基于改进DeepLabv3plus算法的遥感图像海岛建筑提取方法

目前海岛经济快速发展，为避免海岛建筑无序扩建，了解海岛建筑分布特征尤为重要。机器学习方法是从高分遥感影像提取地物目标的常见方式，然而建筑物遥感特征复杂，机器学习方法出现鲁棒性差、难以充分挖掘深层次特征的弊端。文章提出基于DeepLabv3plus网络模型的深度学习语义分割方法提取海岛建筑，并对网络结构进行改进，使用组归一化（GN）方法替代批归一化（BN）以适合小batch size下的语义分割操作。针对海岛建筑数据量较少的问题，采用迁移学习策略，设计基于多源数据的国内城市建筑数据集的预训练样本智能采集和标注方法，再人工标注中国部分海岛建筑进行算法实验。结果表明，在batch size较小时，基于GN的DeepLabv3plus语义分割算法的平均精度和mIoU均得到提升，能够获得更为精确的像素级海岛建筑提取结果。

Extraction Method of Island buildings in Remote Sensing Images Based on Improved DeepLabv3plus Algorithm

As a typical land element, buildings play a dominant role in land use, urban planning, and disaster risk assessment, which is an important goal of remote-sensing-based surveying and mapping. Presently, with the rapid development of island economies, it is particularly important to extract island buildings accurately to avoid disorderly expansion. Machine learning is a common method for extracting terrain objects from high-resolution remote sensing images. However, the remote sensing features of buildings are complex: traditional machine learning methods suffer from poor robustness and have difficulty in fully mining deep-seated features. The overall accuracy and Mean Intersection over Union (mIoU) score of the deep learning architecture DeepLabv3plus is higher than those of other semantic segmentation models, making it suitable for remote sensing building extraction. In this study, the DeepLabv3plus segmentation architecture is adapted, based on the Xception network and hole convolution, to reduce the large number of pictures (batch size) and video memory required to train a deep learning model. Group Normalization (GN) is used to replace the original Batch Normalization (BN), to improve the feature extraction accuracy for a small batch size. Compared with the original DeepLabv3plus architecture, the DeepLabv3plus-G method can train models with a smaller batch size under limited Video Graphics Array (VGA) memory, and the segmentation accuracy is improved. DeepLabv3plus-G can effectively extract island buildings, which provides a new method for Chinese island construction supervision. As a result of limited island building data, this study designs an intelligent collection of pre-training sets and an annotation method based on multi-source data of domestic urban building datasets. Transfer learning strategy is also adopted to train the network model on the automatically annotated domestic urban building datasets. The final pre-training model was saved in the form of a network architecture and weights. The model size was 158 MB, the number of parameters was 22910480, and the depth was 65 layers. Remote sensing images of buildings from 29 temperate islands (including Chrysanthemum Island, etc.), 48 subtropical islands (including Chongming Island, etc.), and six tropical islands (including Haidian Island) in China from Google Earth, with a spatial resolution of 0.6 m, were employed. After automatic clipping, 891 images of island buildings with dimensions of 512 × 512 pixels were obtained. The open source labeling software Labelme was used for manual labeling before the algorithm experiments. Results showed that both the mIoU and mean Average Precision (mAP) of DeepLav3plus-Gare improved under the same batch size, compared to traditional methods. When the batch size was two, the mIoU and mAP improved by 8.95% and 10.45%, respectively, compared to the scores for the DeepLabv3plus architecture. Under the same network, the accuracy of DeepLabv3plus pixel segmentation improved with increasing patch size. The mIoU and mAP scores for a model with a batch size of eight were 10.51% and 10.26% higher than those of models with a batch size of two, respectively. However, the mAP of DeepLabv3plus-G pixel segmentation hardly changed with an increase in the batch size. In conclusion, when the batch size is small, the mAP and mIoU of the DeepLabv3plus building classification algorithm based on GN increase, and more accurate pixel-level island building extraction results are obtained.