Satellite image classification is essential for applications in environmental monitoring, urban development, and agriculture, as it enables the identification of land-use and land-cover features from high-resolution, wide-area imagery. Nevertheless, the inherent variability and temporal dynamics of satellite data introduce significant challenges, since images acquired at different times or under varying conditions may display subtle structural changes due to noise or genuine evolution. This work introduces Kalman-SLDA, a hybrid framework combining convolutional neural network (CNN) feature extraction with enhanced Streaming Linear Discriminant Analysis (SLDA) and Kalman Filtering. The pretrained CNN initializes robust feature representations from optical satellite imagery, while Kalman-SLDA dynamically adapts these features in streaming settings, addressing three core challenges: (1) computational complexity of high-resolution data streams, (2) temporal consistency under class imbalance, and (3) concept drift mitigation compared to other classifiers. Experimental results on the Functional Map of the World - Time dataset show that Kalman-SLDA outperforms conventional classifiers in accuracy. These findings emphasize the necessity of efficient, adaptive algorithms for real-time satellite image analysis, ultimately supporting faster and more reliable data-driven decision-making in dynamic operational contexts.