ISPRS Urban Segmentation Remote Sensing Dataset

Dataset background

One of the main topics in photogrammetry is the automatic extraction of urban objects from data acquired by airborne sensors. This task is challenging due to the fact that objects such as buildings, streets, trees and cars have a very non-uniform appearance in high-resolution data, which leads to large intra-class variance and low inter-class variance. The focus is on detailed 2D semantic segmentation, assigning labels to multiple object categories. Further research drivers are high-resolution data from new sensors and advanced processing techniques relying on increasingly mature machine learning techniques. Despite great efforts, these tasks cannot be considered solved yet. To the best of our knowledge, there are currently no fully automated methods for 2D object recognition used in practice, despite at least two decades of research trying to solve this task. One of the main issues hindering scientific progress is the lack of standard datasets for evaluating object extraction, making it difficult to experimentally compare the results of different methods. This dataset aims to address this issue.

To this end, the research team provided two state-of-the-art airborne image datasets consisting of very high-resolution real orthophoto (TOP) tiles. Both areas cover urban scenes. While Vaihingen is a relatively small village with many individual buildings and small multi-story buildings, Potsdam is a typical historical city with large building complexes, narrow streets and dense settlement structures.

Dataset structure

The dataset contains 2D semantic segmentations of urban areas such as Vaihinge, Potsdam, and Toronto.
Each dataset was manually classified into the 6 most common land cover classes:

• Impervious Surface (BGR: 255, 255, 255)
• Buildings (BGR: 0, 0, 255)
• Low vegetation (BGR: 0, 255, 255)
• Tree (BGR: 0, 255, 0)
• Car (BGR: 255, 255, 0)
• Clutter/Background (BGR: 255, 0, 0)

The clutter/background class includes water bodies (appearing in two images with parts of a river) and other objects that look very different from other objects (e.g., containers, tennis courts, swimming pools) and are generally not of interest for semantic object classification in urban scenes.