Soil moisture Modeling and Forecasting using Spatiotemporal Machine Learning Based Models

Abstract:

Soil moisture plays a significant role in determining the probability of flooding at a given location. Currently, soil moisture is commonly modeled using physically based numerical models. These numerical models can consume high computational time and resources with increasing spatial and temporal model resolutions. In this study, we propose a data-driven modeling approach using Machine Learning (ML) models. ML models such as Convolutional Neural Network (CNN) are well-suited to capture and learn spatial patterns in datasets, while other model types such as Long Short-Term Memory (LSTM) are designed to utilize time-series information and learning from past observations. Recently, machine learning models that combine the capabilities of CNN and LSTM were developed. In this study, we investigate the applicability of one of these models in predicting soil moisture over our study area located in south Louisiana. This study reveals that the combined model significantly outperformed CNN alone in predicting soil moisture. Moreover, we tested the model’s performance using a combination of different sets of predictors and different temporal sequence lengths. Our results show that the combined model can predict soil moisture with mean errors below or equal to 2.5%. The model can also predict inter-observation values making them very useful for applications such as filling the gaps between overpasses of soil moisture observing satellites. Another interesting aspect about our ML models is that they were developed using free web-based resources, and can be easily shared and edited in an online collaborative environment.