SUMLP: A Siamese U-shaped MLP-based network for change detection

Abstract

Remote sensing image change detection (CD) technology is an important means to understand the changes of the earth surface. In recent years, many excellent CD methods based on convolutional neural networks (CNN) have been proposed. Although these methods have achieved good results, CD is still a very challenging task. Recently, many visual algorithms based on multi-layer perceptron (MLP) have been proposed and achieved similar or even better results than CNN-based algorithms in many visual fields. Inspired by the success of these MLP-based algorithms, in this paper we propose a siamese U-shaped MLP-based CD network named SUMLP. First, SUMLP uses CycleMLP block as the basic unit of the network and combines the characteristics of the CD task to construct a MLP-based CD network with siamese U-shaped structure. Secondly, to better optimize the CD network, we design a new loss function, weighted cosine cross entropy (WCCE) loss. WCCE loss has two advantages: 1. WCCE loss gives different weights to the changed samples and the unchanged samples, which can cope with the data imbalance problem in the CD task; 2. WCCE loss uses cosine margin to increase the inter class distance and the intra class compactness, so as to better distinguish the changed samples from the unchanged samples. For comprehensively verifying the validity of our proposed method, we carried out many experiments on three public CD datasets: LEVIR-CD dataset, CDD dataset and WHU-CD dataset. On these datasets, SUMLP’s F1−score are at least 0.33%, 1.41% and 3.56% higher than the other relevant CD methods respectively.