A Cloud-based Machine Learning Approach for Surface Soil Moisture Mapping Using Multi-sensor Remote Sensing Data for the Sher River Watershed
Anoop Patel
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
S. K. Pyasi
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
A. K. Bajpai
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
Anay Rawat
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
Y. K. Tiwari
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
Archana Kaushal
Department of Soil and Water Engineering, College of Agricultural Engineering, JNKVV, Jabalpur - 482004, Madhya Pradesh, India.
Akshita Tomar
SRLM Division, SR & LUM Group, RSA-A, National Remote Sensing Centre (NRSC), Hyderabad, 500042, India.
V. S. Yadav *
Department of Agricultural Engineering, Institute of Agricultural Sciences, Banaras Hindu University, Varanasi - 221005, Uttar Pradesh, India.
*Author to whom correspondence should be addressed.
Abstract
Soil moisture is often referred to as the "blood of the soil" because, just as blood sustains life, soil moisture is essential for plants to survive and thrive. It plays an importance role in maintaining ecological balance, supporting agricultural productivity, and regulating the hydrological cycle. This study developed a soil moisture model using Machine Learning (ML) techniques, focusing specifically on the Sher River Basin located in the Narsinghpur district of Madhya Pradesh, India. This region exhibits diverse hydro-climatic conditions and land-use patterns. 32 in-situ soil moisture observations were collected from 0-15 cm depth soil surface on April 21, 2025, to calibrate the model. The Sentinel-1 synthetic aperture radar parameters (10m), namely, backscatter coefficients (VV, VH, and VV/VH), Sentinel-2 (10m), (Normalized Difference Vegetation Index (NDVI), Optimized Soil-Adjusted Vegetation Index (OSAVI), Normalized Difference Water Index (NDWI)), land surface temperature from Landsat-9 imagery (30m), rainfall sums from CHIRPS datasets (5 km), and MODIS evapotranspiration and terrain features were used as predictor variables. Three modelling tools, namely Random Forest (RF), Classification and Regression Tree (CART) and Multiple Linear Regression (MLR), were evaluated for their effectiveness in modelling soil moisture. The dataset is split into two parts: training (70%) and testing (30%) for independent verification. Furthermore, the model exhibited robust predictive capabilities on the test dataset, achieving an R² of 0.73 and a minimum equation error of 0.07. The CART model demonstrated marginally reduced accuracy compared to the top random forest model, while the linear regression method was less adept at identifying intricate relationships among predictors. This Calibrated model give Soil moisture map at 10 m spatial resolution. The model is really useful for planning of watershed area, Drought assessment, crop heath monitoring and Precision agriculture.
Keywords: Soil moisture, random forest, machine learning, NDVI, NDWI, multi-sensor image, classification and regression tree