EfficientNet-based leaf image analysis and gradient boosting models for predicting sugarcane irrigation management under microclimatic conditions

Deep learning irrigation decision support systems (DSS) enhances sustainable farming and increases crop yield. DSS system increases crop productivity through fertigation and foliar spray management. Existing DSS performs depends on macroclimatic condition and not based on microclimatic conditions, leaf-level nutrient measurement, soil properties and crop growth stages. The above parameters are very important for daily irrigation decisions. Also, most systems are not designed for one specific crop. This article Sugarcane Irrigation Decision Support System (SIDSS) is proposed using microclimatic parameter integrates smartphone-based leaf image acquisition using MATLAB Mobile, classical image-processing techniques, extracts for structural features. The LeafGAN, a generative adversarial networks for dataset augmentation, deep learning model like EfficientNet for leaf nutrient estimation are implemented. The MobileNetV3 for growth-stage classification followed by CatBoost gradient boosting models for irrigation and yield prediction. Visual structural metrics such as texture statistics, edge density, vein density, spatial frequency and local contrast are extracted to characterize leaf surface patterns associated with stress conditions. It is designed to mitigate water stress in the Karan variety of sugarcane, arises due to lack of water, when sugarcane does not get enough water, it releases a hormone called abscisic acid (ABA) secretion and subsequently reduce yield. SIDSS collects real-time leaf nutrient information to monitor plant health and manage irrigation effectively. Experimental result shows that irrigation prediction and yield estimation are 81.2% and 94.6%, respectively. SIDSS performs irrigation management for sugarcane to under different local climate conditions and growth stages.