Authors: Sushant Mehan*, Manoj Lamichhane (* Presenting Author)
Title: Surface soil moisture simulations for crop lands using Remote Sensing and Machine Learning Approaches
Abstract: Soil moisture is pivotal for hydrological modeling and managing agricultural water, especially crucial in addressing the challenge of securing optimal crop yields to meet the food demands of a growing population in the face of widespread water scarcity issues. This challenge is particularly pronounced in arid and semi-arid regions. Despite the availability of various soil moisture monitoring products, their utility at the field scale is limited due to their relatively coarse spatial resolution. Our study addresses this gap by aiming to predict soil moisture at a much finer spatial resolution (10 meters) using the Synthetic Aperture Radar (SAR) Sentinel-1 C-band and Harmonized Landsat Sentinel (HLS) data, with the help of in-situ soil moisture measurements from a field located in the semi-arid locale of Akron, CO, USA. To accomplish our goal, we employed four machine learning models: support vector machines (SVM), random forests (RF), gradient boosting machines (GBM), and 1D Convolutional Neural Network (CNN). These models were evaluated based on their performance metrics, including root mean square error (RMSE), mean square error (MSE), and coefficient of determination (R2), to identify the most proficient model in predicting soil moisture using backscattering coefficients derived from the SAR Sentinel-1 C-band data and spectral indices from HLS. Our findings indicated that the 1D CNN model excelled, demonstrating superior accuracy with an R2 of 0.79 and an RMSE of 0.26 m3/m3. This suggests that the 1D CNN model holds significant promise for field-scale soil moisture prediction in semi-arid regions, offering a valuable tool for enhancing agricultural water management and hydrological modeling in the face of global water scarcity challenges.