Classification And Target Detection In Remote Sensing Images

With the development of remote sensing,a large number of remote sensing images are produced every day.Because of its high resolution and wealth of spectrum information,remote sensing images can be widely used in urban planning,environmental governance,agricultural production and other aspects.However,a large amount of data cannot be fully utilized by humans alone,so machine learning is increasingly used in the field of remote sensing.The purpose of this article is to trace the source of pollution using remote sensing images.The research includes land use segmentation and target detection.The main contributions of this article include:1.In the land use segmentation task,this thesis designs and implements a Landsat data preprocessing system based on Google Earth Engine.Due to the characteristics of the data,this thesis proposes a new preprocessing method,which can train neural networks with ideal performance on weakly labeled datasets,solving the problem of satellite data labeling difficulties.The experiments on real-time images prove the effectiveness of this method.2.In the object detection task,this thesis proposes a few-shot detection framework.In the feature extraction stage,this thesis proposes a new training strategy for training the attention module in the network,so that the network can distinguish the importance of different features in new class samples.In the detection stage,this thesis proposes a detection method based on metric learning,and optimizes the detection results through a new postprocessing method.The results on the both NWPU-RESISC45 dataset and the Gaofen2 satellite images prove the effectiveness of our proposed method.