24 April 2024 in ISPRS Journal of Photogrammetry and Remote Sensing
18 July 2023 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Urban areas predominantly consist of complex building structures, that are assembled of multiple building sections. From very high resolution (VHR) remote sensing imagery, not only roof-tops but also the separation lines between them are visible. Since fully convolutional neural network (FCN)-based methods have become the primary choice in segmentation approaches, they have been extensively used for automatic building footprint extraction. But each of the previous works on building segmentation either lacks separation of building blocks into sections or does not produce sections of regular appearance. We propose a two stage approach to overcome these limitations. The first step segments building and separation lines using an FCN model and the second step produces building instances by using a learning-free method. Our model receives a top-down image and a digital surface model (DSM) patch in two separate encoders. The encoder features are summed before the skip-connections, which utilize the encoder features from the current and higher resolution feature maps. We train our model with regularization losses for building shapes and separation lines on both satellite and aerial imagery. We test our model on a city that was not previously included in the training phase to show that it has the capacity to generalize across different geographical locations and architectural styles. Furthermore, we use our building section instance predictions to generate (a) vectorized building maps and (b) a level of detail (LoD) 1 DSM.
5 December 2023 in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Deep learning is a powerful tool to extract both individual building and roof plane polygons. But deep learning requires a large amount of labeled data. Hence, publicly available level of detail (LoD)-2 datasets are a natural choice to train fully convolutional neural networks (FCNs) models for both building section and roof plane instance segmentation. Since publicly available datasets are often automatically derived, e.g. based on laser scanning, they lack on annotation accuracy. To complement such a dataset, we introduce manually annotated and synthetically generated data. Manually annotated data is domain-specific and has a high annotation quality but is expensive to obtain. Synthetically generated data has high-quality annotations by definition, but lacks domain-specificity. Moreover, we not only detect individual building section instances, but also roof plane instances. We predict separations not only between individual buildings, but also by a class that describes the line which separates roof planes. The predicted building and roof plane instances are polygonized by a simple tree search algorithm. To achieve more regular polygons, we utilize the Douglas-Peucker polygon simplification algorithm. We describe our dataset in detail to allow comparability between successive methods. To facilitate future works in building and roof plane prediction, our Roof3D dataset is accessible at https://github.com/dlrPHS/GPUB.
25 November 2024 in PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science
Building extraction and polygonization is important for urban studies, such as urbanization monitoring, urban planning. Remote sensing images, especially in RGB bands, provide sufficient semantic information which is useful for the task of building extraction and polygonization. Deep learning using Convolutional Neural Networks (CNNs) is proven to be successful in many fields, including building extraction from remote sensing images. In this paper, we propose a two-stage method to solve the task of building polygonization from remote sensing images based on deep learning. Firstly, we decompose a 2‑D building footprint model into three basic geometry primitives. Leveraging stacked Multi-Branch Modules (MBMs), we separate the task of building extraction into tasks of predicting the three geometry primitives using our proposed CNN. At the second stage, we propose an efficient enhanced building polygonization and adjustment algorithm to generate the final building polygons. This algorithm is able to handle both building blocks and individual buildings. We evaluate our model on three open datasets. For building blocks, our model achieved average precision of 62.7% and average recall of 73.6% on the CrowdAI mapping challenge dataset, and 13.9% and 24.4% respectively on the Urban Building Classification (UBC) dataset which contains mainly individual buildings. On the Inria aerial image dataset, the proposed method achieved Intersection over Union (IoU) over 71%.