With the introduction of artificial intelligence (AI), current deep learning-driven change detection methods generally regard changes as a type of specific land cover information and try to detect them simply by using existing semantic labeling models. In practice, the changes that occur on different land cover categories may appear completely different for remote sensing images, making it difficult to detect multiple categories of changes with an end-to-end model. In addition, training such networks requires a large amount of prelabeled references, which are labor-intensive and time-consuming. Motivated by this observation, we integrate dual learning algorithm and disentangled representation theory to develop a novel approach, named content-invariant dual learning (CiDL), for either supervised or unsupervised change detection in remote sensing images. In our framework, two opposite Y-shaped networks, each of which consists of two encoders and one decoder, are introduced to translate bitemporal images from their original domains to each others, where their intrinsic content features are retained, while their style features are consistent. By training our hybrid framework, even without references, this method learns a category-wise cross-domain translation to suppress the discrepancies in paired unchanged regions, and meanwhile highlight those in paired changed regions. The experimental results on two typical change detection data sets and the comparison with other state-of-the-art deep learning-driven methods verify the effectiveness and competitiveness of our proposed CiDL.

Coastal land cover mapping is a significant yet challenging pixel-level segmentation task. Domain shift between optical remote sensing imagery will give rise to remarkable performance degradation for deep supervised methods. Besides, the ground objects characterized with interclass variance and class imbalance may further aggravate the adverse effect. Traditional adversary-based domain adaptation algorithms always leverage a binary discriminator to conduct global adaptation, ignoring the detailed class information. In this article, we develop a novel class-aware domain adaptation method to address these issues. Unlike the naive single one, we propose a joint local and global adversarial adaptation framework to separately execute class-specific and global domain alignment on feature and output spaces. For the former, the introduced classwise discriminator possesses different strategies to extract labels for both data domains. Meanwhile, we restore to entropy minimization to produce high-confident target prediction rather than using the early generated pseudo label with high confidence. Furthermore, we additionally adopt comprehensive reweighting on the supervised segmentation loss to track the class imbalance problem. This manner mainly comprises the sample-based median frequency balancing and the focal loss function for the minority and hard classes. We measure the proposed method on two typical coastal datasets and compare it with other state-of-the-art models. The experimental results confirm its excellent and competitive performance on cross-domain land cover mapping.

High-quality land-use/land-cover mapping with optical remote sensing images yet presents significant work. Even though fully convolutional semantic segmentation models have recently contributed to popular solutions, the lack of annotation data may lead to severe degradations in their inference performance. Besides, the category confusion in high-resolution representations will further exacerbate the adverse effects. In this paper, we propose a category-sensitive semi-supervised semantic segmentation framework to address these weaknesses by employing massive unlabeled data. With the perturbations from adopted hybrid data augmentation structures, we first focus on the output space and execute regularization constraints to learn category-specific discriminative features. It is formulated with a consistency self-training procedure where a dynamic class-balanced threshold selection scheme is proposed to provide high-confident pseudo supervisions for each category. In addition, we introduce pixel-wise contrastive learning on the common embedding space from both labeled and unlabeled data domains to further facilitate the semantic dependencies among category features, in which the reliable labels are leveraged as guidance for pixel sample selection. We verify the proposed framework on two benchmark land-use/land-cover datasets, and the experimental results demonstrate its competitive performance to other state-of-the-art semi-supervised methods.