Change detection in synthetic aperture radar (SAR) images is a hot yet highly challenging task in remote sensing. Existing unsupervised SAR change detection methods often struggle with inherent speckle noise and insufficiently utilize pseudo-labels, particularly neglecting uncertain areas. In this paper, we propose a multilevel difference-enhanced denoise dual-branch network (MDDNet), comprising representation learning and change detection branches. First, fuzzy c-means clustering is employed to generate pseudo-labels, categorizing the image areas as changed, nochanged, and uncertain. Second, we design a denoise representation loss function in the representation learning branch to maximize the use of pseudo-labels, while mitigating speckle noise. Furthermore, a multilevel difference computation module is proposed to focus on changes in ground objects and capture more comprehensive change information. Experimental results on three public SAR datasets show that the proposed method outperforms six state-of-the-art methods, achieving the best performance with an average overall accuracy of 98.86% and an average Kappa coefficient of 89.36%.

Large Language Models (LLMs) have shown promising results in repository-level code completion, which completes code based on the in-file and cross-file context of a repository. The cross-file context typically contains different types of information (e.g., relevant APIs and similar code) and is lengthy. In this paper, we found that LLMs struggle to fully utilize the information in the cross-file context. We hypothesize that one of the root causes of the limitation is the misalignment between pre-training (i.e., relying on nearby context) and repo-level code completion (i.e., frequently attending to long-range cross-file context).To address the above misalignment, we propose Code Long-context Alignment - CoLA, a purely data-driven approach to explicitly teach LLMs to focus on the cross-file context. Specifically, CoLA constructs a large-scale repo-level code completion dataset - CoLA-132K, where each sample contains the long cross-file context (up to 128K tokens) and requires generating context-aware code (i.e., cross-file API invocations and code spans similar to cross-file context). Through a two-stage training pipeline upon CoLA-132K, LLMs learn the capability of finding relevant information in the cross-file context, thus aligning LLMs with repo-level code completion. We apply CoLA to multiple popular LLMs (e.g., aiXcoder-7B) and extensive experiments on CoLA-132K and a public benchmark - CrossCodeEval. Our experiments yield the following results. ❶ Effectiveness. CoLA substantially improves the performance of multiple LLMs in repo-level code completion. For example, it improves aiXcoder-7B by up to 19.7% in exact match. ❷ Generalizability. The capability learned by CoLA can generalize to new languages (i.e., languages not in training data). ❸ Enhanced Context Utilization Capability. We design two probing experiments, which show CoLA improves the capability of LLMs in utilizing the information (i.e., relevant APIs and similar code) in cross-file context. Our datasets and model weights are released in [1].

Heterogeneous change detection (HCD) is a process to determine the change information by analyzing heterogeneous images of the same geographic location taken at different times, which plays an important role in remote sensing applications such as disaster response and environmental monitoring. However, the different imaging mechanisms result in different visual appearances in heterogeneous images, making it difficult to accurately detect changes through direct comparison. To address this problem, we propose a inter-modal and cross-modal self-supervised dual branch learning framework (ICSF) for HCD that incorporates inter-modal and cross-modal learning. First, in the inter-modal branch, we perform contrastive learning on heterogeneous images within their respective modalities to learn the robust and discriminative features, rather than relying on the raw spectral or spatial information from these images. Second, in the cross-modal branch, we perform cross-modal reconstruction to ensure the obtained features exhibit consistent comparability, thereby facilitating the extraction of rich information on the real changes within the images. Next, the difference images (DIs) computed from both branches are further refined using a superpixel segmentation strategy to preserve the consistency of differences within the same ground object. Experimental results on five public datasets with different modality combinations and change events demonstrate the effectiveness of the proposed approach in comparison to ten state-of-the-art (SOTA) methods, achieving the best performance with an average overall accuracy (OA) of 95.88% and an average Kappa coefficient (KC) of 74.20%.

Change detection in synthetic aperture radar (SAR) images is a vital application in remote sensing image processing. Existing unsupervised SAR change detection methods often rely on pre-classification to generate pseudo-labels for classifying the image regions into three classes: nochanged, changed, and uncertain. However, these methods do not fully exploit the pseudo-labels by focusing only on changed and nochanged regions. In this letter, we propose a wavelet transform-based multi-scale self-supervised network (WS2Net), which maximizes the utilization of pseudo-labels and incorporates discriminative feature learning. First, we employ clustering as pre-classification to obtain the aforementioned pseudo-labels. Second, we propose a self-supervised triple loss inspired by contrastive and representation learning. This loss comprises the nochanged and changed losses in the feature domain along with the uncertain loss in the source domain. Furthermore, to extract valuable information from SAR images and to improve the noise robustness of the network, we design a wavelet transform-based multi-scale feature extraction module. Finally, a difference image is generated by comparing the features output from the network, which can be further analyzed through segmentation to obtain the final change map. Comparative experiments are conducted with five state-of-the-art methods on three public SAR data sets, showing that the proposed WS2Net achieves the best performance with an average percent correct classification of 97.89% and an average kappa coefficient of 90.24%.