AgriGen-XL: a physics-aware multi-scale generative framework integrating EfficientNet, DeepLabV3+, and YOLOv5 for precision agricultural imaging

Abstract Computer vision has become indispensable to precision agriculture; however, the scarcity and imbalance of labeled data across crops, livestock, and aquaculture domains continue to limit the robustness and generalization of conventional convolutional neural networks (CNNs). We present AgriGen-XL, a physics-aware generative augmentation framework that integrates generative modeling with task-specific CNN backbones, including EfficientNet, DeepLabV3 +, and YOLOv5. The framework combines conditional diffusion models informed by physical and sensor priors with efficient adversarial sampling and couples these generators with domain-adaptive CNNs for classification, detection, and segmentation. AgriGen-XL is evaluated across three representative agricultural domains: (i) crop disease diagnosis and phenotyping from field and UAV imagery, (ii) livestock activity and welfare monitoring using RGB and thermal video, and (iii) aquaculture fish segmentation, counting, and biomass estimation under variable turbidity. The proposed system introduces three key contributions: a cross-domain synthetic data engine that encodes physical priors (e.g., illumination, canopy geometry, and turbidity) and sensor characteristics (e.g., spectral response and thermal noise); a rare-class up-sampling and scene-composition curriculum to address long-tail data distributions; and a precision analytics layer that converts model outputs into parcel-level disease risk, animal-level welfare indices, and cage-level biomass forecasts. Relative to real-only training with standard augmentations and non-physics generative baselines, synthetic augmentation with AgriGen-XL improves downstream performance by + 3–11 absolute points across classification, detection, segmentation, and counting tasks, reduces rare/long-tail class error by 22–38% (e.g., early disease stages and lameness), and improves cross-season crop generalization by + 7.4 accuracy points. Ablation studies show that gains saturate at a 2 × synthetic-to-real data ratio and that physics-informed generators consistently outperform style-based alternatives. Code, configuration files, and example synthetic asset recipes will be made available upon publication. All real datasets are sourced from publicly accessible repositories to support reproducibility.