Adversarial domain adaptation for deforestation detection in remote sensing imagery

Abstract

Semantic segmentation models aim at classifying images at the pixel level. In general terms, training such models with the traditional supervised approach requires sufficient amount of images and corresponding class label maps. While state-of-the-art deep semantic segmentation networks offer high classification performance, producing the references for supervised training often proves to be quite laborious and costly. Additionally, the accuracy delivered by those networks is directly impacted by the quality and volume of training data. Moreover, the resulting classifiers are, in general, domain specific, what means that after being trained with specific domain data, a significant performance drop is expected when evaluating them on data from another domain, even when dealing with the exact same classification task. In the context of remote sensing applications, a domain is represented by images from different sites, related to different landscapes and/or captured at different dates, likely with different acquisitions conditions. Alike other remote sensing applications, deforestation detection tends to present a poor accuracy when evaluated in a cross-domain scenario. As solution to mitigate such a problem, this work investigates the use of unsupervised domain adaptation techniques combined in a novel method. Despite requiring source domain data alongside the respective class labels, the devised method needs no references for the target domain data during training. Our solution, specialized for deforestation detection, combines two domain adaptation strategies, namely, appearance adaptation and representation matching. In the experimental analysis, we assess the performance of different variants of the proposed method, and compare their outcomes with those delivered by state-of-the-art domain adaptation methods for deforestation detection, over forest areas in the Brazilian Amazon and Cerrado biomes.