Abstract
This research confronts the escalating challenges in precision irrigation driven by climate change by proposing an advanced machine learning framework. The system integrates high-resolution multispectral imagery from Sentinel-2 satellites, specifically leveraging vegetation indices such as NDVI, NDMI, and LAI, with comprehensive meteorological data from OpenMeteo to forecast soil moisture content. A multibranch deep learning architecture was developed, which evaluates LSTM recurrent neural networks, to effectively model the complex temporal dynamics inherent in agricultural environments. Analysis of a substantial dataset revealed that the LSTM-based model demonstrated superior capability in capturing long-term dependencies and showed a moderate predictive strength. However, the model also highlighted difficulties in accurately predicting extreme values. Feature importance analysis indicated that satellite-derived vegetation indices (e.g., fcover count, ndvi count) exerted a more significant influence on predictions than meteorological inputs. This study contributes to enhancing the precision of irrigation scheduling and optimizing resource use for sustainable agriculture. Future research directions include the integration of detailed soil characteristics, specific indicators for extreme weather events, and the exploration of hybrid model architectures to improve predictive accuracy, particularly for outlier conditions, and to ensure robustness across varied agro-ecological zones.