Extraction Of Built-up Areas From High-resolution Remote Sensing Images Based On Lightweight Neural Networks

Built-up areas are the primary places for human activities,and the acquisition,analysis,and application of geographic spatial information on built-up areas are critical for the sustainable development of cities.The rapid development of high-resolution remote sensing has made it possible to achieve fine-scale mapping of built-up areas.However,the complex texture and diverse object classes of high-resolution remote sensing images make it challenging to extract built-up areas.Additionally,for remote sensing images from different sensors,transferring the built-up area extraction model trained on sample data from one sensor to image data from another sensor is also a major challenge.To address these challenges,this paper first constructs a lightweight convolutional neural network model for block-based representation and discrimination of built-up areas in high-resolution satellite images,and then achieve refined extraction of built-up areas through the majority voting method based on grid offset.Furthermore,this paper proposes an unsupervised domain adaptation learning method based on adversarial learning,which can achieve transfer extraction of built-up areas on different sensor images by transferring features from the source domain to the target domain without annotating the sample in the target domain.Specifically,the main contributions of this article are two-fold.:(1)A lightweight multi-level feature fusion convolutional neural network has been constructed,significantly improving the accuracy of built-up area mapping in high-resolution images.Given the problem of poor mapping performance for large-scale object classes(e.g.,built-up areas)in high-resolution images due to the complexity of ground features and scenes,this paper adopts a block-based processing strategy and constructs a lightweight multi-scale feature fusion convolutional neural network to achieve block-based feature representation and discrimination of built-up areas.Furthermore,to improve the jagged boundaries caused by block based processing,a post-processing method based on block-shifting is used to achieve refined extraction of built-up areas.The effectiveness of this method was validated by Gao Fen-2 satellite images covering five regions of Shenzhen,China,with F1-score values of 0.9170,0.9235,0.9086,0.8968,and 0.9146,respectively.Compared to current representative built-up area detection algorithms,the proposed method achieved higher recognition accuracy and produced more complete extraction results.(2)A domain adaptive transfer learning method based on adversarial learning is proposed to extract built-up areas from different sensor images.For the satellite images from different sensors covering the same or other areas,there are significant differences in the characteristics of the built-up areas.This paper constructs a domain adaptive method based on adversarial learning to address this problem.This method can transfer source domain feature knowledge to the target domain through adversarial training,enabling efficient extraction of built-up areas in the target domain without additional target domain samples.We selected Gaofen-2 images as the source domain data and tested the proposed method on Sentinel-2 images.The results showed that the proposed method outperformed other transfer learning methods.The above experimental results demonstrate that a lightweight neural network that fuses multi-scale features has significant advantages in extracting built-up areas from high-resolution images,and the block-shifting method considering contextual information can also better optimize the extraction results.In addition,compared to traditional transfer learning methods,the adversarial domain adaptation method can achieve better built-up area extraction with less effort.