Automatic Building extraction and quantification from Remote sensing images by machine learning

ML algorithms and artificial intelligence are increasingly applied in the field of urban studies. Detailed information about building footprints is required for most urban analyses and modeling. Such an information is not yet readily available for a number of cities, especially Global South cities. Our study aims at comparing the quality of outputs produced by ML using CNN to detect buildings footprints from medium resolution satellite images (Planetscope) and compare these to cadastral parcel data published by Belgian Land registry. In order to decrease computational time and resources, three subsets of areas were taken into account, with varying levels of urban density, namely, Brussels (urban core), Leuven and Nivelles. Our experiments show that CNN-based segmentation networks, and GAN-based generative upsampling networks. detects effectively building units in urban core areas like Brussels with a completeness of 36% which shows a fair agreement with our cadastral data, here considered as the ground truth. However, the result of ML algorithm for peri-urban areas such as Leuven and Nivelles comparatively shows a certain level of underestimation of building, with lower kappa agreement. Landscape metrics and statistical assessment has been carried out at class and landscape level in order to understand the correlation and deviation between the output from ML and the provided cadastral data. These results allow a detailed analysis of both data sets with a variable cell size (eg,100,200,400 and 800m).