Deep learning para identificação precisa de desmatamentos através do uso de imagens satelitárias de alta resolução

• Autor(a) principal: Taquary, Evandro Carrijo
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFG
• dARK ID: ark:/38995/001300000b5mn
• Texto Completo: http://repositorio.bc.ufg.br/tede/handle/tede/10131
• topic: Aprendizado profundo; LSTM; UNET; Sensoriamento remoto; Desmatamento; Paralelismo; Deep learning; LSTM; UNET; Remote sensing; Deforestation; Parallelism; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

One of the most remarkable advances in Remote Sensing is the devise of the CubeSat satellite building standard. This technology opens up a myriad of possible applications that benefit from the higher spatiotemporal resolutions provided by standard-compatible nanosatellite constellations. In this scenario, one need to investigate the new challenges and how to address them to take advantage of this new type of Remote Sensing Big Data. Among these challenges is the development of means to extract useful information from pixel observations over time in a fine-grained manner. This paper is a seminal study on the use of a special Deep Learning approach, Recurrent Neural Networks, to classify long time series of land cover observations. The method was tested against the problem of identifying areas of deforestation that occurred in a contiguous Cerrado region (17,810 km2 ) over 13 months using high resolution images from PlanetScope, a constellation of CubeSat nanosatellites. In addition to temporal analysis, a solution was needed to make mapping more spatially coherent, which was achieved through the use of a Convolutional Neural Network architecture known as U-Net, in order to perform the semantic segmentation of the temporal analysis result performed in the previous step. The accuracy analysis of the model obtained an F1-score index of 0.9 in identifying deforestation areas of the region of interest over the analyzed period. Given the high performance requirements demanded by the volume of data that this new reality imposes on us, the computational power of parallel processing of a cluster of low cost computers has been explored, enabling the mapping of the studied region to be accelerated up to six times. A discussion of limitations and capabilities of the proposed approach is also presented.