Suggestive Data Annotation for CNN-Based Building Footprint Mapping Based on Deep Active Learning and Landscape Metrics

Abstract

Convolutional neural network (CNN)-based very high-resolution (VHR) image segmentation has become a common way of extracting building footprints. Despite publicly available building datasets and pre-trained CNN models, it is still necessary to prepare sufficient labeled image tiles to train CNN models from scratch or update the parameters of pre-trained CNN models to extract buildings accurately in real-world applications, especially the large-scale building extraction, due to differences in landscapes and data sources. Deep active learning is an effective technique for resolving this issue. This study proposes a framework integrating two state-of-the-art (SOTA) models, U-Net and DeeplabV3+, three commonly used active learning strategies, (i.e., margin sampling, entropy, and vote entropy), and landscape characterization to illustrate the performance of active learning in reducing the effort of data annotation, and then understand what kind of image tiles are more advantageous for CNN-based building extraction. The framework enables iteratively selecting the most informative image tiles from the unlabeled dataset for data annotation, training the CNN models, and analyzing the changes in model performance. It also helps us to understand the landscape features of iteratively selected image tiles via active learning by considering building as the focal class and computing the percent, the number of patches, edge density, and landscape shape index of buildings based on labeled tiles in each selection. The proposed method was evaluated on two benchmark building datasets, WHU satellite dataset II and WHU aerial dataset. Models in each iteration were trained from scratch on all labeled tiles. Experimental results based on the two datasets indicate that, for both U-Net and DeeplabV3+, the three active learning strategies can reduce the number of image tiles to be annotated and achieve good model performance with fewer labeled image tiles. Moreover, image tiles with more building patches, larger areas of buildings, longer edges of buildings, and more dispersed building distribution patterns were more effective for model training. The study not only provides a framework to reduce the data annotation efforts in CNN-based building extraction but also summarizes the preliminary suggestions for data annotation, which could facilitate and guide data annotators in real-world applications.