Semantic-Explicit Filtering Network for Remote Sensing Image Change Detection

Remote sensing image change detection (RSI-CD) aims to explore surface change information from aligned dual-phase images. However, RSI-CD currently encounters two major challenges. The first issue is the inadequate object-level semantic representation during the feature extraction in CD networks. The other issue is the spectral resolution of the RS image is limited, which leads to a mixture of pseudochange and real change. In order to explore the above-mentioned two challenges, we propose a semantic-explicit filtering network (SFNet) based on a neighborhood feature attention module (NFAM) and multiple-receptive-field semantic filtering mechanism (MSFM). First, the NFAM exploits the correlation of multiscale features and fuses features from the proximity layer to enhance the semantic-explicit representation of the object level. Then, the MSFM takes the weight map after the enhanced semantic representation as input and progressively refines the weight map through a multiple-receptive-field parallel convolution (MPC). This process filters out pseudochange from the predicted result while retaining the real-change information. The experiments on two benchmark datasets demonstrate that the proposed approach presents satisfactory performance over the existing methods.

I. Introduction

Change detection (CD) is one of the key research topics in remote sensing (RS). The goal of CD is to assign binary labels (i.e., with or without change) to each pixel within the same region by analyzing objects present in that region from images captured at disparate time intervals [1]. In recent years, the rise in demand for high-resolution RS image CD (RSI-CD) has made its applications more important. Therefore, it is greatly significant in many fields, such as urban planning [2], resource management [3], damage assessment [4], [5], and so on.