Mengxi Liu

With the continuous development of remote sensing sensors and image processing technology, methods typified by deep learning are leading remote sensing change detection (RSCD) toward an increasingly refined direction. Due to the significant spatiotemporal heterogeneity in high-resolution remote sensing images (RSIs), current RSCD methods have attempted to enhance the robustness of models in dealing with easily confused backgrounds through data augmentation or attention mechanisms. However, these methods are often limited to modeling the differences in dual-temporal images and fail to address complex temporal perturbations, such as changes in imaging conditions and seasonal changes. To tackle this challenge, this article introduces a contrastive learning framework for change detection based on pseudotemporal images, utilizing the visual state space model (named MambaCL). Specifically, MambaCL proposes a three-branch pseudotemporal input model designed to explicitly model complex temporal change features. A FusionMamba encoder based on a visual state space model is constructed to efficiently extract and fuse long-term contextual features. Besides, a global dense joint contrast learning strategy is proposed to bolster the model’s robustness against temporal variations. Comparative experiments and ablation studies were conducted in this article on the BCDD, CLCD, and LEVIR-CD datasets. The quantitative results indicated that MambaCL performs effectively on the RSCD task, achieving state-of-the-art results on the above three benchmarks. Specifically, it obtained an F1-score of 89.84% on the BCDD dataset, an F1-score of 81.29% on the CLCD dataset, and an F1-score of 91.42% on the LEVIR-CD dataset, respectively.

Super-resolution image reconstruction (SRR) is an important task in remote sensing image (RSI) processing, which can serve the application fields, such as natural resources exploration and agricultural monitoring. Recently deep learning based methods achieved great success in RSI-SRR, making RSI-SRR more efficient and accurate. However, homogeneous low-resolution (LR) and high-resolution RSI pairs from the same sensor are often difficult to obtain, and the use of simulated image pairs makes it difficult to apply these methods in real scenarios. Some RSI-SRR methods attempt to reconstruct image resolution from heterogeneous RSIs. But these methods, on the one hand, are difficult to deal with large-scale RSI-SRR; and on the other hand, they do not take into account the domain differences of heterogeneous RSIs. To address the above challenges, this article proposes a scale-decoupling super-resolution network with domain transfer for heterogeneous RSIs (SDDT-SR). The SDDT-SR breaks down the large-scale SSR (x8) into two cascaded smaller-scale steps (x2, x4), progressively restoring spatial details of the images. Specifically, in SDDT-SR, an SR module with domain transfer (DTSR) is designed to transfer the LR image from a specific domain to another one, and an SR module with unpaired learning (ULSR) is designed to further up-sample the output of DTSR to the target resolution. Experiments on SGSRD and satellite-UAV heterogeneous image change detection dataset (HTCD) dataset show that SDDT-SR outperforms existing methods in both metrics and visualization, with the fréchet inception distance (FID) of 181.8679 on SGSRD and of 163.3831 on HTCD.

Address matching is the key to improve the efficiency of location-based services (LBS). Most current methods mainly focus on encoding features from address texts and have not fully explored the geospatial information of address. Although there are methods that incorporate longitude and latitude, they neglect the absolute coordinate information and neighborhood distribution characteristics within them. The insufficient encoding of geographic information seriously affects the accuracy of address text matching. Therefore, this paper proposes a Multi-Modal Geospatial Encoding Network (MGENet) for Chinese address matching. Specifically, the MGENet uses Geohash encoding to model the spatial information of longitude and latitude, and constructs spatial features to express the neighborhood distribution of address entities. Additionally, a multi-modal semi-interactive layer is incorporated to solve the problem of insufficient interaction between text and spatial features. Experiments on large-scale datasets show that the MGENet can achieve 80.3% in F1 score for Chinese address matching, which is 3.1% higher than that of the state-of-the-art (SOTA) model, Chinese-BERT. Moreover, the integration of semi-interactive layer allows the model to require only 7.48M floating-point operations (FLOPs) during prediction, which greatly promotes the efficiency of downstream services.

Urban renewal has led to the proliferation of informal urban habitats, such as slums, shanty towns, and urban villages (UVs). As an important component of urban renewal, UVs influence urban spatial structure and land use patterns. Therefore, the fine extraction of UV is of great theoretical and practical significance. Existing UV classification techniques mostly employ machine learning and convolutional neural network based models, which struggle to perceive long-range global semantic information. In this article, based on high-resolution remote sensing images, we propose a multiscale mask transformer model for UV (MaskUV). It can extract both local texture features and global features. The multiscale mask transformer module with mask attention can aggregate different levels of pixel and object features, enhancing the model's recognition and generalization abilities. We extracted UV in seven cities in the Pearl River Delta (PRD) using MaskUV and analyzed the spatial pattern and accessibility of UV. Due to the scarcity of fine-grained UV detection datasets, we also provide a novel dataset (UVSet) containing 3415 pairs of 512 × 512 high-resolution UV images and labels, with a spatial resolution of 1 m. Comparative experiments with several UV extraction models demonstrate the effectiveness of MaskUV, achieving an F 1 score of 84.39% and an IoU of 73.00% on UVSet. Besides, MaskUV achieves highly accurate detection results in seven cities in the PRD, with average F 1 and IoU values of 84.41% and 72.44%, respectively.

As high temperature and heat wave have become great threats to human survival, social stability, and ecological safety, it is of great significance to master the spatial and temporal dynamic changes of temperature to prevent high temperature and heat wave risks. The meteorological station can provide accurate near ground temperature, but only within a specific space and time. In order to meet the needs of large-scale research, spatial interpolation methods were widely used to obtain spatially continuous temperature maps. However, these methods often ignore the influence of external factors on temperature, such as land cover, height, etc., and neglect to supplement temporal-wise information. To deal with these issues, a joint spatio-temporal method is proposed to obtain dense temperature mapping from multi-source remote sensing data, which combining a geographically weighted regression (GWR) model and a polynomial fitting model. Besides, a heat wave risk model is also built based on the dense temperature maps and population data, in order to evaluate the heat wave risk of different areas. Accuracy evaluations and experiments have verified the effectiveness of the proposed methods. Case study on the four cities of Zhejiang Province, China have demonstrated that areas with higher degree of urbanization are often accompanied by higher heat wave risks, such as the northern part of the study area. The study also found that the heat wave risks have presented a centralized distribution and spatial autocorrelation characteristics in the study area.

Urban green space (UGS) is an important component in the urban ecosystem and has great significance to the urban ecological environment. Although the development of remote sensing platforms and deep learning technologies have provided opportunities for UGS mapping from high-resolution images (HRIs), challenges still exist in its large-scale and fine-grained application due to insufficient annotated datasets and specially designed methods for UGS. Moreover, the domain shift between images from different regions is also a problem that must be solved. To address these issues, a general deep learning (DL) framework is proposed for UGS mapping in the large scale, and fine-grained UGS maps of 31 major cities in mainland China are generated (UGS-1m). The DL framework consists of a generator and a discriminator. The generator is a fully convolutional network designed for UGS extraction (UGSNet), which integrates attention mechanisms to improve the discrimination to UGS, and employs a point-rending strategy for edge recovery. The discriminator is a fully connected network aiming to deal with the domain shift between images. To support the model training, an urban green space dataset (UGSet) with a total number of 4544 samples of 512×512 in size is provided. The main steps to obtain UGS-1m can be summarized as follows: (a) first, the UGSNet will be pre-trained on the UGSet in order to obtain a good starting training point for the generator. (b) After pre-training on the UGSet, the discriminator is responsible for adapting the pre-trained UGSNet to different cities through adversarial training. (c) Finally, the UGS results of 31 major cities in China (UGS-1m) are obtained using 2179 Google Earth images with a data frame of 7^′30^′′ in longitude and 5^′00^′′ in latitude and a spatial resolution of nearly 1.1 m. An evaluation of the performance of the proposed framework by samples from five different cities shows the validity of the UGS-1m products, with an average overall accuracy (OA) of 87.56 % and an F1 score of 74.86 %. Comparative experiments on UGSet with the existing state-of-the-art (SOTA) DL models proves the effectiveness of UGSNet as the generator, with the highest F1 score of 77.30 %. Furthermore, an ablation study on the discriminator fully reveals the necessity and effectiveness of introducing the discriminator into adversarial learning for domain adaptation. Finally, a comparison with existing products further shows the feasibility of the UGS-1m and the great potential of the proposed DL framework. The UGS-1m can be downloaded from https://doi.org/10.57760/sciencedb.07049 (Shi et al., 2023).

Derris trifoliate , one of the most notorious invaders of Mangroves in South China, seriously threatens the growth of Mangroves and the stability of local ecosystem. In order to effectively control the spread of Derris trifoliate and stabilize the Mangrove ecosystem, it is necessary to identify the distribution and simulate the occurrence of Derris trifoliate in Mangrove forests. While previous methods for invasive plant mapping based on satellite images were limited by low temporal and spatial resolution, the newly born Unmanned Aerial Vehicles (UAV) images have provided a fast, fine-grained, and low-cost alternative for rapid invasive plant monitoring. However, extracting Derris trifoliate in mangrove forests based on UAV images is still a challenge. For one thing, it is difficult to collect enough data for model training, since the Derris trifoliate is hard to be distinguished in Mangroves. For another thing, few existing methods can meet the requirements of high computing efficiency and low memory consumption of UAVs. Therefore, we proposed two lightweight deep learning networks based on DenseNet and VGG, namely, the lightweight DenseNet (LDN) and the lightweight VGG (LVG), and investigated the capability of LDN and LVG in identifying Derris trifoliate from Mangrove forests with small amounts of data. This study have verified the effect of the lightweight deep learning algorithms for the accurate detection of Mangrove invasive plants Derris trifoliate from UAV images.