AgbajeCity/clisense

ML-Powered Predictive Climate Intelligence & Early Warning System for Smallholder Farmers in Rural Nigeria — ALU Mission Capstone 2026 # Clisense - Osun River Corridor Pilot (Osogbo, Osun State) ML-powered predictive climate intelligence and early warning for smallholder farmers in **Osogbo, Osun State, Nigeria**, in the Osun River flood corridor. The system benchmarks four machine-learning architectures on flood classification and serves the champion through a REST API and a browser-based forecast interface. - **Forecast interface + API**: clisense.onrender.com - **API docs (Swagger)**: clisense.onrender.com - **Health check**: clisense.onrender.com - **Repository**: github.com ## What it does One prediction task for the Osogbo gauge / Osun River corridor: 1. **Flood classification** - is a given day a flood-risk day? (binary) Four architectures are benchmarked against a naive persistence baseline: **Random Forest**, **XGBoost**, **Decision Tree**, and a **Multi-Layer Perceptron**, on real historical rainfall (see Dataset below). **MLP Neural Network** wins (highest F1) and is the deployed champion. ## Results (2024 held-out test set, 366 records) | Model | Flood accuracy | Flood F1 | |-------|---------------:|---------:| | Random Forest | 91.8% | 82.8% | | XGBoost | 90.2% | 79.3% | | Decision Tree | 86.9% | 73.3% | | **MLP Neural Network (champion)** | **91.8%** | **83.9%** | | Persistence baseline | 89.6% | 79.1% | Random Forest and MLP tie on accuracy (91.8%); champion selection uses F1, where MLP leads. Discharge features (current discharge and its 1- and 3-day lags) account for ~74% of the champion's importance (permutation importance, since MLP has no native `feature_importances_`). Full metrics are in `models/benchmark_metrics.json`; figures are in `assets/`. ## Architecture `app/model_core.py` is the single source of truth: it assembles the dataset, engineers the features, and trains the champion flood-classification model (whichever architecture the benchmark found best - see `CHAMPION_FLOOD_MODEL`). `app/benchmark …