A satellite-driven, machine-learning pipeline for monitoring and forecasting crop health in Jammu & Kashmir, India.
This repository contains the complete research pipeline and deployment code for a dissertation project on spatio-temporal crop health assessment using remote sensing and machine learning in the Jammu region of India. The workflow ingests multi-source satellite data through the Google Earth Engine (GEE) API, derives vegetation and hydro-climatic indices over 150 fixed sampling points, models a continuous Crop Health Score (CHS) with classical and deep-learning regressors, validates the best model with interpretability tooling, performs recursive multi-step forecasting, and exposes the trained model through an interactive Streamlit application that accepts any user-supplied coordinate.
The project is intended both as a reproducible scientific artefact for examiners and as a starting point for downstream agricultural decision-support tools.
The study area covers the agricultural landscape of the Jammu region in the Union Territory of Jammu & Kashmir. One hundred and fifty fixed spatial sampling points were generated across the region and tracked monthly from 2017 to 2025. Three open satellite archives are used: Sentinel-2 surface reflectance (COPERNICUS/S2_SR) for optical bands, MODIS MOD11A2 for daytime land-surface temperature, and CHIRPS daily precipitation for rainfall and a one-period lag.
From these sources the pipeline derives seven predictive features — NDVI, EVI, NDMI, NDWI, LST, Rainfall, and Rainfall_lag1 — and a composite CHS target that normalises crop vigour onto a 0–1 scale.
The end-to-end workflow is split into six numbered notebooks under notebooks/, each corresponding to one stage of the dissertation:
| Stage | Notebook | Purpose |
|---|---|---|
| 1 | 01_data_extraction.ipynb |
Authenticates GEE, defines the Jammu AOI and 150 sampling points, and exports monthly Sentinel-2, MODIS, and CHIRPS features. |
| 2 | 02_preprocessing_pca.ipynb |
Cleans the long-format feature table, engineers the rainfall lag, builds the CHS target, and runs PCA for exploratory variance analysis. |
| 3 | 03_model_building.ipynb |
Builds 6-month sequence windows and benchmarks Linear Regression, Random Forest, XGBoost, and an LSTM regressor on a temporal hold-out. |
| 4 | 04_validation_shap.ipynb |
Re-loads the winning model, computes metrics on the 2024+ test set, and produces SHAP-based interpretability plots. |
| 5 | 05_multi_step_forecasting.ipynb |
Implements recursive N-step forecasting on top of the trained model for forward-looking CHS projection. |
| 6 | 06_coordinate_prediction.ipynb |
Demonstrates a single-coordinate prediction workflow that wraps GEE feature extraction around the saved model — the prototype that became the Streamlit app. |
The flow can be summarised as: GEE feature extraction → preprocessing & CHS construction → model benchmarking → temporal validation & SHAP → multi-step forecasting → coordinate-level inference → web app.
All models were trained on observations up to and including 2023 and evaluated on a held-out test set spanning 2024 onwards. Sequences are six-month sliding windows of the seven features, flattened to a 42-dimensional vector for the classical learners and kept as a (6, 7) tensor for the LSTM. Metrics are reported on the unscaled CHS target in its native 0–1 range.
| Model | R² | RMSE | MAE |
|---|---|---|---|
| Linear Regression | 0.589 | 0.1033 | 0.0739 |
| Random Forest (default) | 0.599 | 0.1020 | 0.0741 |
| XGBoost | 0.579 | 0.1045 | 0.0753 |
| LSTM (64 units, dropout 0.2) | 0.476 | 0.1167 | 0.0862 |
| Random Forest (tuned) ⭐ | 0.608 | 0.1009 | 0.0723 |
The tuned Random Forest was retained as the production model and is serialised as Best_model.pkl. On the same temporal hold-out, its Pearson correlation between predicted and observed CHS is 0.78.
Recursive multi-step forecasting was then evaluated on the same hold-out by feeding each prediction back into the input window. The one-step horizon retains the in-sample accuracy, but error accumulates rapidly thereafter — an honest baseline that motivates the future-work direction discussed in docs/METHODOLOGY.md.
| Horizon | RMSE | MAE | R² |
|---|---|---|---|
| t + 1 | 0.1018 | 0.0740 | 0.599 |
| t + 2 | 0.1841 | 0.1374 | −0.312 |
| t + 3 | 0.2166 | 0.1764 | −0.797 |
SHAP analysis on the tuned Random Forest places NDVI and NDMI as the dominant contributors to predicted CHS, with LST and Rainfall_lag1 acting as significant secondary modifiers — a result that is consistent with the agronomic literature for the region.
The deployed application in app/app.py lets a user pick any latitude/longitude in the study area — either by typing coordinates or by clicking on a Folium map — choose a target month and year, and run on-demand prediction. The app pulls a small 500 m buffer of Sentinel-2, MODIS, and CHIRPS imagery for the requested period, computes the seven indices server-side, scores them with the saved Random Forest, and renders a coloured marker (Low / Moderate / High) on the map alongside a results table.
To launch it locally:
streamlit run app/app.pyA live deployment can be hosted on Streamlit Community Cloud or Hugging Face Spaces with no code changes.
# 1. Clone
git clone https://github.com/DivyamVerma22/crop-health-jammu.git
cd crop-health-jammu
# 2. Create an isolated environment
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
# 3. Install dependencies
pip install -r requirements.txt
# 4. Authenticate Google Earth Engine (one-time, see docs/SETUP.md)
earthengine authenticate
# 5. Download the trained model weights
python scripts/download_model.py
# 6. Launch the app
streamlit run app/app.pyTo re-run the research notebooks, open them with Jupyter Lab or in Google Colab. Setup details, including the GEE project-name override, are documented in docs/SETUP.md.
crop-health-jammu/
├── app/ # Streamlit production application
│ └── app.py
├── notebooks/ # Numbered research notebooks (stages 1–6)
├── data/sample/ # Small CSV samples for reproduction
├── models/ # Trained model artefacts (downloaded, see scripts/)
├── scripts/ # Helper scripts (model download, env checks)
├── docs/ # Methodology, setup, and Zenodo guide
├── reports/figures/ # Architecture diagram and dissertation figures
├── requirements.txt
├── CITATION.cff
├── LICENSE
└── README.md
Only lightweight sample CSVs are committed under data/sample/ so the repository stays small and clone-friendly. The full Sentinel-2 / MODIS / CHIRPS feature table can be regenerated from notebooks/01_data_extraction.ipynb once Earth Engine access is configured.
The trained model file (Best_model.pkl, 228 MB) is not stored directly in Git. Two supported retrieval options are documented in docs/SETUP.md: pulling it through Git LFS, or fetching it from an external host (Hugging Face Hub, Zenodo, or Google Drive) using the provided helper script. The recommended path — and the one that produces a citeable DOI — is described in docs/ZENODO.md.
Every notebook fixes the random seed where applicable, all temporal splits are deterministic (≤ 2023 train, ≥ 2024 test), and exact package versions are pinned in requirements.txt. The full methodology — feature definitions, sequence construction, evaluation protocol, and SHAP analysis — is described in docs/METHODOLOGY.md.
If you use this code or methodology in academic work, please cite the dissertation via the metadata in CITATION.cff. GitHub renders this as a one-click "Cite this repository" button. The DOI badge at the top of this README resolves to the Zenodo-archived release of the model and code: 10.5281/zenodo.20259081
This project is released under the MIT License. The Sentinel-2, MODIS, and CHIRPS datasets retain their respective providers' licensing terms.
This work was carried out as part of a bachelor's dissertation in the Jammu & Kashmir region. It builds upon the Google Earth Engine platform, the open Sentinel-2 mission (Copernicus / ESA), NASA MODIS, and the UCSB CHIRPS climate record. Sincere thanks to the dissertation supervisors and reviewers for their guidance. Special thanks to my parents who are my support pillars in every situation.
For questions, issues, or collaboration enquiries, please open a GitHub issue.