What question did this study set out to answer?

This study aims to estimate wheat production using satellite and meteorological data through machine learning.

June 22, 20260 citationsOpen Access

ML-Based Wheat Production Estimation using Two Decades of Remote Sensing & Meteorological Time-Series Data across Five Major Indian States

RPRashi Pareek HSHari Shanker Srivastava

Key Points

This study aims to estimate wheat production using satellite and meteorological data through machine learning.
Integrated satellite-derived data from MODIS and GLDAS over two decades (2003-2025).
Applied decision-rule based thresholding to NDVI for classifying wheat and non-wheat areas.
Utilized tree-based machine learning algorithms (DTR, RFR, ETR, XGBR) for production prediction.
XGBR model outperformed others with an average error of 5.6% compared to 10.3% for DTR, 8.3% for RFR, and 7.4% for ETR.
Predicted production statistics were validated against UPAg data, confirming robustness of XGBR.
The integration of remote sensing with machine learning shows promise for agricultural forecasting challenges.

Abstract

Estimating production prior to harvest is a crucial practice that enables both the government and farmers to manage the harvest process effectively. This study focuses on predicting wheat production in five states of India by integrating satellite-derived data with machine learning techniques. The temporal coverage of the study extends from 2003 to 2025, incorporating crop production statistics and meteorological parameters from visual crossing collected during the peak wheat growth window in its phenological cycle, i.e., from November to March. Terra MODIS data products such as NDVI, LAI and thermal anomaly, along with soil moisture from GLDAS, were obtained. For classifying wheat and non-wheat areas, decision-rule based thresholding was applied to NDVI values. Subsequently, wheat production was predicted using tree-based ML algorithms, such as Decision Tree Regressor (DTR), Random Forest Regressor (RFR), Extra Trees Regressor (ETR), and XGBoost Regressor (XGBR). The models' performances were compared using evaluation metrics and plots. The findings demonstrated the superior performance of XGBR over all others. The average error percentages for the DTR, RFR, ETR and XGBR models were 10.3%, 8.3%, 7.4%, and 5.6%, respectively. For validation, the predicted production was compared with the production statistics on UPAg. Comparison with official production statistics confirmed the robustness and reliability of the XGBR model. The work highlights the potential of integrating remote sensing data with machine learning techniques to address the challenges in agriculture.

KI fragen

Bookmark

View Full Paper