Senapati et al. (2025) High-resolution agricultural drought hazard mapping using the potential of geospatial data and machine learning approaches

Identification

• Journal: Environmental Monitoring and Assessment
• Year: 2025
• Date: 2025-10-10
• Authors: Ujjal Senapati, Aman Srivastava, Rajib Maity
• DOI: 10.1007/s10661-025-14538-w

Research Groups

• Department of Civil Engineering, Indian Institute of Technology Kharagpur, Kharagpur, West Bengal, India

Short Summary

This study developed a machine learning-geospatial framework to map high-resolution agricultural drought hazard (ADH) zones in semi-arid, rainfed basins, demonstrating that the Random Forest model achieved superior performance (AUC-ROC of 97.8%) and identified 31.77% of the Upper Dwarakeshwar River Basin as very high hazard zones.

Objective

• To develop a robust, transparent, and transferable machine learning-geospatial framework for high-resolution agricultural drought hazard (ADH) mapping.
• To assess the relative importance of eight geo-environmental predictors for ADH modeling using correlation attribute evaluation.
• To systematically compare the predictive accuracy of Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Adaptive Regression (AR) algorithms for generating ADH maps.

Study Configuration

• Spatial Scale: Upper Dwarakeshwar River Basin (UDRB), West Bengal, India, covering 1934 square kilometers, with ADH maps generated at 30 meters resolution.
• Temporal Scale: Data primarily from 2015–2020, with specific datasets like groundwater from 2014–2017 and paddy analysis for 2018 (normal) and 2019 (slight drought).

Methodology and Data

• Models used: Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), Adaptive Regression (AR).
• Data sources:
  • Satellite-derived: Soil Moisture Index (SMI) from SMAP L-band radiometer data, Temperature Condition Index (TCI) from MODIS Land Surface Temperature (LST), Vegetation Condition Index (VCI), Normalized Difference Vegetation Index (NDVI), and Land Use Land Cover (LULC) from Landsat-8, paddy field distribution from Sentinel-2 MultiSpectral Instrument (MSI) imagery.
  • Observation/Reanalysis: Meteorological Drought Intensity (MDI) from India Meteorological Department (IMD) gridded rainfall data, Soil Drainage (SDR), Soil Depth (SDE), and Soil Texture (ST) from National Bureau of Soil Survey and Land Use Planning (NBSS&LUP) soil survey sheets, Mean Groundwater Level (MGL) from Central Ground Water Board (CGWB) monitoring wells.
  • Input variables: SMI, TCI, SDR, MDI, SDE, MGL, LULC, ST.
  • Target variable: Agricultural drought areas derived from VCI (pixels with VCI < 0.5).

Main Results

• Multicollinearity analysis confirmed the absence of significant collinearity among the eight geo-environmental predictors (tolerance values > 0.1, VIF values < 5).
• Correlation Attribute Evaluation (CAE) identified Soil Moisture Index (SMI) as the most influential factor (Average Merit = 0.357), followed by Temperature Condition Index (TCI, 0.322) and Soil Depth (SDE, 0.305).
• During the testing phase, the Random Forest (RF) model exhibited superior performance with an Area Under the Curve-Receiver Operating Characteristic (AUC-ROC) of 97.8%, a correlation coefficient (R) of 0.869, and a Root Mean Square Error (RMSE) of 0.26.
• The RF model classified 31.77% of the Upper Dwarakeshwar River Basin (UDRB) as "very high" agricultural drought hazard zones and 19.95% as "high" hazard zones.
• A comparative analysis showed a 4.16% decline in total cultivated rice paddy area in a slight drought year (2019) compared to a normal year (2018), with the most significant reductions in very high (7.94%) and high (6.11%) drought-prone zones.

Contributions

• Introduces a novel machine learning-geospatial framework that integrates eight diverse geo-environmental variables (hydrological, pedological, thermal, and land-use factors) to map agricultural drought hazard at a high resolution of 30 meters.
• Systematically compares four established machine learning algorithms, demonstrating the superior predictive accuracy and robustness of the Random Forest model for agricultural drought hazard mapping in semi-arid, rainfed basins.
• Provides high-resolution, implementable agricultural drought hazard maps and an operational workflow designed to guide policy and adaptive resource management for various stakeholders, including farmers, local governments, and NGOs.
• Addresses critical limitations of conventional single-index or static multi-criteria drought assessment methods by effectively capturing complex, non-linear interactions between geo-environmental drivers.
• Establishes a framework with demonstrated transferability to similar agroecological regions characterized by rainfed agriculture and recurrent drought events.

Funding

• Ministry of Earth Sciences (MoES), Government of India (Project No. MoES/PAMC/H&C/124/2019-PC-II).

Citation

@article{Senapati2025Highresolution,
  author = {Senapati, Ujjal and Srivastava, Aman and Maity, Rajib},
  title = {High-resolution agricultural drought hazard mapping using the potential of geospatial data and machine learning approaches},
  journal = {Environmental Monitoring and Assessment},
  year = {2025},
  doi = {10.1007/s10661-025-14538-w},
  url = {https://doi.org/10.1007/s10661-025-14538-w}
}