Seleção de variáveis para otimização de classificação de desmatamento na plataforma Google Earth Engine

Abstract

The irregular deforestation and suppression of native biomes for the sake of development can have consequences in various areas of the environment, such as the climate. Remote sensing combined with image classification is an important tool for monitoring the environment, but the input information must be curated in order to reduce computational effort and obtain the greatest possible precision and accuracy. This work aimed to verify the cost-effectiveness of different sets of optical and radar images in the classification, using the Random Forest algorithm, of areas of deforestation in the Brazilian Amazon. The study area is located in the municipality of Portel, in the state of Pará, Brazil. The image sets were derived from the Sentinel 1 (SAR) and Sentinel 2 (MSI) constellations for the year 2023. The classifications using all the images per set differed visually and, in their accuracies, with the set composed only of SAR polarizations having an overall accuracy (EG) of around 92%, the set of MSI bands having an average EG of 94%, the set of indices having an EG of around 94.5% and the Complete set having an EG of 95%. The computational resources used on the GEE platform differed due to the use of SAR images or not, with the sets containing SAR images using a greater processing load due to the filters needed to reduce speckle noise (Frost and Quegan&Yu filters). The number of images influenced the amount of memory used for processing, with the classifier using around 8 times more memory when comparing the set with the fewest bands (Sentinel 1 - 4 images) with the set with the most images (Complete - 234 images). The most cost-effective set was the Complete 25% set (58 images), using bands and indices derived from both sensors, with high accuracy and average processing consumption compared to the others. A classification was carried out with this set for the year 2022, which was subtracted from the 2023 classification, generating a layer of deforestation alerts for the year 2023, which when visually compared with the official data released by PRODES 2023, there was agreement in the location and shape of the alerts, performing the classification function well with optimized use of processing.