A hybrid attention-based ST-CNN and Penman-Monteith framework for high-resolution soil moisture recharge prediction using satellite driven data

Abstract

In monsoon-affected tropical deltas such as Bangladesh, where precipitation, evapotranspiration, and anthropogenic use of water combine complexly, accurately predicting soil moisture recharge (SMR) is essential for efficient agricultural water management. Most of the time, conventional procedure-based or machine learning approaches fail to do an adequate job of illustrating how SMR changes over time and space at high resolution. To deal with this problem, we present an innovative hybrid approach that brings together the Penman-Monteith evapotranspiration model with a spatiotemporal convolutional neural network that is based on attention (ST-CNN-Attention). Bangladesh's all 64 districts have been included in the model, which utilizes the complete NASA POWER dataset encompassing the years 2004-2024. It includes 22 daily climatic and geographic parameters. To maintain the reliability of the hydrological system, it employs CNN-based pattern recognition from satellite data, LSTM networks with attention mechanisms that focus on vital monsoon dynamics, and physically controlled evapotranspiration. This is the very first high-resolution SMR forecast system built specifically for Bangladesh's predominant polder hydrology. It is more accurate yet simpler to utilize. This structure is versatile so it can be utilized to arrange proper irrigation, alert of droughts ahead of time, and take ways to deal with climate change. Also, it could be beneficial in other delta regions that have been affected by rainfall and have water-management problems that are equivalent to those in the above locations.