Research On Key Technologies Of Remote Sensing Satellite Image Interpretation Based On Few-shot Learning

Thanks to remote sensing satellites’ global observation ability,stable and consistent sensor parameters and operation for many years,the information of the earth’s surface has been recorded in detail.The information consists of urban road information with the trajectory of the GPS vehicle navigation,such as multi-source remote sensing information.The information can be used in environmental monitoring,resource management,disaster forecast,major engineering supervision,unexpected events,such as national defense security.With proper interpretation of these information,the results can help the government to make the decision,to study the cause of events and state of law enforcement,to help local governments for local construction planning.With the rapid development of deep learning in the field of computer vision in recent years and the rapid upgrading of GPU,the image semantic segmentation technology in deep learning has gradually shifted from experimental research to practical application in the field of remote sensing images,so as to assist human in the interpretation of remote sensing satellite images and accelerate the efficiency of image processing.Aiming at the few shot learning task that often appears in remote sensing satellite image tasks,this paper proposes solutions from the following aspects:1.In terms of data generation for rare data,this paper proposes a"Label to Data" generation method to generalize the few-shot data according to the rare data,so as to help the model learn the key features of rare data and improve the model performance;2.In terms of model construction,this paper attempts to compare a variety of optimization modules by analyzing the characteristics of remote sensing satellite data and the performance of baseline model on the dataset of remote sensing satellite,and finally builds a lightweight attention module to enhance local features.3.By Organizing and training the model through the generalization of few-shot learning method,this paper improves the generalization ability of the model on rare data.Finally,through the result of simple integrated classification model,this paper realizes the interpretation of the damage level of building in disaster in the high-resolution remote sensing satellite data.