Adaptação adversarial de domínio para a detecção de desmatamento em imagens de sensoriamento remoto

Abstract

Deforestation detection in a specific region can be achieved from semantic segmentation with convolutional neural networks. However, given the differences in forest coverand other aspects related to image acquisition, such as solar incidence, view angle, among others, using a network trained in a geographic region usually results in a considerable loss of accuracy when the network is used to classify a different region. We can refer to each region as a domain. The training domain, with labeled samples (ground truth) available during training, is the source domain, and the other is the target domain. It is observed that the production of training labels for the possible different domains, considering the large extensions of forest biomes, would have a prohibitive cost. However, it is possible to use domain adaptation methods to minimize the differences between domains and obtain better classification accuracy in the target domain. With those methods, it is possible to adopt two strategies: the first is visual adaptation so that images from the target domain, after being adapted, resemble those from the source domain, and then a better classification result is achieved on the adapted images with the classifier trained with the source domain; the second consists of using domain adaptation to train a classifier for the target domain also. The objective of this work is to compare the different domain adaptation strategies in the context of deforestation detection, and to evaluate the impact of alternative components of loss functions. For the evaluation, images belonging to 3 different regions in Brazil were used, two belonging to the Brazilian Legal Amazon (Rondônia and Pará) and another in the transition between the Cerrado biome and the Amazon forest (in Maranhão). Each of those regions represent a domain. For visual domain adaptation, a model inspired by CyCADA was used, which in turn is based on the CycleGAN architecture. With the CycleGAN, visual domain adaptation was performed in a completely unsupervised way. With the model inspired by CyCADA, labels from the source domain are used during the training in order to obtain better results in the classification step. A network with the DeepLabV3+ architecture, with the focal loss function, was used for the classification. In the experiments, it was observed that when the visual domain adaptation of the domain pairs promotes good results with the source classifier (first strategy), the use of a specific target classifier (second strategy) is usually associated with worse results. While the reverse behavior can also be observed: domain pairs for which poor results are obtained with the first strategy generally promote better results when the second strategy is used. As expected, none of the results using domain adaptation achieved the result obtained with supervised training on the target domain. But experiments show that it is possible to achieve good and, in almost all cases, better results than is possible without any domain adaptation.