We propose a CPU-friendly knowledge distillation framework that merges the compact efficiency of MobileNetV3 with a Squeeze-and-Excitation (SE) block to create a powerful yet ultra-lightweight model for remote sensing. By blending standard cross-entropy training with softened teacher outputs, our student network gains enhanced channel attention at minimal extra overhead—only around 2.3M parameters and a 9MB on-disk size. Despite its modest footprint, it achieves competitive accuracy on representative benchmarks, including 95.24 % on UC Merced Land Use and 94.78 % on EuroSAT. This fusion of SE-driven attention and knowledge distillation addresses resource-constrained deployment scenarios (such as real-time edge inference or on-board satellite processing) where memory and compute are at a premium. Our results highlight that a streamlined, distilled architecture can approach state-of-the-art performance in aerial scene classification without sacrificing the portability essential for practical, real-world applications.