SatNet-B3: A Lightweight Deep Edge Intelligence Framework for Satellite Imagery Classification

Abstract: Accurate weather prediction plays a vital role in disaster management and minimizing economic losses. This study presents SatNet-B3, a novel lightweight deep learning model for classifying satellite-based weather events with high precision. Built on the EfficientNetB3 backbone with custom classification layers, SatNet-B3 achieved 98.21% accuracy on the LSCIDMR dataset, surpassing existing benchmarks. Ten CNN models, including SatNet-B3, were experimented with to predict eight weather conditions: Tropical Cyclone, Extra-tropical Cyclone, Snow, Low Water Cloud, High Ice Cloud, Vegetation, Desert, and Ocean, with SatNet-B3 yielding the best results. Hyperparameter analysis was conducted to ensure optimal performance. The model addresses challenges like class imbal- ance and inter-class similarity through extensive data preprocessing and augmentation. To support scalability in big data contexts, the pipeline is designed to handle high-resolution geospatial imagery efficiently. Post-training quantization reduced the model size by 90.98% while retaining accuracy. Deployment on a Raspberry Pi 4 achieved a 0.3s inference time, enabling real-time decision-making in edge environments. By integrating explainable AI tools such as LIME and CAM, the model highlights influential image regions, improving interpretability and cognitive situational awareness for intelligent climate monitoring.