Deforestation Detection With Weak Supervised Convolutional Neural Networks in Tropical Biomes

Abstract

Abstract. Deep learning methods are known to demand large amounts of labeled samples for training. For remote sensing applications such as change detection, coping with that demand is expensive and time-consuming. This work aims at investigating a noisy-label-based weak supervised method in the context of a deforestation mapping application, characterized by a high class imbalance between the classes of interest, i.e., deforestation and no-deforestation. The study sites correspond to different regions in the Amazon and Brazilian Cerrado biomes. To mitigate the lack of ground-truth labeled training samples, we devised an unsupervised pseudo-labeling scheme based on the Change Vector Analysis technique. The experimental results indicate that the proposed approach can improve the accuracy of deforestation detection applications.

Weakly Supervised Domain Adversarial Neural Network for Deforestation Detection in Tropical Forests

“…al. [14]. In that work, a DL-based change detection classifier is trained on target domain samples using noisy labels computed by a CVA-based algorithm, proposed in [15].…”

“…Recently, with the rise of Deep Learning (DL) [8] as the dominant trend in most image analysis applications, including those based on RS imagery, many DL-based methods have been proposed for change-detection, e.g., [9], [10], including deforestation monitoring, e.g., [11]- [13]. However, despite the success of DL methods in those and other application fields, such techniques are known to demand extensive training data for suitable training [14]. Furthermore, for the deforestation mapping application, the geographic variability of forest cover and deforestation practices conspire against the good gen-eralization of DL models.…”

“…In a more recent work, which was not concerned with domain adaptation, the same group of authors [14] explored a weakly-supervised approach [33] using such pseudo-labels in the training of a DL model for deforestation detection. In both [14], [15], the pseudo-labels generated by CVA proved valuable in mitigating both the negative effect of class imbalance in DA training and the scarcity of labeled samples for training deforestation detection models. Specifically, to face the scarcity of labeled samples for training, diverse initiatives such as weakly supervised and semi-supervised learning have been proposed.…”

“…Specifically, to face the scarcity of labeled samples for training, diverse initiatives such as weakly supervised and semi-supervised learning have been proposed. In short, weakly-supervised learning intends to train classification models using noisy or low-quality labels in the target domain [14]. Potential improvements are directly related to the confidence level on such noisy labels and the CNN's capacity to consider the spatial context in classifying a single pixel.…”

“…Motivated by the findings of [15] and [14], we propose a weakly-supervised Domain Adversarial Neural Network (DANN WS) for deforestation detection. We aim to improve the DANN's feature space discriminability using the pseudolabels generated by the algorithm proposed in [15].…”