Efficient AI in remote sensing application

Research question: Most remote sensing devices are deployed on an aircraft or satellite, which only has low level of CPUs and GPUs embedded. Can we reduce cost in edge devices without significant loss of performance in remote sensing application?

In this study, we present an in-depth evaluation of neural network compression techniques used particularly for a remote sensing application, i.e., land cover classification. We examine several model compression techniques applied to convolutional neural networks, including pruning, quantization, and knowledge distillation. The evaluation is performed on two benchmark datasets, taking into account both the accuracy and memory consumption of the compressed models. Furthermore, we investigate the trade-offs between compression ratios, inference time, and classification performance to gain insight into the practical application of various compression algorithms in remote sensing contexts. Our experimental results show that neural network compression approaches reduce model size and computational cost while still performing competitively good in land cover classification tasks.