A HYBRID CONVLSTM–HIERARCHICAL GNN FRAMEWORK FOR SOIL MOISTURE PREDICTION USING IOT AND SENTINEL-2

Abstract

Machine Learning (ML) technology in Indian agriculture targets the improvement of crop supplies, reducing resource usage, and enhancing farmer efficiency through various applications such as Soil Moisture Prediction and crop health monitoring. Soil moisture prediction helps to reduce water usage and supports personalized advisory systems for farmers. Soil moisture prediction has a major impact on agricultural water resource management. In Traditional ML approaches based up on Convolutional Long Short-Term Memory (ConvLSTM) and Graph Neural Networks (GNN) to implement soil moisture prediction using Sentinel-2 images and in-situ soil moisture data. In-situ soil moisture observations were collected from networks such as the International Soil Moisture Network (ISMN). These point-based measurements are insufficient for building, training, and validating soil moisture prediction models over state- and district-level scales.  Recent Advances in ML have enabled the integration of multisource data to enhance predictive accuracy.  In this paper, we propose a model that combines multi-source data, that is Internet of Things (IoT), Sentinel-2, and Andhra Pradesh Water Resources Management (APWRMS), to make predictions across district-to-state scales. In this research, ConvLSTM was used to extract spatiotemporal features from Sentinel-2 images, and a Hierarchical GNN was used to capture multi-scale spatial dependencies. The main advantages are expanded spatial coverage and improved accuracy through multi-source integration. The model used six Sentinel-2 spectral bands (B2, B4, B6, B8, B10, and B12) for district level soil moisture estimation in Andhra Pradesh. Model performance was evaluated using the coefficient of determination (R²) and the root mean square error (RMSE). Experimental results demonstrate that the proposed ConvLSTM–Hierarchical Graph Neural Network (CHGNN) model achieves R² = 0.88 and RMSE = 0.0403, significantly outperforming the traditional ConvLSTM–GNN approach (R² = 0.692, RMSE = 0.0645). This study highlights the potential of multi-source, deep, and multi-fusion learning frameworks for accurate and efficient soil moisture and soil health monitoring.