Evaluation of Quantized CNN Architectures for Land Use Classification for Onboard Cube Satellite Computing

Nanosatellites have increased in popularity, but its limited communication bandwidth and computing resource constraints remain a challenge. Orbital edge computing has been explored to address this issue by utilizing neural networks onboard the satellites for land use classification. This study explores the performance of pre-trained quantized neural networks in land use classification by fine-tuning ResN et50, MobileNetV3, and GoogLeNet using the EuroSAT RGB dataset with an 80–20 split. The study evaluated the three models based on its accuracy and file size and against its non-quantized counterparts. It was found that the best model is the MobileNetV3, with an 89.04% accuracy at a 3,090 KB file size. MobileNetV3 was also found to be the best architecture for quantization with the least impact to accuracy, with a reduction of only 5.05%, and having the most significant reduction in file size of 85.71 %.