RAVA: a robust recurrent active vision agent for plant disease diagnosis under severe environmental noise
Abstract Deep learning models have achieved near-perfect accuracy in plant disease classification within controlled laboratory settings; however, their deployment in real-world agricultural environments is severely hindered by the “deployment gap”—a critical vulnerability to environmental corruptions such as sensor noise, motion blur, and occlusion. To bridge this gap, we propose the Recurrent Active Vision Agent (RAVA), formulating the disease detection task as a Partially Observable Markov Decision Process (POMDP). Unlike passive Convolutional Neural Networks (CNNs) that process images globally, RAVA mimics the active inspection behavior of human agronomists. Our architecture integrates a lightweight ResNet-18 backbone with a Recurrent Neural Network (RNN) and a Spatial Transformer Network (STN). Driven by Proximal Policy Optimization (PPO), the agent learns a sequential policy to intelligently navigate and zoom in on informative “glimpses,” effectively bypassing background clutter. To stabilize the reinforcement learning process and enforce noise-invariant feature representations, we introduce a hybrid objective incorporating Supervised Contrastive Learning (SupCon). Comprehensive experiments on a combined PlantVillage and PlantDoc dataset demonstrate RAVA’s overwhelming superiority under extreme conditions. In a “Severe Degradation” stress test, standard ResNet-50 accuracy collapses to 26.3%, whereas our active agent maintains a robust 77.0%. Under extreme noise and occlusion, RAVA preserves 54.6% accuracy compared to the baseline’s 17.2%. Notably, this resilience is achieved with merely ∼12M parameters—significantly fewer than large-scale Vision Transformers—proving that active visual attention, coupled with contrastive learning, offers a computationally efficient and highly robust pathway for field-ready precision agriculture.