Unsupervised and self-supervised deep learning networks for semantic segmentation of images have made impressive progress in the last years. They can be trained without any labelled data and yet are able to effectively segment RGB images into meantingful semantic groups. In remote sensing, supplementary information, such as elevation, improves class separation by differentiating classes based to their height above ground.
We take SmooSeg and guide its training process by infusing elevation information into its projector and smoothness prior, to ensure global label consistency across the entire dataset. This improves the segmentation performance, since patches of the same semantic group often exhibit similar elevation characteristics.
We also extend the Conditional Random Field (CRF) to refine the low resolution segmentation results in a post-processing step with elevation information by introducing a second pairwise potential that encourages neighboring pixels with similar elevation to have the same label, ensuring local label consistency.
Our multi-modal training strategy remains unsupervised and improves the segmentation performance on the ISPRS Potsdam-3 dataset by +4% in mIoU over the RGB-only SmooSeg and by 4.3% using the multi-modal CRF post-processing. Collectively, our approach surpasses all state-of-the-art unsupervised segmentation networks that rely solely on RGB data for the Potsdam-3 dataset, highlighting the important role of elevation data in label-free segmentation for remote sensing applications.

In this paper, we explore the application of domain adaptation techniques for semantic land cover segmentation using aerial remote sensing data. We leverage canonical correlation and histogram matching to facilitate the transfer of knowledge from pre-trained classification models to new datasets without the need for additional labeled data. Specifically, we perform Canonical Correlation to align feature distributions between the source and target domains and Histogram Matching to enhance the correspondence of pixel distributions across datasets.
The effectiveness of these domain adaptation techniques is assessed through improvements in semantic segmentation performance of the Random Forest and DeepLabV3+ classifiers on German city datasets. Our results indicate a substantial increase in segmentation accuracy when using domain adaptation methods. Furthermore, we examine the role of elevation data, represented by Normalized Digital Surface Models (NDSM), which enhances segmentation performance on unseen datasets. These findings underscore the efficacy of domain adaptation and the value of elevation data in remote sensing classification, particularly in dynamic environments where models encounter new datasets.

This article presents a method for the automated detection and quantification of cracks on masonry surfaces. The core of the approach is a neural network trained for semantic segmentation, which enables the identification of cracks in image data. To facilitate a physically meaningful analysis, the image data is combined with 3D geometric information. A 3D point cloud is projected onto the image plane to establish correspondences between 2D image points and 3D spatial coordinates. These 2D–3D correspondences are utilized to evaluate the detected cracks in a geometrically accurate manner. Based on the segmentation results and the projected 3D data, cracks can be classified within the point cloud and analyzed metrically. The Crack length is determined using a graph-based model, in which the crack structure is represented as a network and the longest continuous crack path is computed using Dijkstra’s algorithm. The Crack width is measured in the images based on the segmentation masks and a scaling factor derived from the 2D–3D correspondences. The proposed method enables a precise and automated assessment of crack patterns in masonry structures by leveraging both image and 3D data.