Hailun, Bozhou, McLean, and Cass, respectively, with corresponding Kappa coefficients of 0.895, 0.887, 0.883, and 0.916. Complexity analysis showed that Transformer-KAN required 0.8142 M parameters and 1.1967 M FLOPs, with an average inference time of 3.0191 ms per sample, compared with 0.6130 M parameters, 0.7971 M FLOPs, and 1.4009 ms per sample for the conventional Transformer, indicating that the improved feature representation was accompanied by increased computational complexity. Cross-year classification performance remained generally high from 2021 to 2025, although interannual variations were observed due to differences in crop growth conditions, phenological timing, and image acquisition quality. Discrepancies between remote-sensing-derived and officially reported soybean areas were mainly related to differences in statistical definitions and residual classification uncertainties, rather than model transferability. Overall, Transformer-KAN provides accurate and transferable soybean mapping and can serve as a spatially explicit complement to official agricultural statistics. Full article

Multi-angle measurements provide essential data on the anisotropic reflectance properties of vegetation, enabling more robust retrievals of leaf area index (LAI), clumping index (CI), and canopy gap fraction compared to conventional single-view remote sensing. While Unmanned Aerial Vehicles (UAVs) offer unprecedented centimeter-level spatial resolution and flexible deployment, their application exposes a fundamental scale mismatch between ultra-high-resolution imagery and traditional bidirectional reflectance distribution function (BRDF) models. This review explicitly identifies that conventional 1D radiative transfer models (RTMs), which rely on the assumption of a statistically homogeneous canopy, suffer from severe scale-dependent biases, such as systematically underestimating hotspot reflectance (e.g., observed biases of 25% to 40% in 5-cm resolution UAV studies over specific vegetation canopies), and structural-optical confounding when directly applied to UAV data. At centimeter scales, macroscopic structural heterogeneity disrupts this homogeneity, necessitating the use of 3D RTMs that can explicitly simulate geometric occlusion and complex multiple scattering processes in highly heterogeneous environments. To bridge these theoretical and operational gaps, this review uniquely synthesizes UAV-specific multi-angle methodologies, systematically correlating canopy architectural types with optimal sensor configurations, flight strategies, and BRDF modeling frameworks. By evaluating recent advancements in multimodal data fusion, physics-informed machine learning, and physiological parameter retrieval, this review provides a comprehensive roadmap for decoupling structural and biochemical traits, highlighting how multi-angle directional signatures can substantially elevate classification accuracy, with specific experiments on spectrally similar crops and mixed tree species demonstrating improvements from roughly 40% to over 89%. Ultimately, it establishes practical, decision-oriented guidelines for overcoming transient illumination and co-registration errors, advancing high-fidelity quantitative monitoring and stress detection in complex forest ecosystems. Full article

Maintaining rice terraces as living agricultural heritage requires distinguishing interruptions of paddy production from changes in settlement fabric, yet evidence on associated factors remains limited in the Qinba transition zone. This study compared terrace abandonment, paddy-to-dryland conversion, and composite change in the settlement architectural landscape in five villages of the Fengyan Ancient Terraces using 260 household questionnaires, 189 changed-plot records, GIS, and binary logistic models. Overall, 77 (29.62%) experienced abandonment, 65 (25.00%) dryland conversion, and 198 (76.15%) composite settlement change; 31 dwellings (11.92%) were vacant or physically deteriorated. A higher share of farmland served by gravity-fed irrigation was associated with lower odds of abandonment and conversion, whereas greater maximum walking time to cultivated plots was associated with higher odds of abandonment. Core-area location showed no stable association with either production outcome, and the exploratory settlement model was not statistically significant. Because the cross-sectional models identify associations rather than motives or causal sequences, the results do not establish that irrigation deterioration caused conversion. By distinguishing cessation of cultivation, adaptive crop conversion, and settlement change, this study identifies functional irrigation continuity and site-specific accessibility as more informative conservation targets than static boundaries alone, while supporting context-sensitive settlement renewal. Full article

Foundation models for Earth observation are commonly transferred to downstream tasks without explicit attention to the preprocessing mismatch between pretraining and target distributions. This study isolates preprocessing mismatch as a controlled experimental factor in transferring the Hydro multispectral foundation model, a Swin Transformer V2 Base encoder pretrained with SimMIM on Sentinel-2 Level-2A water-body imagery, to the Marine Debris and Oil Spill (MADOS) marine pollution benchmark dataset, processed through ACOLITE Rayleigh reflectance and providing 11 of the 12 spectral bands used during pretraining. The two datasets are therefore produced by different atmospheric correction algorithms under different reflectance conventions, and the resulting per-band statistical discrepancy is quantified as the starting point of the analysis. Three preprocessing dimensions are then systematically varied while all other settings are held constant: input normalisation, spectral band adaptation for the missing B09, and encoder transfer mode. From this, four findings emerge. Normalisation mismatch between training and inference is the single largest source of performance degradation, reducing the mean Intersection over Union (mIoU) by 0.458, more than seven times the largest radiometric perturbation tested. A zero-parameter Frobenius-matched column crop of the patch embedding adapts the 12-band pretrained encoder to the 11-band target, at least as effectively as any learnt linear or nonlinear adapter, at a lower cross-seed variance. Under limited target supervision (1433 training patches against an 87.9 million-parameter encoder), freezing the encoder outperforms both fine-tuning in full and random initialisation training from scratch. The gains of partial unfreezing are attributable to augmented training (very simple copy–paste (VSCP) augmentation, exponential moving average (EMA), and test-time augmentation (TTA)) rather than to encoder adaptation. With matched preprocessing, the frozen encoder reaches 0.600 mIoU and matches the published MariNeXt baseline within seed variability. Mechanistic analysis via band-occlusion attribution and feature-space separability shows that input normalisation determines which spectral bands the encoder relies upon, with the magnitude of the shift correlated to the per-band gap between the source and target distributions. Operationally, preprocessing alignment, rather than architectural modification, carries most of the practical effort in transferring a multispectral foundation model to marine surface segmentation. These results are established for a single encoder–benchmark pair under limited target supervision. The mechanism they identify is more portable than the magnitude reported. A frozen encoder’s representations remain bound to the normalisation statistics of its pretraining dataset, so any transfer that departs from these statistics at inference is predicted to degrade sharply in proportion to the per-band distance between the two distributions. Full article

Rice is a staple grain crop central to China’s food security. As the core production region of single-cropping rice, the middle and lower reaches of the Yangtze River face escalating high daytime and nighttime temperatures and compound drought–heat stress amid global warming. The accurate assessment of heat hazards is therefore pivotal for regional yield stability and disaster mitigation. Based on meteorological, remote-sensing, and soil data, together with county-level rice yield statistics from 150 major producing counties spanning 1991 to 2024 (5009 county-year calibration units), we first constructed a composite heat damage index (CHI) by integrating daytime harmful accumulated temperature (Ha), nighttime harmful accumulated temperature (HNa), and the Vegetation Health Index (VHI). We then implemented a gradient boosting decision tree (GBDT) machine learning framework in which yield loss was imposed as a physical constraint. This framework was benchmarked against convolutional neural network (CNN), random forest (RF), and support vector machine (SVM) models, with the Shapley additive explanations (SHAP) method used for attribution analysis and an independent temporal partitioning strategy applied for model validation. The results indicate the following: (1) compared to the single daytime heat damage index, the CHI elevated the yield correlation coefficient from 0.52 to 0.63; (2) with yield constraint calibration, the model attained a balanced accuracy of 92.6% and 94.0% consistency with historical disaster records; (3) regional heat hazard presents a spatial pattern of “high in inland areas and low in coastal areas,” with the heading–flowering stage as the critical sensitive period; and (4) high nighttime temperature accounts for approximately 20% of the model’s relative importance, with higher discriminative sensitivity for high-grade hazards, while the amplifying effect of water deficit on heat stress maintains a stable relative importance of around 16%. In this study, the coupled optimization of traditional assessment paradigms and data-driven approaches is achieved, providing a methodological reference for refined growth stage–specific heat hazard assessment. Its cross-regional portability and independent predictive validity require further validation. Full article

Agriculture is pivotal to China’s dual-carbon goals, yet the nonlinear and spatiotemporally heterogeneous relationships between agricultural carbon emissions (ACE) and their associated socioeconomic, agricultural-production, public-investment, and climatic factors remain poorly understood. Here, we compile a multi-source provincial ACE inventory covering cropland use, rice cultivation, and livestock production for 2000–2023 and combine spatial trend and autocorrelation diagnostics with explainable machine learning using XGBoost–SHAP and geographically and temporally weighted regression (GTWR). We find that national ACE increased from 254.48 to 270.25 Mt, while emission intensity fell by 59.6%, indicating that the carbon efficiency of agricultural production improved substantially, although total emissions did not achieve an absolute decline. ACE exhibited persistent spatial imbalance, significant spatial clustering, and gradual diffusion of high-emission areas. Model benchmarking showed that XGBoost achieved the best overall performance, with a mean cross-validated R² of 0.9131 and an independent temporal-test R² of 0.7056. SHAP importance aggregated across repeated cross-validation identified agricultural public investment (21.3%) and urbanization rate (17.9%) as the two leading factors, followed by precipitation, temperature, agricultural industrial structure, and industrial agglomeration level, with the same six factors consistently identified by the three tree-based models. SHAP dependence plots further revealed nonlinear relationship patterns and approximate transition locations around an agricultural public investment level of CNY 11.84 billion, an urbanization rate of 54.69%, annual precipitation of 698.4 mm, and annual temperature between 13.01 and 22.17 °C. GTWR provided stronger statistical evidence of spatiotemporal nonstationarity for urbanization rate, temperature, and industrial agglomeration level, whereas the coefficient patterns of agricultural public investment, agricultural industrial structure, and, particularly, precipitation received more limited local statistical support. These findings provide support for region-specific agricultural carbon mitigation by improving agricultural input-use efficiency, optimizing the allocation of public investment, facilitating the transition toward low-emission crop and livestock management, and integrating climate adaptation with sustainable agricultural production and food security. Full article

European agriculture is increasingly required to remain economically viable while improving resource-use efficiency and reducing exposure to rising input and energy costs. This study examines the relationship between farm profitability, asset-use efficiency and energy-cost intensity across different European farming systems. The analysis is based on FADN/FSDN economic indicators reported by country and farm type for the period 2015–2023. Four types of farming were analysed separately: field crop farms, dairy farms, mixed farms and wine farms. Return on assets was used as the dependent variable, while asset turnover, economic size, fixed asset share, productivity indicators and energy intensity were included as explanatory variables. The empirical analysis combined farm-type-specific panel models with several robustness checks, including common time effects, first-difference specifications and a unified farm-type interaction model. The results show that asset turnover is the most consistently significant correlate of profitability, with a positive association across all the analysed farming systems. Energy intensity is negatively associated with profitability, with the most robust evidence identified in field crop and mixed farms. In dairy and wine farms, the association between energy intensity and profitability was not statistically significant at the 5% level in the main fixed-effects models. Fixed asset share was negatively and statistically significantly associated with profitability only in mixed farms. These findings indicate that profitability-related relationships differ across farming systems and that asset-use efficiency and energy-cost exposure should be interpreted within the structural and production characteristics of each farm type. The study does not aim to identify causal effects, but provides comparative evidence on conditional associations between profitability, accounting-based efficiency indicators and resource-use indicators in European agriculture. Full article

Insect-proof nets provide a non-chemical alternative within integrated pest management (IPM), but their effectiveness against thrips may depend on crop species, season, mesh size, and treatment-associated microclimate conditions. We compared an uncovered and unframed control (CK) and 40-, 60-, and 80-mesh insect-proof nets on thrip abundance in eggplant and potato across two planting seasons in a single-span plastic greenhouse. Weekly plot-level thrip surveys were conducted using a five-point sampling method. Thrips were not identified to species, and the response variable therefore represented total thrip abundance. Treatment effects were estimated with negative binomial (NB1) generalized linear mixed models incorporating a random intercept for plot and an offset for sampling effort; microclimate variables were analyzed with linear mixed models, and exploratory analyses additionally adjusted the NB1 model for measured microclimate covariates to examine whether accounting for these variables altered the estimated net-cage treatment associations. Net-cage treatments generally reduced thrip abundance relative to the control, but the magnitude and consistency of suppression varied markedly with crop and season. The 60-mesh net showed the most consistent direction of suppression, with estimated reductions ranging from 36.0% to 44.4% across the four crop–season contexts; this suppression was statistically supported in both eggplant contexts, but only marginal in the two potato contexts, and this directional consistency did not imply universal superiority in all contexts. The 80-mesh net produced stronger suppression in some cases but was inconsistent, with no detectable suppression in eggplant during the second season. Treatment-related differences in 7-day temperature and relative humidity were generally statistically uncertain, whereas illuminance showed some treatment-associated variation. Adjustment for the measured microclimate covariates did not consistently attenuate the estimated treatment associations, indicating that the measured variables did not provide a simple statistical explanation for the observed context-dependent pattern. Overall, the 60-mesh treatment showed the most consistent estimated suppressive direction under the tested conditions, whereas increasing nominal mesh count did not produce a simple monotonic increase in suppression. The microclimate-adjusted analysis was exploratory and did not provide a causal decomposition of net-cage effects. Full article

Agricultural nonpoint source pollution mainly causes lake eutrophication in China, largely affected by variations in crop distribution. To analyze the multiscale relationships between the long-term evolution of cropping patterns and lake water quality at the macro scale, this study analyzed nationwide datasets for 2000 and 2020 covering 420 relatively large lakes. We systematically examined the spatial restructuring of six major food and cash crops—wheat, rice, maize, soybean, peanut, and rapeseed—and evaluated their multiscale associations with lake total nitrogen (TN) and total phosphorus (TP) concentrations and how these associations changed over time. The results showed the following: (1) From 2000 to 2020, the spatial distributions of the six major crops underwent substantial restructuring. The dominant production areas of rice, wheat, and maize were maintained or further reinforced, whereas soybean, rapeseed, and peanut exhibited varying degrees of regional redistribution and localized concentration. (2) Lake water quality differed between the flood and non-flood seasons. TN exhibited pronounced seasonal differences between the two study periods, whereas temporal changes in TP were generally limited; both nutrients nevertheless showed marked regional heterogeneity among the five major lake regions. (3) The crop–water quality relationship exhibits significant scale dependence and crop-specific variations. The XGBoost model demonstrated a certain degree of out-of-field (OOF) predictive capability for both TN and TP, with OOF R² values of 0.448 and 0.447, respectively. For TN, the highest OOF R² values were observed in the 1000–2000 m buffer zone in both 2000 and 2020; the optimal prediction scale for TP shifted from 1000–2000 m in 2000 to 2000–5000 m in 2020. SHAP results showed that corn maintained a high and relatively stable predictive importance in the TN model, followed by wheat, peanuts, and rice; in the TP model, corn and rapeseed were the crop predictors with the highest relative SHAP importance. PDP results further indicate that there are generally nonlinear or non-monotonic relationships between different crop coverage proportions and TN and TP. (4) Pronounced spatial heterogeneity was observed across the five lake regions. The Eastern Plain Lake Region was characterized by associations involving multiple crops, whereas maize was the most prominent crop in the Northeast Plain and Mountain Lake Region. In the Inner Mongolia–Xinjiang Plateau Lake Region, maize predominated, with wheat and rapeseed also showing notable importance. In the Tibetan Plateau Lake Region, TN was associated with multiple crops, whereas TP was primarily related to maize and rapeseed. The Yunnan–Guizhou Plateau Lake Region exhibited particularly strong scale-dependent differences. This study provides a nationwide analytical framework for comparing the scale differences and regional variations in the statistical associations between the spatial distribution of crops and lake water quality at the specific crop level. The findings can provide a scientific basis for formulating differentiated agricultural nonpoint source pollution control strategies that are adapted to the evolving characteristics of crop planting structures. Full article

Water resources are essential for sustainable agricultural development, and accurate crop water requirement prediction is important for improving irrigation efficiency and optimizing water allocation in irrigation districts. This study focused on the Chushandian Irrigation District in Henan Province, China. Reference evapotranspiration (ET₀) was calculated using the FAO Penman–Monteith equation, and the monthly crop water requirements (ET_C) of wheat, peanut, rapeseed, corn, rice, and vegetables were estimated using crop coefficients (K_c). XGBoost feature importance, Pearson correlation, Mantel, and SHAP analyses were used to examine the meteorological drivers of crop water requirement. Atmospheric pressure showed high nonlinear predictive importance, whereas mean air temperature, relative humidity, and sunshine duration exhibited more consistent physical and statistical relationships with crop water requirement. A process-informed TCN–Transformer framework was then developed for joint and crop-specific prediction. The TCN module extracted local temporal variations, while the Transformer module captured long-term dependencies. In the joint prediction task, the proposed model achieved an R² of 0.9487 and an RMSE of 33.24 mm, outperforming the LSTM, GRU, and CNN–LSTM baselines. The crop-specific results further demonstrated that the model effectively represented seasonal variations and periods of relatively high water requirement across the six crops. The proposed framework can support monthly water-allocation planning and seasonal irrigation scheduling in multi-cropping irrigation districts. Full article

Hidden adulteration of stale berries in dried goji berry batches is difficult to detect by manual inspection or batch-level quality assessment. This study developed a high-throughput method for kernel-level spoilage adulteration quantification in dried goji berries. It addressed three practical challenges in the image processing of densely arranged dried-fruits, including scalable label generation for deep learning segmentation without pixel-level manual annotation, separation of densely touching small berries, and full-size quality level distribution map reconstruction. SAM-assisted pseudo-label generation combined with multi-scale image cropping was used to overcome the limitation of manual pixel-level annotation, while YOLO-based instance segmentation was further employed for efficient berry localization in dense scenes. The freshness labels of segmented single berries were assigned by a statistical RGB-HSV grading rule. Specifically, adaptive multi-scale image cropping for segmentation was applied to improve local separability of berries under dense adhesion and occlusion conditions. The crop-level segmentation and grading outputs were subsequently reconstructed into the original image coordinate system to generate complete quality distribution maps. Results showed that YOLO models trained based on the pseudo-labels achieved a precision of 0.953, a recall of 0.951, an mAP₅₀ of 0.960, and an mAP_50-95 of 0.846. The full-size grading map reconstruction method produced a mean duplicate-suppression rate of 4.31%. In the full freshness-grading test dataset, 4850 berries were detected, including 449 stale berries. The mean absolute counting error was 1.61%. The proposed framework reduces manual annotation requirements while enabling berry-level freshness classification and quantitative stale-berry proportion estimation, providing objective information for dried fruit quality screening and adulteration control. Full article

Hydrothermal variability, rising evaporative demand, and drought extremes increasingly threaten crop production in semi-arid regions, yet their relative contributions to maize yield variability remain unclear. Here, we examined maize yield responses to growing-season climatic conditions in Lyuliang City, North China, during 2005–2024 using yield statistics and ChinaMet climate data. Trend analysis, Pearson correlation, candidate regression models, standardized coefficients, and generalized additive models were used to identify dominant climatic predictors. Maize yield showed no significant temporal trend during the study period (Sen’s slope = 0.01 t ha⁻¹ yr⁻¹, p = 0.58), whereas growing-season potential evapotranspiration tended to increase (2.81 mm yr⁻¹, p = 0.06). Diurnal temperature range declined significantly (−0.04 °C yr⁻¹, p = 0.01), and minimum SPEI also decreased significantly (−0.04 yr⁻¹, p = 0.01), indicating intensified extreme dry conditions. Maize yield was most strongly correlated with aridity index (r = 0.72, p < 0.001) and water deficit (r = 0.72, p < 0.001), suggesting that hydrothermal balance explained yield variability better than precipitation or temperature alone. The highest-ranked regression model included aridity index, growing-season temperature, diurnal temperature range, and minimum SPEI, explaining 70% of interannual yield variation. Aridity index was the strongest positive predictor, whereas diurnal temperature range had a significant negative association with yield. Although extreme dry conditions intensified over time, minimum SPEI was not directly associated with annual yield, implying that drought impacts may depend on drought timing, crop phenology, and management buffering. These findings highlight the importance of maintaining favorable hydrothermal balance and reducing risks from increasing evaporative demand and temperature variability to support climate-resilient maize production in Lyuliang City and climatically similar rain-fed semi-arid regions of North China. Full article

Public cattle re-identification benchmarks remain limited in scale and rarely provide leakage-controlled open-set and cross-camera evaluation. Using 25,209 curated crops of 197 Holstein cows recorded by four synchronized side-view cameras in a passageway of a working farm, we investigate viewpoint-related shifts in the embedding space and evaluate cattle re-identification under camera-disjoint gallery/query construction. The extractor combines an animal-pretrained MegaDescriptor Swin-B backbone with multi-level projections, learned gating, quality weighting, and two-level aggregation. Under the strictly inductive identity-disjoint cross-camera protocol without camera centering, the proposed model achieved Rank-1 of 55.7% (95% CI: 47.7–63.8%), Rank-5 of 90.0% (86.3–93.7%), and mAP of 70.2% (66.3–74.1%). Under the same protocol, TransReID ViT-B/16 achieved 38.6%, 77.9%, and 55.9%, while OSNet-x1.0 achieved 12.9%, 47.9%, and 30.7%, respectively. The comparison with the pure-GAP configuration was not statistically resolved (Holm-adjusted p = 0.91). In held-out open-set evaluation, the validation-selected global cosine threshold obtained the strongest mean AUROC, F1, and AUOSCR (63.2%, 66.9%, and 44.4%); the corresponding results were 60.2%, 64.4%, and 38.8% for Mahalanobis rejection and 56.9%, 63.2%, and 41.7% for EVT. Zero-shot evaluation at a second corridor on the same farm yielded Rank-1 of 26.6%, Rank-5 of 41.9%, and mAP of 35.3%. These results establish cross-camera feasibility under the defined protocols and same-farm cross-installation transfer, but do not establish cross-farm, cross-device, seasonal, longitudinal, or full production robustness. Full article