Unsupervised Domain Adaptation for Cloud Detection Based on Grouped Features Alignment and Entropy Minimization
ABSTARCT :
Most convolutional neural network (CNN)-based cloud detection methods are built upon the supervised learning framework that requires a large number of pixel-level labels. However, it is expensive and time-consuming to manually annotate pixelwise labels for massive remote sensing images. To reduce the labeling cost, we propose an unsupervised domain adaptation (UDA) approach to generalize the model trained on labeled images of source satellite to unlabeled images of the target satellite. To effectively address the domain shift problem on cross-satellite images, we develop a novel UDA method based on grouped features alignment (GFA) and entropy minimization (EM) to extract domain-invariant representations to improve the cloud detection accuracy of cross-satellite images. The proposed UDA method is evaluated on ``Landsat-8 ? ZY-3'' and ``GF-1? ZY-3'' domain adaptation tasks. Experimental results demonstrate the effectiveness of our method against existing state-of-the-art UDA approaches. The code of this paper has been made available online (https://github.com/nkszjx/grouped-features-alignment).
EXISTING SYSTEM :
? Existing works for domain adaptation in semantic segmentation focus on RGB data.
? Therefore, when applying existing models on data with a different distribution than the training data, i.e., from a different domain, the performance is considerably degraded.
? Coping with these issues would enable the use of large existing labeled LiDAR datasets for more realistic use-cases in robotic applications, reducing the need for data labeling.
? SqueezeSegV2 based the adaptation on existing adaptation works like correlation alignment .
? This work investigates different UDA strategies (both existing and novel) to improve UDA for the particular case of LiDAR semantic segmentation.
DISADVANTAGE :
? The first strategy addresses this problem by applying a set of simple steps to align the data distribution reducing the domain gap on the input space.
? As far as deep learning methods are concerned, there are two main types of approaches to tackle the 3D LiDAR semantic segmentation problem.
? Regarding segmentation on LiDAR data, very few works have studied the problem of domain adaptation.
? Depending on the problem tackled and prior knowledge, we can hypothesize which of these differences can be neglected and assumed not to affect to the models we are learning.
PROPOSED SYSTEM :
• They propose to apply the KL divergence between the output probabilities of both modalities as the main loss function. Besides, they also apply previously proposed methods like entropy minimization.
• The proposed approach achieves better results than the other baselines in the three different scenarios for unsupervised domain adaptation in LiDAR Semantic Segmentation.
• These two proposed strategies can be applied in conjunction with current state-of-the-art approaches boosting their performance.
• Besides, we propose a learning-based approach that aligns the distribution of the semantic classes of the target domain to the source domain.
ADVANTAGE :
? As evaluated on extensive domain adaptation tasks, our proposed method achieves stateof-the-art classification performance on both vanilla unsupervised domain adaptation and partial domain adaptation.
? Domain adaptation attempts to boost the performance on a target domain by borrowing knowledge from a well established source domain.
? These approaches are suitable for those scenarios where the source and target domains share the same support, thus they cannot achieve satisfactory performance in the wild scenarios.
? Compared with these two novel methods, our method still achieves the best performance, demonstrating the advantage and necessity of considering the relation between the global and local distribution alignment.
|