A multi-criteria optimization approach for the geometric integration of vector-based land cover datasets

Land Use/Land Cover (LULC) mapping plays a pivotal role in understanding the complex interplay between human activities and the natural environment. Here, we present a novel approach for fusing LULC information from two different data sources. Specifically, we aim to combine the advantages of LULC information extracted from remote-sensing imagery and of LULC information mapped by volunteers, especially LULC information from OpenStreetMap. Since OpenStreetMap is based on vector representations of objects and since we aim to generate a harmonized vector layer as output, our approach is to conduct the information fusion in a vector environment, using methods from computational geometry. Our method starts by vectorizing the remote sensing data, yielding a set of polygons. We then overlay the boundaries of these polygons with the boundaries of the LULC polygons from OpenStreetMap and refine the resulting planar subdivision using a Constrained Delaunay Triangulation (CDT). This triangulation is the starting point for the main step of our approach. In this step, for a land use type of interest, we generate the output polygons by computing an optimal selection of triangles from the CDT and resolving the boundaries between adjacent triangles in the selection. The novelty of our approach is that we optimize an objective function that combines five explicitly modeled criteria: (1, 2) adherence to the boundaries of the first/second input layer, (3, 4) adherence to the areas of the first/second input layer, and (5) generalizing the output polygons to a desired level of detail. We show how to efficiently solve our model using an approach based on graph cuts. Application of the methodology to forest land cover portions in North Rhine-Westphalia, Germany, reveals promising results, with outputs closely resembling reference data. With an average Intersection over Union (IoU) of 0.831 and a discrete Hausdorff distance ([Formula: see text]) of 1251.37m, across seven regions of interest, we demonstrate the effectiveness of our approach in accurately mapping LULC through dataset fusion in a vector framework, achieving optimal geometric harmonization.