Deforestation Detection in the Amazon Using DeepLabv3+ Semantic Segmentation Model Variants

Abstract

The Amazon rainforest spreads across nine countries and covers nearly one-third of South America, being 69% inside Brazilian borders. It represents more than half of the remaining tropical forest on Earth and covers the catchment basin of the Amazon river on which 20% of the surface fresh water on the planet flows. Such an ecosystem produces large quantities of water vapor, helping regulate rainfall regimes in most of South America, with strong economic implications: for instance, by irrigating crops and pastures, and supplying water for the main hydroelectric plants in the continent. Being the natural habitat of one-tenth of the currently known species, the Amazon also has enormous biotechnological potential. Among the major menaces to the Amazon is the extension of agricultural and cattle farming, forest fires, illegal mining and logging, all directly associated with deforestation. Preserving the Amazon is obviously essential, and it is well-known that remote sensing provides effective tools for environmental monitoring. This work presents a deforestation detection approach based on the DeepLabv3+, a fully convolutional deep learning model devised for semantic segmentation. The proposed method extends the original DeepLabv3+ model, aiming at properly dealing with a strong class imbalanced problem and improving the delineation quality of deforestation polygons. Experiments were devised to evaluate the proposed method in terms of the sensitivity to the weighted focal loss hyperparameters—through an extensive grid search—and the amount of training data, and compared its performance to previous deep learning methods proposed for deforestation detection. Landsat OLI-8 images of a specific region in the Amazon were used in such evaluation. The results indicate that the variants of the proposed method outperformed previous works in terms of the F1-score and Precision metrics. Additionally, more substantial performance gains were observed in the context of smaller volumes of training data. When the evaluated methods were trained using four image tiles, the proposed method outperformed its counterparts by approximately +10% in terms of F1-score (from 63% to 73%); when the methods were trained with only one image tile, the performance difference in terms of F1-score achieved approximately +18% (from 49% to 67%).