logicmatters06/agrimitra-crop-model

AgriMitra Crop Recommendation Model

A scikit-learn RandomForestClassifier that recommends a crop from seven soil and climate inputs:

N, P, K, temperature, humidity, ph, rainfall

The model predicts one of 51 crop labels. AgriMitra only displays the 22 crops that also have a full 15-section guide in its knowledge base; the remaining labels are still part of the trained classifier and are exposed through the response's coverage value.

Files

File	Description
crop_recommendation_model.pkl	Joblib artifact (RandomForestClassifier, 51 classes)
metrics.json	Holdout metrics and top confusions
confusion-matrix.png	Full 51 × 51 holdout confusion matrix

Holdout metrics

Evaluated on the same 80/20 stratified holdout split (random_state=42) used in crop-ml/trainingModel.ipynb.

Metric	Value
Test samples	1,320
Correct	1,316
Accuracy	0.99697
Macro precision	0.99635
Macro recall	0.99608
Macro F1	0.99607
Weighted F1	0.99696

Top confusions: jute → rice (2), blackgram → maize (1), mothbeans → pigeonpeas (1).

Training data

The 6,596-row merged dataset is in the AgriMitra repository (crop-ml/merged_crop_dataset.csv) and the source corpus is on Hugging Face: logicmatters06/agrimitra-dataset.

The first 22 crop classes (100 rows each) match the widely used public Kaggle Crop Recommendation dataset. The remaining 29 classes are additional Indian crop records compiled for this project. Verify the original dataset licences before redistributing the training CSV.

Usage

import joblib
import pandas as pd

model = joblib.load("crop_recommendation_model.pkl")
features = ["N", "P", "K", "temperature", "humidity", "ph", "rainfall"]
row = pd.DataFrame([[90, 50, 50, 27, 75, 6.5, 1200]], columns=features)

proba = model.predict_proba(row)[0]
labels = model.classes_
ranked = sorted(zip(labels, proba), key=lambda item: item[1], reverse=True)
print(ranked[:5])

The artifact was trained with scikit-learn 1.8 and is loaded here with scikit-learn 1.9.1. Loading emits an InconsistentVersionWarning; the AgriMitra backend logs that warning rather than hiding it. Pin the runtime to the version in requirements.txt for reproducible results.

Limitations

• The model is trained on tabular agronomic ranges, not local soil tests.
• It does not account for season, irrigation method, market price, or variety.
• Rare classes have very few samples; the aggregate holdout accuracy is dominated by the larger classes.
• Recommendations are suggestions, not a substitute for local extension advice.
• Pesticide and fertilizer doses are intentionally not produced by this model.

Licence

The model artifact and training code are released under the MIT licence. The source PDFs and derived corpus remain subject to their original authors' rights; see the dataset card for details.