Open access to Earth observation data, such as Copernicus and Landsat program data, along with remarkable advancements in deep learning, has played a significant role in computer vision tasks. Consequently, there is an increasing demand for well-labeled datasets and the development of optimized models that can be executed on resource-limited hardware. In this study, the Hybrid Multi-Block Convolutional Neural Network (HCNN) was employed for land use and land cover (LULC) classification at the patch level. HCNN was designed to achieve accuracy comparable to or higher than state-of-the-art models while reducing complexity to ensure faster training and optimized performance. To evaluate the model, paired datasets from Sentinel-1 and Sentinel-2 imagery were constructed using Google Earth Engine, covering study areas in Iran. The datasets were co-registered to ensure spatial alignment between the SAR and optical data. The performance of HCNN was benchmarked against three widely used CNN architectures, including DenseNet-121, ResNet-50, and VGG-19. HCNN demonstrated high classification accuracy on both optical and SAR datasets, achieving 99.56% and 99.14%, respectively. These results consistently surpassed the performance of the other CNN architectures, highlighting HCNN's superior accuracy and efficiency across both data types. Moreover, HCNN required substantially less training time, completing in approximately one-eighth of the time needed by the benchmark models, highlighting its efficiency in both accuracy and computational cost.