2022: Returning the favor - Leveraging quality insights of OpenStreetMap-based land-use/land-cover скачать в хорошем качестве

2022: Returning the favor - Leveraging quality insights of OpenStreetMap-based land-use/land-cover

https://media.ccc.de/v/state-of-the-m... The fitness of OSM for multi-label classification is proven. A workflow to enhance OSM-based multi-labels using machine learning is established. The results are provided to the OSM community via the HOT Tasking Manager. Introduction Land-use and land-cover (LULC) information in OSM is a challenging topic. On the one hand, this information provides the background for all other data rendered on the central map and is used by applications like https://osmlanduse.org. On the other hand, this information has a difficult position within the OSM ecosystem. LULC information can be quite cumbersome or even difficult to map e.g. due to natural ambiguity. The growing tagging scheme provides a collection of sometimes ambiguous or overlapping tag definitions that are not fully compatible with any official LULC legend definition [1]. Furthermore, the data is highly shaped by national preferences and imports. This diversity of the LULC data in OSM is a fundamental principle of OSM that enabled the success of the project. Yet, this can create considerable usage barriers or at least caveats for data users unfamiliar with the projects' ecosystem. The remote sensing community for instance has started to use OSM LULC information as labels in their classification models. Frequently, OSM LULC data has thereby been taken at face value without critical reflection. And, while the quality and fitness for purpose of OSM data has been proven in many cases (e.g. [2,3]) these analyses have also unveiled quality variations e.g. between rural and urban regions. The quality of OSM therefore can be assumed to be generally high, but remains unknown for a specific use-case. The proposed work first assesses the impact of these challenges on a use-case of multi-label remote sensing (RS) image classification and then provides a machine learning (ML) based workflow to overcome and finally mitigate them. Multi-labels are a type of image classification where a satellite image is labeled with multiple containing LULC classes. In the presented study these labels are extracted from OSM and used to train the ML algorithm. Methods and Results The fitness for purpose of OSM for multi-label RS image classification was tested on a Sentinel 2 scene with a resolution of 10m and four bands in south west Germany recorded in June 2021. The area was chosen for its estimated high completeness and low amount of imported data. OSM data was grouped by its tags into the four LULC classes 'forests', 'agricultural areas', 'build-up area' and 'water bodies'. 18 tags that could unequivocally be mapped to these classes were used and small elements below the image resolution or the classes minimal mapping unit were filtered. The chosen scene was then tiled into a 1.22 x 1.22 km grid of 8100 image patches. Zero to four labels were assigned to each patch, based on the OSM LULC elements therein. Evaluation was performed manually on 910 random patches, of which 80% had a correct OSM-based multi-label, thereby proving the assumed high completeness and quality in the region. The proposed workflow provides a method to enhance this OSM-based RS image multi-label classification and extend it to areas of lower OSM quality and completeness using ML (specifically deep learning (DL)). The main obstacle for ML and especially DL is the required amount of labeled training data. Volunteered geographic information (VGI) like OSM offers a potential solution to this challenge by providing an overabundance of LULC information that is suitable for this purpose if data quality is sufficiently high. The workflow uses the multi-label information extracted from OSM for training and then detects discrepancies between its predictions and OSM. Using this information and pinpointing the exact location of error within the patches provides valuable OSM data quality information. Apart from facilitating a fast quality estimation for large areas, the workflow can make its findings automatically available to the OSM community in a feedback loop using the HOT Tasking Manager framework. Thereby the valuable service by the OSM community of providing large amounts of free and generally high quality training data is recognised in the form of quality feedback including mapping hints to the OSM community. The five workflow stages are: 1) RS data collection and preprocessing, 2) OSM data collection and preprocessing, 3) LULC multi-label modeling, 4) OSM data issue flagging and 5) the community feedback loop. While each step is an atomic use case and application, the combination of all four steps creates a tool that is useful for the RS and the OSM community likewise. The tool is openly available at https://gitlab.gistools.geog.uni-heid... under the GNU Affero General P