MtSCCD：Land-use scene classification and change-detection dataset for deep learning

Abstract

Land-Use Scene Classification and change Detection (LUSCD) aim to recognize land-use types and monitor their changes by using Remote-Sensing (RS) images, which play an important role in urban planning and land-use optimization. In the era of RS big data, conventional hand-crafted feature-based methods are infeasible for LUSCD because the extracted features are not sufficiently discriminative for RS images with high complexity. As a novel data-driven paradigm for information extraction from RS images, deep learning provides a new solution for LUSCD. However, the existing publicly available datasets have limited samples and is thus unable to train a successful deep-learning model. Therefore, it has great significance in constructing an open and large-scale LUSCD benchmark. To advance the progress of LUSCD using deep-learning methods, this paper releases a large-scale scene classification and change-detection dataset termed Multi-temporal Scene Classification and Change Detection (MtSCCD). The RGB images in MtSCCD are cropped from large-size high-resolution RS images captured from the central areas of five China cities, namely, Hangzhou, Shanghai, Wuhan, Nanjing, and Hefei. The size of the cropped images is 300×300 pixels with the spatial resolution of around 1 m. MtSCCD has 10 land use classes, which are residential land, public service and commercial land, educational land, industrial land, transportation land, agricultural land, water body, green space, woodland, and woodland. Based on the cropped land-use images in MtSCCD, this paper constructs two sub-datasets termed MtSCCD_LUSC (MtSCCD Land Use Scene Classification) and MtSCCD_LUCD (MtSCCD Land Use Change Detection) for land-use scene classification (LUSC) and land-use change detection (LUCD), respectively. MtSCCD dataset has the following characteristics. (1) It is currently the largest publicly available LUSCD dataset, and both of the two sub-datasets (i.e., MtSCCD_LUSC and MtSCCD_LUCD) have 65548 images in total. (2) The images in MtSCCD are split into training set, validation set, and testing set according to the five cities. For example, images from three of the five cities are randomly split into training and validation set, whereas the rest remain to be the testing set. Therefore, MtSCCD has high extensibility, i.e., it can be easily extended to be a larger dataset. (3) For a deep-learning model, the training set and testing set are categorized from different cities, so it is beneficial to demonstrate the model’s generalization ability. (4) MtSCCD has high intra-class diversity, making it a challenging dataset. Based on MtSCCD_LUSC and MtSCCD_LUCD, this paper evaluates several deep-learning feature-based methods for LUSC and LUCD. Specifically, AlexNet, VGG networks (i.e., VGG16 and VGG19), GoogLeNet, and ResNet networks (i.e., ResNet18, ResNet50, and ResNet101) are selected to extract deep-learning features that are then fed into SVM for LUSC. We also evaluate DenseNet, EfficientNet, SENet, ViT, and SwinT for LUSC. Two kinds of LUCD approaches including conventional classification-based methods and current similarity-based methods have been evaluated. Experimental results show that the highest overall accuracy of MtSCCD_LUSC dataset is around 76%, indicating much room for improvement. Regarding LUCD, similarity-based methods particularly similarity learning-based ones outperform classification-based methods by a significant margin, providing a promising research direction for LUCD. This paper presents the currently largest scene classification and change-detection dataset MtSCCD based on high-resolution RS images of the central area of five China cities. MtSCCD contains two subsets MtSCCD_LUSC and MtSCCD_LUCD. Both had 10 land-use types and 65548 images in total. Based on the two sub-datasets, this paper evaluates the performance of several deep networks for scene classification and change detection, expecting to provide baseline results for related researchers. We hope that the MtSCCD dataset can promote this progress in land-use type recognition and monitoring.