Multi-scale automatic building extraction and detection from remote sensing imagery and lidar data based on deep learning models

Building, as the main place of human production and life, is an important component of basic geographic information elements. With the acceleration of urbanization, building information extraction using remote sensing technology is vital in many domains, such as urban planning, real estate management, land use, and intelligent city construction. Automatic building detection and extraction algorithms can provide precise location and spatial range information, significantly reducing time costs and labor. Traditional methods rely on manual design and shallow features with weak generalization ability and automatic interpretation. Recently, deep learning algorithms, especially convolution neural networks (CNNs), have robust feature extraction ability, achieving advanced performance. However, buildings often exhibit multi-scale, spectral heterogeneity and variability with complex geometric shapes. In addition, environmental factors bring obstacles to building feature extraction. CNN-based methods cannot effectively model the global context and ignore multi-scale semantic representation. Therefore, the building extraction results often exhibit fragmentation and lack spatial details using CNN-based methods. Moreover, the semantic segmentation approaches cannot obtain instance information. To address these issues, this study developed two improved CNN-based models: Multiscale Building Extraction Network (MBEN) and Building Instance Segmentation Network (BISN). MBEN improved the residual CNN structure by introducing dilated convolution and using a dual-branch structure to extract LiDAR and image features. Additionally, two novel optimization modules have been developed to improve the feature extraction ability of MBEN. Concretely, a multi-scale contextual optimization module based on CNN was developed to enhance global semantic correlations and multi-scale representation capabilities. This module can effectively alleviate the fragmentation of extraction results due to spectral heterogeneity and variability. The semantic-guided attention module was proposed to refine and restore fine-grained spatial features for buildings. Furthermore, BISN was constructed to obtain object-level detection information for buildings. Meanwhile, local spatial-spectral perceptron and cross-level feature fusion modules were developed to enhance multi-modal feature fusion for BISN. In addition, an adaptive center point detector (ACPD) was developed to improve the accuracy of location and scale prediction. To verify the performance of the proposed methods, three open-source datasets were used in the experiments, including the Wu Han University Building dataset (WHU), the Building Instance Segmentation dataset (BISM), and the Buildings Off-Nadir Aerial Image (BONAI) dataset. The results of the ablation experiments confirmed that the proposed modules can improve average extraction accuracy by 1.5% and F1 score by 3%. Average detection accuracy was improved by about 2% in the test dataset. Overall, quantitative analysis and visual experiments confirmed that the proposed modules are superior to other state-of-the-art methods, with 93.19% intersection of union (IOU) and 97.56% overall accuracy (OA) on WHU dataset, 94.72% IOU and 97.84% OA on BISM dataset, 95.43% IOU and 98.37% OA on BONAI dataset. The proposed instance detection framework improves the inference speed and detection accuracy for building, which is about 37% faster than the two-stage methods with 94% average precision.

Text FK 2024 51 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (911kB) Text FK 2024 51 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (16MB) Text FK 2024 51.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (16MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/19016

Actions (login required)

View Item