Water level forecasting plays a crucial role in water resource management and disaster mitigation; however, traditional methods face significant limitations in regions with complex terrain and highly dynamic hydrological conditions. This study proposes a novel multimodal water level forecasting framework that integrates remote sensing data with advanced deep learning models. A key contribution is the use of YOLOv9 (You Only Look Once, version 9) to automatically detect reservoir surface areas from Sentinel-2 satellite imagery, ensuring precise reservoir boundary extraction without manual intervention. Image features are subsequently extracted using VGG19 (Visual Geometry Group 19), while hydrological time-series features are processed through a fully connected neural network. These heterogeneous feature sets are projected into a unified latent space and integrated into a SARIMAX (Seasonal AutoRegressive Integrated Moving Average with eXogenous variables) model to improve forecasting accuracy. The system is deployed on a web-based platform, providing intuitive interaction and ease of use. Experimental results at An Khe Reservoir demonstrate superior performance, with improvements of 14.5% (Root Mean Square Error-RMSE) and 11.7% (Mean Absolute Error-MAE), confirming the feasibility and practical value of the proposed approach.