Deep neural networks for above-ground detection in very high spatial resolution digital elevation models

Abstract

Deep Learning techniques have lately received increased attention for achieving state-of-the-art results in many classification problems, including various vision tasks. In this work, we implement a Deep Learning technique for classifying above-ground objects within urban environments by using a Multilayer Perceptron model and VHSR DEM data. In this context, we propose a novel method called M-ramp which significantly improves the classifier's estimations by neglecting artefacts, minimizing convergence time and improving overall accuracy. We support the importance of using the M-ramp model in DEM classification by conducting a set of experiments with both quantitative and qualitative results. Precisely, we initially train our algorithm with random DEM tiles and their respective point-labels, considering less than 0.1% over the test area, depicting the city center of Munich (25 km2). Furthermore with no additional training, we classify two much larger unseen extents of the greater Munich area (424 km2) and Dongying city, China (257 km2) and evaluate their respective results for proving knowledge-transferability. Through the use of M-ramp, we were able to accelerate the convergence by a magnitude of 8 and achieve a decrease in above-ground relative error by 24.8% and 5.5% over the different datasets.