Crop Yield Prediction
in Nepal

Nepal's agricultural sector feeds over 65% of the population, yet farmers and policymakers have almost no data-driven tools to anticipate crop yield fluctuations. Climate variability — erratic monsoons, temperature shifts, unexpected frost — makes season-to-season yield prediction genuinely difficult.

This project built a full machine learning pipeline that predicts district-level crop yields across Nepal by combining historical agricultural statistics with meteorological data from NASA's POWER satellite system. The goal: give farmers and agricultural planners a tool that actually works with Nepal's limited data infrastructure.

The core model evaluation loop — all 8 regression algorithms evaluated on both log-scale and original scale, with the target variable log1p-transformed before training:

import numpy as np
from sklearn.linear_model import LinearRegression, Ridge, Lasso
from sklearn.tree import DecisionTreeRegressor
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor
from sklearn.svm import SVR
from sklearn.metrics import mean_squared_error, r2_score

# Target: log1p transformed for better regression performance
y = np.log1p(df_encoded['h/ha_yield'])

# Evaluate each model on log-scale and original scale
def evaluate_model(model, name, X_train, X_test, y_train, y_test):
    model.fit(X_train, y_train)
    y_pred_log = model.predict(X_test)

    # Convert back to original scale
    y_pred_orig = np.expm1(y_pred_log)
    y_test_orig = np.expm1(y_test)

    r2  = r2_score(y_test_orig, y_pred_orig)
    mse = mean_squared_error(y_test_orig, y_pred_orig)
    return r2, mse

models = [
    ("Linear Regression",          LinearRegression()),
    ("Ridge Regression",            Ridge()),
    ("Lasso Regression",            Lasso()),
    ("Decision Tree",               DecisionTreeRegressor(random_state=42)),
    ("Random Forest",               RandomForestRegressor(random_state=42)),
    ("Gradient Boosting",           GradientBoostingRegressor(random_state=42)),
    ("Support Vector Regression",   SVR()),
]
# XGBoost trained separately using xgb.train() with early stopping
# Best result → R² = 0.8175, MSE = 0.2031

All models evaluated on a held-out 20% test set. Metrics computed on the original scale (after inverse log1p transform). XGBoost achieved the best R² and lowest MSE:

The paper presents the full methodology, dataset construction, feature engineering decisions, and model evaluation framework. It contextualizes the work within Nepal's agricultural data scarcity and argues for satellite-derived meteorological inputs as a scalable alternative to ground station networks.

Presenting at ICRTAI 2025

The final XGBoost model was wrapped in a Streamlit web application that lets users input district, crop type, and meteorological parameters to get a predicted yield. Designed to be accessible to agricultural officers and researchers without requiring any coding knowledge.