A Knowledge Distillation-Based Ground Feature Classification Network With Multiscale Feature Fusion in Remote-Sensing Images

As a fundamental task in remote-sensing interpretation, semantic segmentation of remote-sensing images intends to allocate a definite class to each pixel in the image. Fast and efficient semantic segmentation of high-resolution remote-sensing images provides help to capture the real surface covering and plays an essential role in urban planning and dynamic monitoring. However, there are still some limitations in the previous remote-sensing image semantic segmentation model for urban scenes, such as the low weight of small target pixels and the tiny target size leading to the unsatisfactory recognition and segmentation results of the model for small target features. Meanwhile, the deeper and broader feature extraction module in the semantic segmentation network usually leads to more redundant parameters, which takes a lot of computation time. Thus, we propose a lightweight semantic segmentation network based on the knowledge distillation combined with a multiscale pyramidal pooling module and attention mechanism named KD-MSANet, which enhanced the ability to fuse and focus on shallow features. Then, we trained teacher—student models to obtain lightweight network models through a model pruning and distillation framework. Experiments on Vaihingen and Potsdam datasets demonstrated that the network we designed significantly reduces the number of parameters while ensuring almost constant accuracy. Compared with the precompression model, the student model reduced in size by 43.6% and the training efficiency was improved by 22.3%, while the accuracy reached 99.30% of the teacher model.