Exploring the Application of Machine Learning for Soil Moisture Forecasting over In-situ Soil Moisture Sensors Network

This research explores the application of machine learning (ML) for soil moisture data prediction within a vast soil moisture sensor network (WITSMS-Network) deployed by WIT across the Indus Basin. While the network provides real-time soil moisture data, sensor malfunctions can lead to data gaps and sub-optimal irrigation practices. To address this challenge, we propose an approach utilizing Long Short-Term Memory (LSTM) models to ingest and analyze continuous time series soil moisture data from the network alongside multi-modal data from land surface models. This end-to-end sequential ML pipeline allows for multi-step forecasting with variable history across diverse sites. The research serves two primary purposes: 1) to assess the viability of ML methods for soil moisture forecasting within WIT’s extensive IoT network, and 2) to evaluate the quality of short-term (few months), fragmented (with missing values) multi-site forecasts against benchmark dataset from COSMOS-UK consisting of continuous, long-term data from a single location. Our findings suggest the ML model has the capacity to learn the variations in soil moisture content, giving dynamic forecasting results of 1.47 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−3</sup> ubRMSD for COSMOS-UK dataset and a high correlation of 98% for the WITSMS-Network dataset with static forecasting. Furthermore, the multi-site temporally sparse data offers limited forecasting accuracy compared to continuous single-site data.