Três décadas de mudanças na planície costeira brasileira: O status dos manguezais, da aquicultura e salicultura a partir de séries temporais Landsat e técnicas de aprendizado de máquina

Abstract

Since the 1980s, land-use and land-cover (LULC) mapping has become a common scientific task. However, the systematic and continuous identification of any terrestrial use or cover, whether on a global or regional scale, demands large storage and processing capacities. This thesis presents two cloud computing pipelines to analyze: 1) the annual status of Brazilian mangroves from 1985 to 2018, along with a new spectral index, the Modular Mangrove Recognition Index (MMRI), which has been specifically designed to better discriminate mangrove forests from the surrounding vegetation, and 2) the annual status of the aquaculture and salt-culture over the Brazilian coastal plains. The mangrove cover showed two distinct occupation periods, 1985-1998 and 1999-2018. The first period shows an upward trend, which seems to be related more to the uneven distribution of Landsat data than to the regeneration of Brazilian mangroves. In the second period, a mangrove loss trend was registered, reaching up to 2% of the mangrove forest. On a regional scale, ~80% of Brazil's mangrove cover is located in the Amazon, Maranhao, Para, Amapa states. In terms of persistence, ~75% of the Brazilian mangroves remained unchanged for two decades or more, especially in the Brazilian Amazon. As for item 2, aquaculture and salt-culture are two of the most classical coastal land-uses worldwide. It isn't different in Brazil, where both land-uses are related to relevant economic activities in the Brazilian Coastal Zone (BCZ). However, to automatically discriminate such activities from other water-related covers/uses is not an easy task. Spectrally speaking, water is water and, unless it presents a high concentration of optically active compounds, not much can be done to dissociate a variety of water-related targets. In this sense, convolutional neural networks (CNN) have the advantage of predicting a given pixel's label by providing as input a local region (named patches or chips) around that pixel. Both the convolutional nature and the semantic segmentation capability allow the U-Net classifier, a type of CNN, to access the "context domain" instead of solely isolated pixel values. Backed by the context domain, the results obtained show that the BCZ aquaculture/saline ponds occupied ~356 km² in 1985 and ~544 km² in 2019, reflecting an area expansion of 52% (~185 km²), a rise of 1.5x in 35 years. From 1997 to 2015, the saline/aquaculture area grew by a factor of ~ 1.7, jumping from 349 km2 to 583 km2, a 67% increase. In 2019, the northeast sector concentrated 93% of the coastal aquaculture/salt-culture surface, 6% in Southeast and 1% in South. Interestingly, despite presenting extensive coastal zones and suitable conditions for developing different aquaculture products, the Amazon coast shows no relevant aquaculture infrastructure sign.