ResNet-GLUSE cuts parameters 33x for satellite image classification

Researchers from the University of Luxembourg and partners have developed ResNet-GLUSE, a lightweight convolutional neural network optimized for onboard satellite Earth observation image classification. The model integrates an adaptive channel-wise attention mechanism called GLUSE (Gated Linear Unit-enhanced Squeeze-and-Excitation) into a ResNet backbone. This dynamic gating improves feature recalibration while keeping computational overhead minimal. Tested on the EuroSAT and PatternNet datasets, ResNet-GLUSE achieves 94% and 98% accuracy respectively, proving its effectiveness for land-use and object classification from satellite imagery.

While the more complex MobileViT transformer model reaches 99% accuracy, ResNet-GLUSE offers drastic efficiency gains: 33x fewer parameters, 27x fewer FLOPs, 33x smaller model size, approximately 6x lower power consumption, and about 3x faster inference time. This makes it feasible for deployment on resource-constrained satellite hardware. Additionally, its simple architecture can be easily ported to neuromorphic chips like the Akida Brainchip, where it consumes only 852.30 mW. The balance of high accuracy and ultra-low resource usage positions ResNet-GLUSE as a practical solution for real-time, onboard Earth observation tasks.