| Aircraft detection based on Synthetic Aperture Radar (SAR) image has important application value in military and civil fields.However,the performance of aircraft detection in SAR images is severely restricted by the aircraft’s size,position difference,speckle noise and complex background interference.Deep neural network has attracted great attention because of its high precision and automatic image processing ability.However,the "black box"characteristic of depth model makes users unable to understand the reasons for its decision-making,which greatly hinders its application in practical engineering.Aiming at these problems,this paper carries out in-depth research,and the main contents and innovations are as follows:1) An Efficient Bidirectional Path Aggregation Attention Network (EBPA2N) is proposed for aircraft detection.In EBPA2N,YOLOv5s is used as the base network and then the Involution Enhanced Path Aggregation module and Effective Residual Shuffle Attention module are constructed.The network can capture the relationship between aircraft backscattering features and better encode multi-scale geospatial information,which greatly improve the accuracy of aircraft detection.2) A three-stage Geospatial Transformer detection framework is proposed,that is,the large scale SAR image is decomposed by sliding window,the aircraft is detected by multi-scale Geospatial Context Attention network (MGCAN),and the final detection result is obtained by recomposition.In MGCAN,the attention mechanism is adjusted by using geospatial analysis method,and then two innovative geospatial attention modules are proposed,which can fully capture multi-scale context information and geospatial information of aircraft,and better solve the interference of target location difference and complex background information.The experiments of multiple SAR images show the high reliability of this method.3) An explainable aircraft detection framework for SAR image is proposed,which includes three parts:backbone network selection,path aggregation network (PANet) and visualization of the detector.The Hybrid Global Attribution Mapping module provides interpretation information and selects the optimal backbone network for aircraft detection in SAR images.PANet provides advanced fusion technology to learn multi-scale features.Class-specific Confidence Scores Mapping is used to visualize the detection head and draw the detection performance to better understand the network.4) A innovative quantitative attribution evaluation network based on geographic statistics is proposed.Based on the local and global geospatial statistical methods,this network uses sliding window matrix blocks to highlight the significant regions of attribution,thus providing quantitative analysis for attribution of different eXplainable Artificial Intelligence (XAI)algorithms.Experiments on multiple SAR images with different XAI algorithms show the effectiveness of the proposed framework.Combined with SAR domain knowledge and deep learning,this paper constructs an aircraft detection network.The experimental performance is better than the existing detection model on 1m resolution Gaofen-3 data,and realizing rapid and high-precision aircraft automatic detection.In addition,the paper interprets the "black box" characteristics of depth network in SAR aircraft detection,which provides an important reference for scholars in SAR field to carry out research based on Geospatial Artificial Intelligence (GeoAI). |
