ChangeNet: Multi-Temporal Asymmetric Change Detection Dataset

TL;DR

This work addresses the limitations of existing change detection datasets—namely small size, short temporal coverage, and symmetrical ground truth—by introducing ChangeNet, a large-scale, multi-temporal remote sensing dataset with up to six temporal phases, 31,000 image pairs, and dense pixel-level annotations across six categories. It also defines the new task of Asymmetric Change Detection and establishes BACD and SACD benchmarks on ChangeNet, demonstrating that the dataset presents meaningful challenges beyond prior datasets and that pretraining on ChangeNet can improve performance on other CD tasks. The results underscore the dataset's practicality, realism (including perspective distortions), and potential to advance robust, transferable change detection methods for real-world applications. Overall, ChangeNet provides a rich resource for evaluating and developing both binary and semantic change detection with temporal asymmetry in remote sensing imagery.

Abstract

Change Detection (CD) has been attracting extensive interests with the availability of bi-temporal datasets. However, due to the huge cost of multi-temporal images acquisition and labeling, existing change detection datasets are small in quantity, short in temporal, and low in practicability. Therefore, a large-scale practical-oriented dataset covering wide temporal phases is urgently needed to facilitate the community. To this end, the ChangeNet dataset is presented especially for multi-temporal change detection, along with the new task of "Asymmetric Change Detection". Specifically, ChangeNet consists of 31,000 multi-temporal images pairs, a wide range of complex scenes from 100 cities, and 6 pixel-level annotated categories, which is far superior to all the existing change detection datasets including LEVIR-CD, WHU Building CD, etc.. In addition, ChangeNet contains amounts of real-world perspective distortions in different temporal phases on the same areas, which is able to promote the practical application of change detection algorithms. The ChangeNet dataset is suitable for both binary change detection (BCD) and semantic change detection (SCD) tasks. Accordingly, we benchmark the ChangeNet dataset on six BCD methods and two SCD methods, and extensive experiments demonstrate its challenges and great significance. The dataset is available at https://github.com/jankyee/ChangeNet.

Figures (2)

• Figure 1: The existing bi-temporal change detection datasets, including LEVIR-CD LEVIR-CD, DSIFN-CD DSIFN-CD, WHU Building CD WHU, SECOND asymmetric. The former three are only for binary change detection (BCD), while the latter is applicable to semantic change detection. Existing change detection datasets usually consist of short (only two) temporal phases with completely symmetrical land-cover appearances.
• Figure 2: The ChangeNet contains images covering a wide range of time phases and the land-cover change annotations. There are up to 6 temporal phases with amounts of practical asymmetric changes. Here, we select four samples with 3 temporal phases.