Osinmen/agrisense_backend

A FastAPI-powered machine learning backend for rainfall forecasting and crop advisory in Ondo State, Nigeria. # 🌧️ AgriSense — Rainfall Forecasting & Agricultural Advisory API > A hybrid machine learning system for seasonal rainfall prediction and crop advisory generation for smallholder farmers in Ondo State, Nigeria. --- ## 📌 Overview AgriSense is a REST API backend for a seasonal agricultural planning system targeting smallholder farmers in Ondo State, Nigeria. The system forecasts monthly rainfall up to 12 months ahead using a multivariate CatBoost gradient boosting model trained on 42 years of NASA POWER climate data (1984–2025), and translates forecast outputs into actionable crop advisory recommendations covering planting windows, expected pest and disease pressures, and estimated crop yield ranges. A key innovation is the **Macro-Seasonal Adaptive Winsorization Pipeline (MSAW)** — a novel post-processing layer that dynamically constrains recursive forecast outputs using historically calibrated monthly volumetric tiers, preventing autoregressive error propagation while preserving full predictive freedom during the main wet season. --- The system follows a three-tier architecture: **Tier 1 — Mobile Frontend (Flutter)** - Cross-platform iOS and Android application - English and Yoruba language support - OpenMeteo API integration for real-time daily weather display - AgriSense seasonal forecast and advisory display **Tier 2 — FastAPI Backend (This Repository)** - CatBoost Multivariate model inference - MSAW post-processing layer - Crop advisory and yield estimation engine - Firebase Firestore data persistence - Monthly scheduled forecast generation with local push notification trigger **Tier 3 — Data and Storage Layer** - NASA POWER climate data (training source) - Firebase Firestore (forecast and advisory storage) - CatBoost model artefact (.cbm file) - 42-year historical climate dataset (inference context) ## 🤖 Model Performance The CatBoost Multivariate model was selected as the deployment model following emp …