SasyaVedha: A Soil-Aware and Climate-Resilient Agronomic Decision Support System for Indian Farms

Abstract: Indian farmers routinely have to make five separate decisions in a single season — what disease is on the leaf, how much yield to expect, when to irrigate, what to feed the soil, and how a coming drought or flood will change all three — and most digital tools answer only one of these at a time. This paper presents SasyaVedha, a web-based agronomic decision support system that brings crop disease identification, yield prediction, irrigation scheduling, fertilizer advisory, and climate-event impact simulation into a single Flask-and-React platform, alongside an enterprise mode for multi-farm operations with role-based access control and an append-only audit log. Disease identification uses a MobileNetV2 transfer-learning CNN trained on the PlantVillage dataset, with a deliberate fallback to a curated disease database when trained weights are unavailable so the feature degrades gracefully rather than failing outright. Yield and irrigation are each modelled with a scikit-learn ensemble — Random Forest for yield, Extra Trees for daily net irrigation need — built inside a Column Transformer pipeline that one-hot encodes crop and soil type and passes four live weather features through unchanged, with weather drawn at request time from the Open-Meteo forecast API. Fertilizer guidance and climate-event projection are implemented as heuristic engines grounded in ICAR-style NPK target ranges and per- state, per-crop vulnerability tables rather than as learned models, a choice made specifically to keep their recommendations traceable to known agronomic thresholds. The system is backed by a file-based SQLite schema covering users, sessions, farms, farm assignments, and an activity log, and the frontend is localized into six Indian languages. We describe the architecture, the methodology behind each of the seven functional modules, the role-based security model, and the engineering trade-offs made along the way, including the current reliance on a bootstrap- synthesized training set for the regression models pending real farm records. We close with the limitations this implies, and the roadmap items identified during development.

Keywords: Agronomic Decision Support; MobileNetV2; Random Forest; Extra Trees; Climate Impact Simulation; Role- Based Access Control; Open-Meteo; Precision Agriculture