Novel Land Cover Change Detection Deep Learning Framework with Very Small Initial Samples Using Heterogeneous Remote Sensing Images

Abstract

Change detection with heterogeneous remote sensing images (Hete-CD) plays a significant role in practical applications, particularly in cases where homogenous remote sensing images are unavailable. However, directly comparing bitemporal heterogeneous remote sensing images (HRSIs) to measure the change magnitude is unfeasible. Numerous deep learning methods require substantial samples to train the module adequately. Moreover, the process of labeling a large number of samples for land cover change detection using HRSIs is time-consuming and labor-intensive. Consequently, deep learning networks face challenges in achieving satisfactory performance in Hete-CD due to the limited number of training samples. This study proposes a novel deep-learning framework for Hete-CD to achieve satisfactory performance even with a limited number of initial samples. We developed a multiscale network with a selected kernel-attention module. This design allows us to effectively capture different change targets characterized by diverse sizes and shapes. In addition, a simple yet effective non-parameter sample-enhanced algorithm that utilizes the Pearson correlation coefficient is proposed to explore the potential samples surrounding every initial sample. The proposed network and sample-enhanced algorithm are integrated into an iterative framework to improve change detection performance with a limited number of small samples. The experimental results were achieved based on four pairs of real HRSIs, which were acquired with Landsat-5, Radarsat-2, and Sentinel-2 satellites with optical and SAR sensors. Results indicated that the proposed framework could achieve competitive accuracy with a small number of samples compared with some state-of-the-art methods, including three traditional methods and nine state-of-the-art deep learning methods. For example, the improvement rates are approximately 3.38% and 1.99% compared with the selected traditional methods and deep learning methods, respectively.