Crop Recommendation System using Machine Learning (Soil and Climate Dataset)

Choosing the right crop for a field is one of the highest-stakes decisions in farming, yet it is still made mostly by tradition, word of mouth or last season's prices. A soil health card gives the farmer NPK and pH numbers, but those numbers are hard to translate into a crop choice, and a mismatch between crop and soil-climate conditions means wasted fertilizer, poor germination and a season of lost income. Machine learning can bridge this gap: crop-requirement knowledge, compiled over years of agronomy into reference tables, can be encoded in a dataset and learned by a classifier that maps seven soil and climate measurements to the most suitable of 22 crops. This project builds exactly that pipeline — the public Crop Recommendation Dataset (2,200 records, 22 balanced crop classes) is explored, split and used to train a Random Forest classifier, compared against SVM, k-NN and Naive Bayes baselines, evaluated on a held-out test set, and served through a web demo where any soil-climate profile returns a ranked recommendation.

1. The 2,200-row Crop Recommendation Dataset is loaded: 7 numeric features (N, P, K, temperature, humidity, pH, rainfall) and a label column with 22 crop classes, 100 rows each, no missing values.
2. Exploratory analysis checks feature distributions and class balance; the data is split into a stratified 80/20 train/test split (1,760 train, 440 test) so every crop is represented in both sets.
3. A Random Forest classifier is trained on the training split, with tree count and maximum depth tuned; SVM, k-NN and Naive Bayes are trained on the same split as baselines.
4. All models are evaluated once on the held-out test set: accuracy, per-class precision/recall/F1 and a 22x22 confusion matrix.
5. The winning model is saved as a .pkl file together with the exact feature schema (names, order, valid ranges).
6. In the web demo, the 7 entered values are validated against the dataset's observed ranges, passed to the saved model, and the app returns the predicted crop with class probabilities and the top-3 ranked crops.
7. The demo additionally shows each input next to the predicted crop's typical profile, giving a plain-language "why this crop" explanation.

Tech stack:

• Python 3, scikit-learn (Random Forest)
• Pandas, NumPy (data handling)
• Matplotlib, Seaborn (EDA & evaluation plots)
• Jupyter Notebook (training & evaluation)
• HTML5, CSS, JavaScript (recommendation web demo)
• Crop Recommendation Dataset (Kaggle, atharvaingle)
• Git

Dataset & model details

• Dataset: Crop Recommendation Dataset (Kaggle, published by atharvaingle) — 2,200 rows, 7 numeric features (N, P, K, temperature, humidity, ph, rainfall), target column "label" with 22 crop classes: rice, maize, chickpea, kidneybeans, pigeonpeas, mothbeans, mungbean, blackgram, lentil, pomegranate, banana, mango, grapes, watermelon, muskmelon, apple, orange, papaya, coconut, cotton, jute, coffee. 100 samples per class (perfectly balanced), zero missing values. Class list verified against the dataset file itself.
• Task: 22-class tabular classification; input = 7 soil/climate numbers, output = predicted crop label plus a probability distribution over all 22 crops.
• Model: Random Forest classifier (ensemble of decision trees; number of trees and max depth tuned in the notebook). SVM, k-NN and Naive Bayes are trained as baselines on the same stratified split. Training (offline, notebook) and inference (live, web app) are separate phases.
• Metrics: Test accuracy, per-class precision/recall/F1 and a 22x22 confusion matrix on the stratified held-out test split. Accuracy is reported as a design target (≥98% expected — this benchmark is well-separated in published notebooks), never as a measured claim; the built-to-order training run's actual figures are documented in the delivered report.

• [Random Forest crop classifier] Ensemble of decision trees trained on the 2,200-row dataset; tree count and depth tuned in the notebook. Outputs the predicted crop plus a full probability distribution over all 22 classes.
• [Live recommendation web demo] Single-file web app: enter the 7 soil/climate values and get the recommended crop instantly, with suitability scores and runner-up crops shown as bars.
• [Explainability breakdown] The demo compares your inputs against the predicted crop's typical profile feature-by-feature, so the recommendation is traceable rather than a black box.
• [Full training & evaluation notebook] Data loading, EDA, stratified train-test split, model training, hyperparameter notes and every evaluation plot in one reproducible Jupyter notebook.
• [Baseline model comparison] SVM, k-NN and Naive Bayes trained on the same split in the same notebook, so the report can justify the Random Forest choice with numbers.
• [Confusion matrix & per-class report] Precision, recall and F1 for each of the 22 crops, making it clear exactly which crop pairs the model confuses.
• [Input validation & range checks] The demo validates every input against the dataset's observed ranges (e.g. pH 3.5-9.9) and flags out-of-range values before prediction.

Optional extensions (configurable at order time, not in the base build)

• [Regional variety mapping + multilingual demo UI] Mapping the 22 generic crop labels to locally grown varieties for a chosen region, and a Hindi/Marathi demo interface alongside English.

Future scope (not implemented)

• Live weather-API inputs, market-price-aware ranking, and a mobile app wrapper.

• Complete training & evaluation Jupyter notebook
• Trained Random Forest model file (.pkl) with the input feature schema
• Working crop-recommendation web demo wired to the trained model
• EDA plots, confusion matrix and per-class metrics
• Project report PDF (background, methodology, evaluation, honest limitations)
• PPT presentation for final review
• Viva Q&A preparation document (Random Forest, overfitting, feature importance, evaluation metrics)