Reconhecimento de padrões de crescimento urbano em imagens de satélite utilizando aprendizagem profunda e teoria dos valores extremos.

Abstract

The recognition of urban growth patterns in satellite images has applications that range from understanding the dynamics of urbanization to inferring future urban expansion. Both the availability of global inventories of land use, based on remote sensing, and advances in deep learning methods, offer an opportunity to boost the state of the art of existing models for this purpose. This task has broad implications for disaster preparedness, the environment, infrastructure development, and epidemic prevention, as well as developing new computer vision methods for time series data. Inspired by sequential models, this work proposes a method for detecting novelties, or anomalies, using the Peaks Over Threshold (POT) algorithm, a parametric probabilistic approach based on the Theory of Extreme Values that does not require manually defined thresholds. and does not presuppose data distribution. The algorithm was applied to representations obtained by a convolutional neural network (U-Net architecture) in order to recognize and detect possible changes in the geography of the regions, taking advantage of a temporal sequence of remote sensing images extracted from the SpaceNet dataset. The results show that, despite the moderate resolution of the data, it was possible to track identifiers of changes in the Earth’s surface temporally. The results validate the effectiveness of the proposed method in detecting anomalies with results of 91.34% and 85% for F-score and recall respectively, as well as an F-beta of 87.42%, a score that represents the average precision weighted harmonica and recall. In comparison with the literature, the reported results are 90%, 71.16% and 69% for the F-score, exclusively, without recording additional metrics. The use of the POT algorithm in conjunction with a convolutional network of U-Net architecture, applied to the SpaceNet dataset, has brought experimental evidence that it is a promising approach to automatically provide detections of spatio-temporal changes or alterations on the Earth’s surface in practical applications.