Cluster-based local modeling (CBLM) paradigm meets deep learning: A novel approach to soil moisture estimation

Abstract

Producing precise soil moisture maps through soil moisture modeling is highly valued for a variety of purposes, such as agricultural productivity, water resource management, climate modeling, environmental monitoring, and disaster management. However, the severe spatial variability of soil moisture makes a single model insufficient for accurately estimating soil moisture in an area. In this study, a deep learning-based soil moisture estimation approach using the Cluster-Based Local Modeling (CBLM) paradigm is presented. The approach involves the utilization of Sentinel-2 imagery and ancillary data such as soil maps, geology maps, land use/cover maps as well as topographic and spectral indices to create a database for modeling. The optical trapezoid model (OPTRAM) was also employed to estimate relative soil moisture content as another ancillary data for soil moisture modeling. Some robust clustering algorithms i.e. K-means, genetic algorithm, particle swarm optimization, and density-based spatial clustering were applied to divide the study area into homogenous sub-areas. Two powerful deep learning methods, i.e., Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), were then trained on the dataset to make accurate soil moisture estimations. Promising results were obtained through the proposed approach in accurately modeling soil moisture. According to the findings, the global modeling approach using both LSTM and CNN models exhibited moderate performance. Isolated pixels with unreasonably high soil moisture values were produced by the LSTM model which may be related to modeling process errors. Significantly enhanced results were observed in both deep learning models with the use of the local modeling approach. The cluster-based local modeling paradigm was able to capture local variations in soil moisture and produce more interpretable results by developing separate models for each cluster. Moreover, the CBLM-based CNN-CLTM hybrid model outperformed both local CNN and LSTM models. Generalization of the results was also confirmed by applying the best model for a different year and different season. Overall, the approach shows promise in improving our understanding of soil moisture dynamics and identifying areas that are vulnerable to drought or flooding, which could help inform targeted interventions.