Crop recommendation from where you are — pick a point on a map, get the crops that will actually grow there.

---

Results

Built for the TSA Software Development event, 2024–2025 season.

Placement	Competition
🥇 1st place	Texas TSA State Conference — Software Development
Semifinalist	TSA National Conference

The problem

TSA's theme statement for the season:

Develop a program that enhances the environment and/or agriculture to be more sustainable and efficient.

Planting the wrong crop for your soil and climate wastes a season. The information needed to make that call — nitrogen, phosphorus, potassium, pH, and the local temperature, humidity and rainfall — is scattered across agronomy references and weather services, and it isn't obvious how to combine them.

Green Acres collapses that into one interaction: click where you farm, get a ranked crop recommendation.

How it works

  Leaflet map            Open-Meteo               Flask API              Keras model
  click a point   ──►    temperature      ──►     /predict       ──►     22-class
  or type lat,lon        humidity                 7 features             softmax
                         rainfall                                            │
                                                                             ▼
                                                                    ranked crops +
                                                                    growing notes

1. Pick a location. A Leaflet map over OpenStreetMap tiles; click anywhere or type latitude, longitude. The coordinates are validated and held in localStorage.
2. Fetch conditions. Open-Meteo supplies temperature, relative humidity and rainfall for that point — no API key required.
3. Predict. Those three plus four soil values go to the Flask service as seven features. A Keras model returns a softmax over 22 crops, and a TFLite build of the same model runs alongside it for comparison.
4. Explain. The top crops are paired with growing notes from backend/cropTexts/.

The model recommends across: apple, banana, blackgram, chickpea, coconut, coffee, cotton, grapes, jute, kidney beans, lentil, maize, mango, moth beans, mung bean, muskmelon, orange, papaya, pigeon peas, pomegranate, rice, and watermelon.

Running it

The frontend is static and the model service is a local Flask app. You need both.

1. Start the prediction service

pip install flask flask-cors numpy tensorflow
cd frontend/response
python ai.py

It serves on http://127.0.0.1:5000. ai.py loads my_model.keras and optimized_model.tflite from its own directory, so start it from frontend/response/ or the model paths won't resolve.

2. Open the frontend

Open frontend/land/land.html in a browser, choose a location, then continue to the results page. CORS is already enabled on the Flask app, so file:// works.

3. Check it directly (optional)

curl "http://127.0.0.1:5000/predict?nitrogen=90&phosphorus=42&potassium=43&temperature=21&humidity=82&ph=6.75&rainfall=203"

Returns the echoed inputs plus keras_prediction and tflite_prediction — each a 22-element probability vector aligned to the crop list above.

Layout

frontend/
├── land/            map + coordinate entry (Leaflet, localStorage)
└── response/
    ├── ai.py                 Flask service — GET /predict
    ├── ai.html / .js / .css  results page, Open-Meteo fetch, prediction display
    ├── my_model.keras        trained classifier
    └── optimized_model.tflite  TFLite build of the same model
backend/
├── cropTexts/       growing notes for each of the 22 crops
└── logo.png
apiLinks.txt         data sources and the fields we needed from each

Known limitations

Honest about what a two-day competition build didn't get to:

• Soil values are hardcoded. Nitrogen, phosphorus, potassium and pH are fixed at 90 / 42 / 43 / 6.75 in ai.js — only temperature, humidity and rainfall come from your location. We couldn't find a free soil API with the coverage we needed (see apiLinks.txt). Predictions therefore vary with climate but not with your actual soil.

Fixed since the competition

• The crop labels were in the wrong order. The model's 22 outputs follow the alphabetical label encoding of its training dataset — the same 22 labels that name the files in backend/cropTexts/ — but allCrops listed them in an arbitrary, pluralised order. So a prediction could be reported as the wrong crop, and the growing-notes lookup (which builds a filename from the display name) 404'd for most of them ("Rices" → rices.txt, which doesn't exist). The list is now in label-encoding order and every name maps to an existing notes file.
• Predictions ran on zeroed weather. makeAIRequest() called getWeather() and then read temp / humidity / rainfall from localStorage on the very next line — but getWeather() resolves asynchronously, and land.js clears localStorage on load. Every prediction was made with temperature, humidity and rainfall all 0. getWeather() now returns the conditions and is awaited.
• The second and third suggestions were wrong. ai.js read allCrops[ind + 2] and allCrops[ind + 4] — neighbours in the array, not the next-highest probabilities — and rendered undefined when the top crop sat near the end of the list. It now sorts the prediction vector and takes the top three.
• GET /predict now validates its inputs and returns 400 on missing or non-numeric parameters and 500 on a failed prediction, instead of 200 with an error body.
• / now serves the results page as a static file; render_template raised TemplateNotFound on every request because there is no templates/ directory. Model paths also resolve relative to ai.py, so the service can be started from any directory.
• Leaflet upgraded to 1.9.4 over HTTPS. It was pinned to 0.7.3 over plain HTTP, which browsers block as mixed content on the HTTPS GitHub Pages site — the map never loaded there.
• The repo root served a "Hello World" placeholder to anyone visiting the Pages site. It now redirects to the app.

Team

Built by three students for TSA Software Development, 2024–2025.

	Role	Grade
Stavya Palassery	Team Lead, Frontend	11
Vamsi Yadagiri	AI Model Developer	9
Suchay Kommisetty	Database and API Manager	11

_{First coding experiences, for the record: a Tic Tac Toe site in HTML/CSS/JS; Minecraft Hour of Code; and Pre-Algebra homework automated with Java and some Python.}