AI-driven multi-agent framework for smart irrigation and crop health monitoring in Indian rice and sugarcane farming

Abstract

Disease prevention and water management are important to all the crops, particularly rice and sugarcane production in India. The article proposes a reinforcement learning (RL) based intelligent irrigation management system that is capable of optimising water consumption and crop nutrition in response to the changing agricultural climatic conditions. Decentralised reinforcement learning (RL) is used in a network of irrigation agents that utilise soil and microclimate sensor networks to set the terms of water allocation, water use efficiency (WUE) and crop health. At the same time, deep convolutional networks can be used to differentiate between plant stress/disease and leaf images and take applicable proactive actions. It is a framework that incorporates satellite-derived indices (NDVI, EVI, land surface temperature) with local sensor measurements and image-based health measurements through multimodal deep learning. Far-reaching simulations (including Indian climate and crop calendars) demonstrate that the multi-agent system lowers water consumption and preserves the yields and properly notifies stressed plants. The scores of disease detection with plantvillage-based fine-tuned on rice (120 (3 disease types) and 3829 (5 disease types) and sugarcane (2569 images for all disease types, Convolutional Neural Network (CNN) yield results of >98 % accuracy. Crop mapping (rice/sugarcane) Satellite/LSTM-based crop mapping (with Sentinel-1 / Sentinel-2) achieves more than 97 % accuracy. The suggested structure provides a data-driven, scalable system for precision agriculture to enhance the management of irrigation periods and crop health. Simulation experiments show that the RL-based controller can reduce water consumption while preserving optimal soil moisture levels when compared to rule-based irrigation strategies.

Plant phenotyping relevance

画像・衛星・センサーを統合して植物ストレス／病害状態を推定するマルチモーダル基盤が提案され、病害検出性能も評価されているため、植物表現型推定が実質的な構成要素である。

Code and data availability

The paper reports simulation-based experiments using public leaf-image datasets. The only paper-specific public asset explicitly identified is the Kaggle rice leaf diseases dataset (vbookshelf/rice-leaf-diseases) cited as a data source for the rice disease fine-tuning set. No authors' code, trained models, or data dép