Deep Learning Based Soil Moisture Downscaling Framework for Precision Agriculture in Data-Scarce Regions

High-resolution soil moisture (SM) information is critical for irrigation decision-making, crop modelling, flood and drought prediction, and water resources management. However, satellite products only provide coarse-resolution data that cannot capture farm-scale spatial variability influenced by factors such as soil heterogeneity, topography, and anthropogenic activities. While downscaling methods offer a potential solution, they currently struggle in data-scarce regions, such as India, where the absence of dense observation networks limits their effectiveness. In this study, we present an irrigation optimisation framework that downscales satellite-derived soil moisture (SM) data to field-scale root zone soil moisture (RZSM) to support data-driven irrigation decision-making in Nashik District, Maharashtra, India. Utilising a Convolutional Long Short-Term Memory (ConvLSTM) network, we integrated sparsely located in-situ data from ground-based sensors with remote sensing predictors, including precipitation, vegetation indices, land surface temperature, and terrain attributes. The ConvLSTM architecture captures non-linear spatial and temporal interactions governing the field-scale SM variability. The models achieved strong performance, with Root Mean Square Error (RMSE) values from 0.02 to 0.08 m³/m³, Mean Absolute Error (MAE) values from 0.02 to 0.06 m³/m³, Correlation Coefficient (r) values ranging from 0.79 to 0.92, and Coefficient of Determination (R²) values between 0.61 and 0.88. These results validate the potential of deep learning for accurate field-scale SM estimation without requiring dense ground networks. Building on this, we are currently extending the framework by coupling the ConvLSTM architecture with a farm-scale ecohydrological model. This hybrid approach enables generalised, field-scale mapping at ungauged locations without in-situ sensors, offering a scalable, scientifically grounded solution for precision agriculture in water-stressed regions. This work can support farmers in making informed irrigation decisions and contribute to improved water management practices.