crop-recommendation-system

An ML-driven crop suggestion system that takes a farmer's location, pulls historical and real-time soil/weather data, and recommends suitable crops. Combines tabular ML (crop selection from soil + climate) with a CNN (plant disease detection from leaf images).

🏆 AgriHackathon Winner — Dr. N.G.P. Institute of Tech. 🥉 Top 30 / 4000 — Hack Of PI, IIT Kanpur (Semifinalist) 📄 Published — IRJMETS, e-ISSN: 2582-5208 · Paper PDF

---

What it does

   farmer's location (GPS)
            │
            ├──► historical soil/weather (tn.data.gov.in)
            ├──► live weather (API)
            ▼
   feature vector: N, P, K, pH, rainfall, temperature, humidity, soil-type
            │
            ▼
   crop recommendation model  → ranked list of suitable crops
            │
   (parallel)
            ▼
   leaf image upload  → ResNet plant-disease classifier (separate notebook)

Two ML problems, one product:

1. Crop recommendation — given soil chemistry + climate, classify which crop is most suitable. Six classifiers compared in Phase 1/Crop Recommendation.ipynb on a 440-sample held-out split:

   Classifier	Accuracy
   XGBoost	99.32%
   Naive Bayes	99.09%
   Random Forest	99.09%
   Logistic Regression	95.23%
   Decision Tree	90.00%
   SVM	10.68% (no feature scaling)

   XGBoost ships as the recommendation model. The dataset is well-separated by N/P/K + climate features, which is why most non-linear classifiers cluster near 99%; SVM's collapse is the canonical "linear kernel without scaling" failure — kept in the comparison as a deliberate negative result.

2. Plant disease detection — given a leaf photo, classify healthy vs. diseased. Separate ResNet-based CNN training pipeline in Static Website/The-Farming-Company/notebooks/plant-disease-classification-resnet-99-2.ipynb, served via TensorFlow Lite for mobile deployment. The two ML pipelines are independent, not glued in code — the product wraps both behind a Flask UI.

Tech stack

• ML: scikit-learn (Decision Tree, NB, Random Forest, XGBoost), PyTorch / ResNet for vision, TensorFlow Lite for deployment
• Backend: Flask
• Frontend / mobile: static web app + TF Lite mobile model
• Storage: Firebase Firestore
• Data sources: tn.data.gov.in (historical), weather API (live)

Project layout

Phase 1/                    crop-recommendation models — 4 classifiers compared
   ├── Crop Recommendation.ipynb
   ├── DecisionTree.pkl, NBClassifier.pkl, RandomForest.pkl, XGBoost.pkl
   └── Crop_recommendation.csv
Phase 2/                    end-to-end notebook + soil/place lookup tables
   ├── Automated_Crop_Recommendation_System.ipynb
   ├── ACSS PAPER.pdf
   └── crops.csv, cropSoil.csv, placeSoil.csv, …
Phase 3/                    final deck / writeup (VyaVas.pdf)
Static Website/             Plant-disease pipeline assets (training notebooks + model components)
   └── The-Farming-Company/
        ├── app/utils/                    ResNet9 architecture (model.py),
        │                                  disease/fertilizer info dictionaries
        ├── Data-raw/                     raw districtwise yield data
        ├── notebooks/                    training notebooks: plant-disease ResNet,
        │                                  data prep, final recommendation model
        └── models/                       trained model artifacts (PyTorch + TF Lite)
Hackathons/                 pitch deck + demo recording (AgriHackathon, HACK OF PI)

Recognition

• Winner, AgriHackathon — Dr. N.G.P. Institute of Tech.
• Semifinalist (Top 30 of 4000), HACK OF PI — IIT Kanpur
• Published in IRJMETS (e-ISSN 2582-5208) — paper

Notes

This repository contains the full set of artifacts from a multi-phase project (research → notebooks → static deployment). It's preserved as the work-of-record rather than a clean library.