Junyan Lu

The scarcity of annotated datasets remains a primary obstacle for remote sensing change detection (RSCD). Most current methods address this by designing entirely new model architectures from scratch, a process that is labor-intensive and often yields only marginal performance gains. While supervised contrastive learning can enhance feature discrimination with limited labels, it suffers from two critical drawbacks in the RSCD domain: first, the computation of fine-grained pixel-wise contrastive loss on lengthy pixel sequences imposes a heavy memory burden. Second, the inherent coupling of gradient contributions from positive and negative sample pairs in changed regions weakens the discrimination of true changes. To overcome these limitations, we propose a balance factor-driven change enhancement block (CEBlock). Our solution incorporates a two-stage feature compression (TSFC) mechanism to efficiently compress lengthy bitemporal sequences, reducing memory requirements and enabling more precise identification of changes. Additionally, we introduce a piecewise loss function that uses a balance factor to replace the correlation calculation of positive and negative samples in real change regions, to separate the gradient contributions of positive and negative sample pairs in real change regions, thereby enhancing the model’s attention to change features. A key advantage of the CEBlock is its plug-and-play design; it can be directly integrated into the Siamese feature extraction networks of existing models [including convolutional neural networks (CNNs), Transformers, and Mamba frameworks] to optimize feature alignment with minimal effort. Experiments show that methods augmented with our CEBlock achieve comprehensive performance enhancements on the WHU-CD, DSIFN-CD, and SYSU-CD datasets and establish new state-of-the-art (SOTA) results on two public benchmarks, with $F1$ /intersection over union (IoU) improvements of 0.51%/0.77% and 3.30%/3.68% on SYSU-CD and DSIFN-CD, respectively.

Horizontal annotation-based weakly-supervised rotating object detection is a research field that has just been explored. This concept is expected to have a transformative effect on the advancement of data-driven rotating object detection methods. Existing pioneering researches directly regress the rotating rectangle based on angle description, which mainly face two limitations under the weakly-supervised framework: i) The regression form of weakly-supervised learning is redundant relative to the training objective, thereby increasing the difficulty of model training; ii) The training objective of self-supervised learning do not have a close logical relationship with the test metrics, which may result in loss of accuracy. Addressing above issues, this paper proposes an angle-free weakly-supervised rotating object detection framework, whose salient points mainly include: i) By improving an angle-free rotating object representation, the decoupling between horizontal and rotating regression parameters in describing rotating objects is achieved; ii) A weakly-supervised learning pipeline that is completely equivalent to the common horizontal object detection is designed to effectively relieve the difficulty of model training; iii) A geometrically intuitive self-supervised learning loss function is introduced to bridge the gap between the training objective and testing metrics. Experimental results on multiple large-scale remote sensing datasets confirm that the accuracy of this method is superior to the state-of-the-art (SOTA) weakly-supervised rotating object detection methods, and is competitive even with SOTA fully-supervised related works.

Forest defoliating pests are significant global forest disturbance agents, posing substantial threats to forest ecosystems. However, previous studies have lacked systematic analyses of the continuous spatiotemporal distribution characteristics over a complete 3–5 year disaster cycle based on remote sensing data. This study focuses on the Dendrolimus superans outbreak in the Changbai Mountain region of northeastern China. Utilizing leaf area index (LAI) data derived from Sentinel-2A satellite images, we analyze the extent and dynamic changes of forest defoliation. We comprehensively examine the spatiotemporal patterns of forest defoliating pest disasters and their development trends across different forest types. Using the geographical detector method, we quantify the main influencing factors and their interactions, revealing the differential impacts of various factors during different growth stages of the pests. The results show that in the early stage of the Dendrolimus superans outbreak, the affected area is extensive but with mild severity, with newly affected areas being 23 times larger than during non-outbreak periods. In the pre-hibernation stage, the affected areas are smaller but more severe, with a cumulative area reaching up to 8213 hectares. The spatial diffusion characteristics of the outbreak follow a sequential pattern across forest types: Larix olgensis, Pinus sylvestris var. mongolica, Picea koraiensis, and Pinus koraiensis. The most significant influencing factor during the pest development phase was the relative humidity of the year preceding the outbreak, with a q-value of 0.27. During the mitigation phase, summer precipitation was the most influential factor, with a q-value of 0.12. The combined effect of humidity and the low temperatures of 2020 had the most significant impact on both the development and mitigation stages of the outbreak. This study’s methodology achieves a high-precision quantitative inversion of long-term disaster spatial characteristics, providing new perspectives and tools for real-time monitoring and differentiated control of forest pest infestations.

Digital surface model (DSM) presents height information of the Earth’s surface and plays an important role in many remote sensing (RS) applications. Since the conventional acquisition of DSM is laborious and expensive, DSM reconstruction from monocular RS images has attracted extensive research in recent years, which is an ill-posed problem and thus rather challenging. Related works have achieved great accomplishments in this regard; however, they still face some limitations in training robustness, accuracy, and efficiency. To address the issues, a semantic joint monocular RS image DSM regression framework is proposed in this article, whose salient points include that: 1) semantic segmentation is integrated into the DSM regression task so that a shared backbone can extract complementary features from each objective to improve the performance of the individual task. Meanwhile, based on the consistency of the two training objectives, a two-stage joint loss function is introduced to improve the convergence and robustness of model training; 2) an encoding–decoding backbone is designed based on feature multiplexing, which simultaneously achieves multiscale feature fusion and information decoupling, thereby greatly reducing model parameters and improving efficiency while ensuring feature extraction effect; and 3) an iterative upsampling approach is introduced to transform the full-scale spatial features into large receptive-field and locally discriminative dynamic kernels, which are used to inpaint coarse-grained features while decoding, thus enhancing regression accuracy. Finally, experiments demonstrate the effectiveness of the proposal. It is easy to train and achieves superior or comparable accuracy compared with state-of-the-art related works while improving the efficiency by a large margin.

The quantity and quality of cropland are the key to ensuring the sustainable development of national agriculture. Remote sensing technology can accurately and timely detect the surface information, and objectively reflect the state and changes of the ground objects. Using high-resolution remote sensing images to accurately extract cropland is the basic task of precision agriculture. The traditional model of cropland semantic segmentation based on the deep learning network is to down-sample high-resolution feature maps to low resolution, and then restore from low-resolution feature maps to high-resolution ideas; that is, obtain low-resolution feature maps through a network, and then recover to high resolution by up-sampling or deconvolution. This will bring about the loss of features, and the segmented image will be more fragmented, without very clear and smooth boundaries. A new methodology for the effective and accurate semantic segmentation cropland of high spatial resolution remote sensing images is presented in this paper. First, a multi-temporal sub-meter cropland sample dataset is automatically constructed based on the prior result data. Then, a fully convolutional neural network combined with contextual feature representation (HRNet-CFR) is improved to complete the extraction of cropland. Finally, the initial semantic segmentation results are optimized by the morphological post-processing approach, and the broken spots are ablated to obtain the internal homogeneous cropland. The proposed method has been validated on the Jilin-1 data and Gaofen Image Dataset (GID) public datasets, and the experimental results demonstrate that it outperforms the state-of-the-art method in cropland extraction accuracy. We selected the comparison of Deeplabv3+ and UPerNet methods in GID. The overall accuracy of our approach is 92.03%, which is 3.4% higher than Deeplabv3+ and 5.12% higher than UperNet.

Point feature-based (PFB) remote sensing image registration methods have achieved great accomplishments. However, such methods based on the hand-crafted point features (HCPFs) have encountered limitations of robustness, accuracy, efficiency, and automation for high-resolution remote sensing (HRRS) image registration. A two-stage registration framework specialized for HRRS images, which combines object and point features is introduced in this article. Specifically, in the first stage, we propose to abstract and represent the macroscopic semantic objects in HRRS images as the deep convolutional object features (DCOFs), which is used to implement the global coarse registration. In the second stage, specific local areas are selected according to the DCOF matching, and the PFB method is performed in the local areas to achieve refinement. The salient points of this proposal include that, 1) the proposed DCOF is more discriminative and expressive than the HCPF at the macroscopic scale so as to effectively achieve automatic and robust global preregistration. 2) We implement the object feature extraction in an object detection fashion which is more-or-less plug and play. 3) Our designed two-stage strategy plays to the advantages of the DCOF and HCPF in different scenarios, thereby improving accuracy and efficiency. Finally, experiments demonstrate the effectiveness of our method. Compared with the state-of-the-art PFB methods, it improves the accuracy by about 3$\%$–6$\%$ and the efficiency by about 15$\%$–30$\%$.

Feature-based remote sensing image registration methods have achieved great accomplishments. However, they have faced some limitations of applicability, automation, accuracy, efficiency, and robustness for large high-resolution remote sensing image registration. To address the above issues, we propose a novel instance segmentation based registration framework specifically for large-sized high-resolution remote sensing images. First, we design an instance segmentation model based on a convolutional neural network (CNN), which can efficiently extract fine-grained instances as the deep features for local area matching. Then, a feature-based method combined with the instance segmentation results is adopted to acquire more accurate local feature matching. Finally, multi-constraints based on the instance segmentation results are introduced to work on the outlier removal. In the experiments of high-resolution remote sensing image registration, the proposal effectively copes with the circumstance of the sensed image with poor positioning accuracy. In addition, the method achieves superior accuracy and competitive robustness compared with state-of-the-art feature-based methods, while being rather efficient.

As object detection is widely adopted in aerial images, scene texts and other fields, rotating object detection plays an important role and draws attention since it can provide highly accurate orientation and scale information. In this article, we propose a novel and simple baseline to effectively conduct rotating object detection. First, we design a brand-new representation for rotating objects by using a circle cut horizontal rectangle (CCH). The CCH ensures that the regression parameters will not exceed the defined domain and avoids vertex sorting, thus solving some problems in current common representations, including the boundary problem and order problem, and improving the robustness. Second, we design a lightweight head based on the CCH to add the rotating regression to classic benchmark in an almost cost-free manner and propose a single-stage anchor-free rotating (SAR) object detection convolutional neural network. Finally, we demonstrate the details of our method by applying it to data sets with different scenarios. The experiments confirm that our method achieves competitive accuracy and state-of-the-art speed in aerial image and scene text detection.

Many studies have shown that, owing to the constraint of working memory capacity, language users prefer shorter dependency distances. However, these studies, which are all based on dependency distances in sentences or texts, leave the question unanswered: does noun phrase structure also demonstrate the tendency of dependency distance minimization? To answer this question, the article, based on a self-annotated English corpus, was carried out to probe the dependency distance in noun phrases between heads and their pre- and post- modifiers. The results show that the dependency distance between pre-modifiers and heads is much shorter than that between post-modifiers and heads. This finding suggests that noun phrases also show the tendency of dependency distance minimization, although they present a much longer mean dependency distance than sentences.