Crop Yield Prediction using Random Forest

Every sowing season, farmers and agri-traders make expensive decisions — what to plant, how much fertilizer to buy, what volumes to contract — without knowing what the harvest will actually deliver. Traditional yield estimation leans on experience, thumb rules, or coarse district averages that ignore how rainfall, temperature and soil nutrients interact on a specific field. Machine learning fills this gap honestly: historical records of weather, soil NPK and yields contain learnable non-linear relationships, and Random Forest is a natural fit for such tabular farm data because it models interactions (rainfall matters differently at different temperatures) without hand-built formulas. This project builds that system as a complete regression pipeline: public crop-yield records in, expected yield in tonnes per hectare out, with feature importances that show exactly which inputs drive each prediction.

1. The public Kaggle crop-yield dataset (country × year × crop records with rainfall, temperature, NPK, pesticide/fertilizer inputs and yield target) is loaded and cleaned; impossible nutrient values and outlier yields are removed.
2. The crop label is one-hot encoded and numeric features are standardized with a scaler fitted only on the training split, then the data is split 80/20 into train and holdout sets.
3. A RandomForestRegressor with 300 trees trains on bootstrap samples; each tree sees a random feature subset at every split, which decorrelates the trees.
4. Hyperparameters (max depth, minimum samples per leaf) are tuned with 5-fold cross-validation on the training data; the best configuration is refit on the full training set.
5. R², RMSE and MAE are computed on the untouched 20% holdout, with actual-vs-predicted scatter and residual plots for the error-analysis section of the report.
6. At inference time the demo app applies the saved scaler, queries the forest, and renders the mean prediction, the tree-vote interval, per-factor suitability scores and the cross-crop comparison.

Tech stack:

• Python 3.10, scikit-learn (RandomForestRegressor, GridSearchCV, metrics)
• pandas, NumPy (data wrangling, feature tables)
• Matplotlib, Seaborn (actual-vs-predicted, residuals, importance plots)
• Jupyter notebook (buyer-run training and evaluation)
• Flask demo app (prediction form, comparison view)
• joblib (serialized model + scaler artifacts)

• [Yield prediction from field inputs] Enter crop, region, season rainfall, temperature, soil N/P/K and pesticide use; the trained forest returns expected yield in tonnes per hectare with a prediction interval derived from the spread of tree votes.
• [Random Forest regressor, 300 trees] Scikit-learn RandomForestRegressor with tuned depth and leaf settings; hyperparameters chosen by 5-fold cross-validation and a small grid search, all logged in the notebook.
• [Feature importance analysis] Mean-decrease-in-impurity importances rank every input — typically rainfall and temperature on top — giving the report a ready-made explainability section.
• [What-if crop comparison] Replay identical field conditions across rice, wheat, maize, sugarcane and cotton to see which crop the model favours, with suitability flags when inputs leave a crop's optimum band.
• [Buyer-run training and evaluation notebook] One notebook cleans the data, trains, tunes, and computes R², RMSE and MAE on the 20% holdout the trees never saw, with actual-vs-predicted and residual plots.
• [Prediction intervals, not just point estimates] The demo shows the ensemble's 90% interval so the report can discuss uncertainty honestly instead of pretending a single number is certain.
• [Serialized model + scaler] The fitted forest and the preprocessing scaler ship as versioned artifacts, so inference applies exactly the transformation training used.

• Complete source code (data prep, training, tuning, evaluation, inference, demo app)
• Jupyter training and evaluation notebook (buyer-run: clean, train, tune, evaluate, plot)
• Serialized model and scaler artifacts from the reference training run
• Project report PDF (background, methodology, evaluation, error analysis, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (ensembles, bias-variance, impurity importance, CV)
• Setup guide (environment, dataset download, running the notebook, using the demo)

• Predictions are reliable only near the training distribution; for input combinations far outside it (e.g. extreme drought years) the model extrapolates and the interval widens — the report documents this openly.
• The public dataset aggregates at country × year × crop grain, so field-level micro-variation (soil patches, local pests) is invisible to the model; district-level IMD data is listed as future scope, not current capability.
• Weather inputs are season aggregates, not forecasts — the system answers "given this season's conditions, what yield?" not "what will the weather be?".
• The model captures correlation, not agronomic causation; it is a planning aid prototype, not a substitute for agronomist advice.
• Yield is reported in tonnes per hectare; converting to revenue needs prices, which the base project does not model.