Prob-POS: A Framework for Improving Visual Explanations from Convolutional Neural Networks for Remote Sensing Image Classification

• DOI: 10.3390/rs14133042
• Publication Date: 2022-06-27T02:50:23Z
• Authors: Xianpeng Guo; Biao Hou; Zitong Wu; Bo Ren; Shuang Wang; Licheng Jiao
• Keywords: Feature (linguistics); Representation; Contextual image classification

During the past decades, convolutional neural network (CNN)-based models have achieved notable success in remote sensing image classification due to their powerful feature representation ability. However, lack of explainability during decision-making process is a common criticism these high-capacity networks. Local explanation methods that provide visual saliency maps attracted increasing attention as means surmount barrier explainability. vast majority research conducted on last layer, where salient regions are unintelligible for partial images, especially scenes contain plentiful small targets or similar texture image. To address issues, we propose novel framework called Prob-POS, which consists class-activation map based probe (Prob-CAM) and weighted probability occlusion (wPO) selection strategy. The proposed simple but effective architecture generate elaborate can be applied any layer CNNs. wPO quantified metric evaluate effectiveness each different categories automatically pick out optimal layer. Variational weights taken into account highlight high-scoring map. Experimental results two publicly available datasets three prevalent networks demonstrate Prob-POS improves faithfulness CNNs images.