Chlorophyll-a (Chl-a) concentration estimation by remote sensing is an important means for monitoring offshore water quality and eutrophication. In-situ hyperspectral data can achieve accurate analyses of Chl-a, but it is not suitable for regional inversion. Satellite remote sensing provides the possibility for regional inversion, but the precision is lower limited to atmospheric correction result. Therefore, this work uses machine learning to fuse in-situ hyperspectral data and Sentinel-2 multispectral instrument images to combine their complementary advantages, so as to improve the precision of regional Chl-a concentration inversion. First, the in-situ spectra were resampled based on the satellite spectral response function to obtain equivalent reflectance. Second, the spectral feature bands of Chl-a were determined by correlation analysis. Then three machine learning models, support vector regression, random forest, and back propagation neural network, were used to establish mapping relationships of feature bands between equivalent reflectance and satellite image reflectance so as to correct the satellite feature bands. Finally, Chl-a inversion models were constructed based on the satellite feature bands before and after correction. The results demonstrate that the corrected inversion model shows an increase in R2 by 0.25 and a decrease in mean relative error by 7.6%. This fusion method effectively improves the accuracy of large-scale Chl-a concentration estimation.

Understanding the impact of climate change on Earth presents a significant scientific challenge. Monitoring changes in terrestrial ecosystems, including leaf water content, is essential for assessing plant transpiration, water use efficiency, and physiological processes. Optical remote sensing, utilizing multi-angular reflectance measurements in the near infrared and shortwave infrared wavelengths, offers a precise method for estimating leaf water content. We propose and evaluate a new index based on multi-angular reflection, using 256 leaf samples from 10 plant species for calibration and 683 samples for validation. Hyperspectral indices derived from multi-angular spectra were assessed, facilitating efficient leaf water content analysis with minimal time and specific bands required. We investigate the relationship of leaf water content using spectral indices and apply linear and nonlinear regression models to calibration data, resulting in two indices for each indicator. The newly proposed indices, (R1−R2)/(R1−R3) for linear and (R1905−R1840)/(R1905−R1875) for nonlinear, demonstrate high coefficients of determination for leaf water content (>0.94) using multi-angular reflectance measurements. Published spectral indices exhibit weak relationships with our calibration dataset. The proposed leaf water content indices perform well, with an overall root mean square error of 0.0024 (g/cm2) and 0.0026 (g/cm2) for linear and nonlinear indices, respectively, validated by Leaf Optical Properties Experiment, ANGERS, and multi-angular datasets. The (R1−R2)/(R1−R3) bands show promise for leaf water content estimation. Future studies should encompass more plant species and field data.

Classification of complex and large hyperspectral images (HSIs) with machine learning (ML) algorithms is an important research area. Recently, explainable artificial intelligence (XAI), which helps to explain and interpret black-box ML algorithms, has become popular. Our study aims to present extensive research on the use of XAI methods in explaining the band effect in HSI classification and the impact of reducing the high band number of HSIs by feature selection on the performance of the classifiers. The importance levels of the spectral bands that are effective in the decisions of the different ML classifiers were examined with the deep reinforcement learning and XAI methods, such as Shapley additive explanations (SHAP) and permutation feature importance (PFI). Our work selects representative bands using SHAP and PFI as XAI analysis techniques. We evaluated the XAI-based band selection performance on three publicly available HSI datasets using random forest, light gradient-boosting machine, and extreme gradient boosting classifier algorithms. The results obtained by applying XAI and deep learning methods were used to select spectral bands. Additionally, principal component analysis, a common dimension reduction technique, was performed on the dataset used in our study. Comparable performance evaluation shows that XAI-based methods choose informative bands and outperform other methods in the subsequent tasks. Thus the global and class-based effects of spectral bands can be explained, and the performance of classifiers can be improved by eliminating features that have a negative impact on classification. In HSI classification, studies examining the decisions of ML classifiers using XAI techniques are limited. Our study is one of the pioneer studies in the usage of XAI in HSI classification.

Hyperspectral image super-resolution (HI-SR) can improve hyperspectral images’ spatial resolution to capture more spatial details from the observed scenario. Recently, fusing multispectral and hyperspectral images to implement HI-SR has become a hot topic in remote sensing. Moreover, the development of deep learning has further promoted the advancement of HI-SR over the past few years, bringing many specific HI-SR networks. However, for most HI-SR networks, it is challenging to obtain global features from different images, which limits the reliability of results. We propose a new HI-SR method that adopts a two-stream self-attention network (TSSA-Net) to address the above issues. The proposed deep network consists of two coupled encoders for acquiring the abundance and endmembers from multispectral and hyperspectral images, respectively. Each encoder has a self-attention mechanism to acquire global features and use a convolutional layer to aggregate local features. Meanwhile, cross-stream self-attention is designed for information exchange among different streams, enhancing the robustness of TSSA-Net. Several experiments are conducted to verify the effectiveness and competitiveness of TSSA-Net.

Hyperspectral remote sensing combined with machine learning represents an important means of soil chemical element monitoring. However, there is a problem regarding the lack of universality in element content prediction models across different soil types. We propose an improved standard-sample calibration transfer method [the direct standardization (DS), piecewise direct standardization (PDS), and spectral space transformation (SST) calibration transfer methods combined with the standard sample difference pairing (SSDP) method], which is derived from the standard-sample calibration transfer method commonly used in spectrometer calibration, to study the transferability of machine learning prediction models between different soil types. To study the effectiveness of the improved standard-sample calibration transfer method, albic black soil and typical black soil from northeast China are selected as the research objects. We established partial least squares (PLS) and backpropagation neural network (BPNN) prediction models for the soil organic matter (SOM) content of albic black soil. Subsequently, the albic black soil model is used to predict the SOM content of typical black soil. The results show that the improved standard-sample calibration transfer method significantly improves the prediction ability of the albic black soil model for typical black soil, with the determination coefficient (R2), root mean square error (RMSE), and relative percent deviation (RPD) of the typical black soil validation set increasing to 0.616 to 0.774, 2.184 to 2.568, and 1.500 to 1.763, respectively. Among the methods utilized, SSDP-PDS, SSDP-DS, and SSDP-SST exhibit the best, second-best, and worst performances, respectively. In addition, the methodological framework is also effectively applicable to the prediction of SOM content in albic black soil using a typical black soil model. Finally, compared with the sample mixing method, the improved standard-sample calibration transfer method demonstrates higher prediction accuracy. This study holds significance for research on the universality of element content prediction models across different soil types.

The extensive development of coal resources has resulted in the formation of tensile fissures, disrupting soil structure and impacting soil quality. Monitoring soil nutrients in such areas is crucial for agricultural productivity. This study focused on the tension fissure area near the 3522 working face of Huaibei coal mine, Anhui province, China. We analyzed the variability characteristics of soil organic matter (SOM) in the fissure zone and assessed the potential of using soil spectra to estimate SOM. The fractional order differentiation (FOD) method was employed to preprocess the spectral data. Traditional correlation coefficient (CC) and coupled algorithms correlation coefficient—successive projection algorithm (CC-SPA) and correlation coefficient—competitive adaptive reweighted sampling (CC-CARS) were used for band extraction. Finally, partial least squares regression (PLSR) and support vector machine regression (SVR) were applied to establish the SOM estimation model. The results show that the development of fissures decreases the SOM content, and using spectroscopy to estimate SOM content in the fissure zone has greater potential. FOD can capture subtle spectral features. As the fractional order increases gradually from 0 to 1, spectral features are enhanced, but further increasing the fractional order introduces minor noise peaks, limiting the accuracy of SOM estimation. Feature selection algorithms effectively extract sensitive spectra related to SOM. The coupling algorithm CC-CARS obtains good estimation results with fewer band variables extracted. After 0.8 order differentiation, the SVR model constructed using feature wavelengths extracted by the CC-CARS algorithm performs best, with an R2 of 0.87 for the test set. The study can provide some technical references for the rapid and accurate estimation of SOM in coal mining fissure zones.

The spatiotemporal distribution information of crops is essential. However, remote sensing mapping still faces challenges, such as fragmented land plots, cloud cover, and changes in crop morphology across the phenological stages. To address these issues, we constructed a time series multi-feature dataset using the flexible spatio-temporal data fusion method, which fused Sentinel-2 and MODIS data, with the addition of Sentinel-1 SAR data. We employed the time-weighted dynamic time warping (TWDTW) algorithm and compared its performance with the random forest (RF) algorithm. By comparing the crop classification results of different lengths of time-series multi-feature datasets, we investigate the earliest identifiable time for crops. The results demonstrate that the incorporation of multi-source time-series remote sensing data in the construction of crop phenological features effectively improves crop classification accuracy. The TWDTW algorithm attained an overall accuracy of 92.28% and a Kappa coefficient of 0.895, outperforming the RF method, which had an accuracy of 90.44% and a Kappa coefficient of 0.869. Early-stage time-series data was enough to identify wheat, whereas complete time-series data for the entire growing season was required for oilseed rape and fallow fields. In the early stages of identification, using spatial-temporal fusion can improve crop recognition accuracy. This study found that the spatio-temporal fusion of multi-source remote sensing data and TWDTW offers the potential for accurate crop classification, especially in areas with complex fragmented cropping systems.

We investigate the potential of Canavalia ensiformis green manure in enhancing soil fertility, rice growth, and environmental sustainability in Thai rice cultivation. The research integrates field measurements with high-resolution PlanetScope satellite data to monitor the environmental changes induced by green manure application. Fourteen vegetation indices (VIs) were derived from satellite data to assess the impact of green manure on rice growth and development. The results demonstrate short-term improvements in soil organic matter, nitrogen, phosphorus, and potassium levels in the green manure-amended plots, with strong correlations observed between VIs and field-measured parameters. The visible atmospherically resistant index showed the strongest correlations with soil moisture across various depths (r=0.621 to 0.935, p<0.01), and the yellow normalized difference vegetation index demonstrated the highest correlation with available phosphorus (r=0.812, p<0.05). For rice growth parameters, the carotenoid reflectance indices (CRI Green550) exhibited the strongest correlation with plant height (r=0.878), and the chlorophyll content estimation indices (Chl Green550) with chlorophyll content (r=0.642). The normalized difference vegetation index showed the highest correlation with rice yield (r=0.885). The study highlights the effectiveness of these indices in monitoring soil and plant parameters in response to green manure application. However, the benefits of green manure diminished over time, emphasizing the need for regular application and complementary soil conservation practices. The findings contribute to developing sustainable rice cultivation practices and underscore the potential of integrating green manure with remote sensing technologies to enhance soil health, crop productivity, and environmental sustainability in Thai rice farming systems.

The coffee agro-ecosystems are increasingly being transformed into small-scale coffee-growing agricultural systems. In this context, the challenge of accurately classifying coffee cropping systems (CSs) becomes more significant, particularly in regions such as Uganda where dense vegetation and diverse topography complicate traditional land surveys. We harness the capabilities of remote sensing to provide hyperspectral data crucial for distinguishing between various coffee CSs and other land covers. Specifically, we focus on the spectral analysis of three types of Robusta coffee CSs—those integrating agroforestry, those combined with banana cultivation, and those in full sun exposure. Using in situ hyperspectral measurements captured by the FieldSpec 2™ spectroradiometer across the 325 to 1075 nm range of the electromagnetic spectrum, we aimed to (1) analyze the unique spectral properties and behaviors of these Robusta coffee CSs and (2) effectively discriminate among them using advanced hyperspectral datasets alongside the machine learning (ML) classification algorithms. The key to this process was the use of narrow spectral bands (NSBs) and various narrow-band vegetation indices (VIs), serving as predictor variables. A selection of critical variables (NSB = 9 and VIs = 8) was identified through the guided regularized random forest (RF) technique and then applied to four ML algorithms—RF, stochastic gradient boosting (GB), linear discriminant analysis, and support vector machine for classification experiments. The findings indicated high discrimination accuracy, with the RF and GB algorithms achieving overall accuracies of 93% and 90.5%, respectively, when using the selected VIs, and 87.3% (RF) and 83% (GB) when applying the chosen NBSs. These results underline the efficacy of integrating hyperspectral datasets and ML algorithms in reliably categorizing Robusta coffee CSs, a crucial step toward enhancing sustainable coffee cultivation practices.

Simultaneous achievement of the detection of small targets in synthetic aperture radar (SAR) images and the visual interpretation of the entire scene remains challenging due to the SAR imaging properties. To address these challenges, a two-branch network, which includes a proposed target detection network of TYSAR-YOLOv5 and a SAR image colorization network of CycleGAN, is developed. TYSAR-YOLOv5 is established on the basis of the YOLOv5 network, which incorporates an additional detection head specifically designed for capturing small targets from shallow features. The BoTNet structure is also deployed in the backbone network to capture global information and locate targets in highly dense scenes accurately with an acceptable number of parameters. To enhance the interpretation of the surrounding environment of detected targets, the CycleGAN network known for style transfer capabilities is employed to colorize SAR images and fuse with detected targets to achieve the enhanced target detection result of the SAR image. We expanded and updated the large-scale multi-class SAR image target detection dataset-1.0 (MSAR-1.0) by adding small aircraft targets using data augmentation techniques. The reference color images corresponding to the SAR images of each target type are also added using Google Earth Engine and public vision data for SAR image colorization. Experimental results demonstrate that the developed network dominantly outperformed state-of-the-art target detection networks in small aircraft target detection and multi-class target detection tasks. Of more importance, the final visualizations of target detection exhibit excellent interpretability, which provides rich semantic information benefiting decision-making.

Mapping native grasses and alien species on river dikes is crucial for vegetation management, which influences native species conservation and river dike maintenance. We aimed to improve alien species distribution mapping of Solidago altissima and Sorghum halepense on river dikes along Tone River, by fusing satellite and drone images. The methodology includes fusing pan-sharpened WorldView-3 satellites with drone images for one sample dike. The object-based approach segmented images into objects and extracted statistical data to form datasets. Regression models were constructed from the sample dike to predict fused satellite–drone datasets for other dikes. The random forest classification model was then applied to map the distribution of alien species. Our findings highlight the enhanced mapping accuracy achieved using a fused satellite–drone dataset approach. For Solidago altissima, the fused dataset reaches the highest overall accuracy (OA) of 98.39% and a kappa coefficient of 0.976. Likewise, for Sorghum halepense, the fused dataset achieved the highest OA of 97.78% and a kappa coefficient of 0.936, compared with that of pan-sharpened and original satellite datasets. These outcomes underscore the contribution of the fused satellite–drone imagery for alien species distribution mapping in the river dike environment.

Agricultural greenhouses (AGHs) are one of the main technological methods for regulating crop growth environments globally. They play a crucial role in food production, resource conservation, and rural economies, yet they also present environmental security threats. The current research on AGH extraction primarily focuses on small scales. However, the problem of confusing greenhouses with other features has not been completely resolved, and accurately obtaining the distribution of AGHs on large spatial scales is a challenge. We extract multi-temporal spectral feature data from Sentinel-2 images using Google Earth Engine (GEE) and build a hierarchical extraction decision tree model to extract AGHs. The model achieves an overall accuracy of 95.65%, with a Kappa coefficient of 0.89. Utilizing this model, a spatial distribution map of AGHs in Jilin Province for 2021 was generated, offering an initial insight into the overall distribution of AGHs. This method also demonstrates the significant potential and extensive application prospects of utilizing Sentinel-2 data and the GEE platform for identifying AGHs. It offers a scientific foundation for the sustainable development of agricultural regions and holds significant importance in advancing the development of facility agriculture by using Sentinel-2 data embedded in GEE at a regional or global scale.

Accurate monitoring of aboveground biomass (AGB) is crucial in preventing grassland degradation and achieving carbon neutrality. Remote sensing data and machine learning-based methods have been widely used to estimate the grassland AGB from national to regional scales due to their unique advantages of low cost and high efficiency. However, in the context of significant spatial heterogeneity, the estimation process for AGB in this category usually has inherent uncertainty. Existing statistical validation methods are unable to characterize the spatial distribution of uncertainty and generally lack consideration of potential uncertainty in grassland AGB. To address this issue, we developed a framework to map the spatial distribution of uncertainty based on the quantile regression forest model. Furthermore, the framework explored the driving factors of uncertainty using the geographical detector model. The research results show that the quantile regression forests model in the framework well-estimated the grassland AGB and characterized its spatial distribution. Also, the spatial pattern of uncertainty was closely related to the AGB and affected the amount of sampling points. Among the multiple factors, the soil-adjusted vegetation indices were the primary driving force of the uncertainty. This research presents an approach for mapping uncertainty in grassland AGB estimation and spatializing estimation error, which could be an effective complement to existing AGB estimation methods and thus facilitate the accurate management of grassland resources.

Cropland is crucial for national food security, maintaining agricultural product quality, and ensuring environmental safety. Consequently, there is an urgent need for cropland change detection (CD) using high-resolution remote sensing images to accurately track cropland distribution and changes. However, the irregular shapes of cultivated areas and challenges in feature fusion complicate boundary localization, posing a risk of losing critical change features. To address these issues, we introduce the dynamic sparse hierarchical attention-driven cropland CD network, which combines a vision transformer with dynamic sparse hierarchical attention (DSHA-Former) and holistic complementation fusion (HCF) modules. DSHA-Former effectively detects targets and refines edge features in images of various sizes. Concurrently, HCF preserves key details by integrating core, setting, margin, and panorama data, supplementing global image-level content comprehensively. This significantly improves the definition of changed areas in the process of merging features from different scales, thus enhancing cropland CD. We evaluate our approach on the CL-CD dataset, achieving an F1-score of 79.49%. In addition, the network demonstrates strong generalization capabilities on both the LEVIR-CD and WHU-CD datasets, with F1-scores of 92.42% and 92.18%, respectively. Our method is effective and demonstrates broad applicability in detecting cropland changes.

Spaceborne tomographic synthetic aperture radar (SAR) (TomoSAR) uses stacks of SAR images acquired at slightly different positions over a certain period in a repeat–pass manner, to obtain absolute 3D positions and 4D deformations of a large amount of building scatters. The double-tier permanent scatterer (PS) network construction method without pre-removal of the atmospheric phase screen has been successfully applied in TomoSAR, but it suffers from the problem that the PS network cannot cover the whole study area when applied to a larger range of building complexes. To address this problem, this paper proposes a new method of double-tier PS network construction combined with a region-growing algorithm, which introduces a region-growing algorithm in the middle of the first-tier network and the second-tier network to realize the construction of a PS network that covers the entire study area after several loop iterations. Experiments are carried out using 26 scenes of TerraSAR-X/TanDEM-X 3-m data stack in Nanjing, China, and the experimental results show that the proposed method improves the PS coverage in the first-tier network from 72.5% to 99.6%, and the network expansion efficiency value increased from 0.12 to 21.48. Its effectiveness and robustness are verified by comparison with other methods.

In the field of remote sensing image (RSI) change detection (CD), existing methods often struggle to balance local and global features and adapt to complex scenes. Therefore, we propose a bidirectional-enhanced transformer network to address these issues. In the encoding part, we introduce a bidirectional-enhanced attention operation that encodes information both horizontally and vertically, as well as deep convolution to improve local contextual connections, thereby reducing computational complexity while improving the network’s perception of global and local information. In the feature fusion part, we propose a channel weighting fusion module, which recalibrates channel-wise features to suppress noise and enhance semantic relevance. We tested the proposed method on two publicly available RSI CD datasets, the LEVIR-CD and DSIFN-CD datasets. Experimental results show that our model outperforms several state-of-the-art CD methods, including one based on convolution, three based on attention, and three based on the transformer.

For many applications such as climatology, agriculture, and fire prevention, information regarding soil moisture can prove very relevant. With a large number of spaceborne or airborne sensors, optical remote sensing can be used to assess the surface soil moisture. In this paper, several indices and the MARMIT4SMC method, relying on the multilayer radiative transfer model of soil reflectance (MARMIT) model, were applied to hyperspectral images of a bare soil in the 0.4 to 2.5 μm spectral range, with a high spatial resolution (4.2 cm), and compared them in terms of robustness. Indeed, only a few experiments were performed with images in the field, and the indices were calibrated due to in situ measurements with probes or sample weighing. In our experiment, only a very small area was available for in situ measurements. This motivated the use of laboratory calibration relying on soil samples and a robustness comparison of the involved metrics. MARMIT parameters and a set of indices were computed for each pixel of a couple of images of bare dry and wet soil, respectively. The indices or parameter values were converted to soil moisture content, using a laboratory calibration. Visual and numerical analyses on the moisture content images obtained for each index or method were performed. The soil heterogeneity and the width of the histograms, reaching 20%, show that the step from the laboratory to the field can be relevant only at the scale of a set of pixels: the large discrepancy prevents from a reliable moisture assessment from a single pixel, but a histogram can bring a fairly reliable one. It also leads to the conclusion that MARMIT4SMC and the Normalized Soil Moisture Index, look more robust than the others when going from the laboratory to the field.

Hyperspectral cameras are useful Earth observation sensors with many practical applications in agriculture and forest remote sensing for detecting and classifying the state of objects. Recently, CubeSats satellites were launched into low Earth orbits. These satellites have certain advantages, such as low cost and short-term development, and their practical applications are continuously increasing. The installation of a hyperspectral camera on CubeSats would significantly increase their observation frequency. However, to fit on a CubeSat, observation instruments must be compact and weigh only a few hundred grams. To address this challenge, we developed a 35-g miniaturized hyperspectral camera by combining a linear variable band-pass filter and an image sensor. This innovative camera was installed on two CubeSats and launched into orbit in 2023 and 2024. The on-orbit experiments conducted using these CubeSats were successful, and valuable hyperspectral data were obtained.

Modern agricultural practices require accurate prediction of crop yields, particularly in the face of a changing climate. Despite the improvement in estimating crop yield over the years through machine learning (ML) algorithms, the increased volatility and complexity of weather patterns continue to make use of conventional ML models unreliable for understanding intricate relationships. We therefore explore the potential of advanced ML, specifically a transformer-based architecture coupled with a temporal convolutional network (TCN) for crop yield prediction. We argue that transformers’ ability to model long-range dependencies within data sequences makes them well suited to handle complex relationships in comprehensive crop datasets. In addition, the integration of a TCN would complement the transformer’s strengths by focusing on temporal feature extraction. Alongside climatic variables, the study integrates soil properties, moderate resolution imaging spectroradiometer (MODIS), and average yield to analyze the complex factors influencing crop growth and yield. The proposed TCN-transformer (TCNT) model is trained and evaluated using corn and soybean yield values selected from 264 counties within Illinois, Iowa, and Wisconsin (the United States corn belt region). Furthermore, experimental results show the superior performance of our TCNT framework over other state-of-the-art models for both in-season and end-of-season yield predictions.

Polarimetric synthetic aperture radar (PolSAR) is an important technology in radar remote sensing. It can capture complex, multidimensional data on the Earth’s surface. This technology is essential for detailed terrain analysis. It enables the differentiation of various land covers by examining scattering patterns. We introduce the multi-dimensional probabilistic voting ensemble network (MD-PVE-Net). It is a semi-supervised framework that combines multiple polarimetric decomposition algorithms with advanced convolutional neural network (CNN) architectures, notably encoder–decoder architectures such as U-Net and residual U-Net. These architectures are designed to maintain spatial relationships and generate dense output maps, addressing the challenges of pixel-wise classification present in contemporary CNN-based models. The incorporation of various decomposition techniques, including Cloude, Huynen, HAAlpha, Freeman-Durden, Vanzyl, and Yamaguchi—significantly enhances the feature extraction process, providing detailed statistical descriptions of scattering mechanisms. MD-PVE-Net leverages these improved features to substantially increase labels’ accuracy for unlabeled data, thus enriching the training dataset. Empirical validation using three distinct PolSAR datasets shows that MD-PVE-Net surpasses current classification methods, especially in complex agricultural environments where accurate crop type discrimination is essential. The ResU-Net architecture, with its deep layers and integration of residual blocks, helps process the intricate, high-dimensional PolSAR data, achieving superior classification results compared with other classification methods. An additional PolSAR dataset is used to validate and confirm the effectiveness of the proposed models in real-world scenarios.

Water body extraction from remote sensing images in complex backgrounds is crucial for environmental monitoring, disaster management, and urban planning. Although existing water body extraction algorithms for remote sensing images offer robust tools for monitoring, they still face challenges, including the inability to extract fine water bodies and the frequent omission of water body edges in complex backgrounds such as dense vegetation, varying terrain, or cloud interference. In response, we introduce an architecture named WatNet to enhance the precision of water body extraction in complex environments. WatNet comprises three main modules: the global multi-attention fusion module (GMAF), the water forward network module (WFN), and the edge focus attention module (EFA). The GMAF module enhances the model’s capability to capture global information through multi-head self-attention and convolutional attention modules, improving the overall feature extraction of water bodies. The WFN module utilizes depth-wise separable convolution and attention mechanisms to enhance the capture of local features in fine water bodies. The EFA module significantly improves the clarity and accuracy of water body boundaries through refined edge detection. Experiments on the LoveDA Remote Sensing Land Cover (LoveDA), Qinghai–Tibet Plateau (QTPL), and Wuhan dense labeling (WHDLD) datasets show that WatNet outperforms the mainstream methods in precision, recall, overall accuracy (OA), F1 score, and mean Intersection over Union (mIoU). On the LoveDA dataset, WatNet’s mIoU improved by 1.24% compared with the second-best method, by 0.4% on the QTPL dataset, and by 0.51% on the WHDLD dataset. These results validate the WatNet’s effectiveness and robustness in water body extraction tasks under various environmental conditions.

Heavy metal stress can lead to morphological and physiological variations in crops. We aimed to distinguish heavy metal stress levels based on the variations of morphological and physiological parameters from radiative transfer and statistical models. Sentinel-2 satellite images and in situ measured data were collected from heavy metal-contaminated soils of rice growing areas in Zhuzhou City, Hunan Province, China. The chlorophyll content (chlorophyll a + chlorophyll b, Cab) and leaf area index (LAI) were calculated using a PROSAIL radiative transfer model and the multilayer perceptron algorithm. A two-dimensional feature space was established from Cab-LAI. Furthermore, a normalized heavy metal stress index (HMSI) from the established Cab-LAI theoretical triangular model was explored to distinguish heavy metal stress levels in rice. The results indicated that (i) the PROSAIL and artificial neural network algorithm were successful at deriving physiological parameters with high estimation accuracy. Pearson’s correlation coefficient between the predicted and measured Cab was 0.85; (ii) the correlation between the measured concentration of cadmium in the soil and the HMSI was 0.84, indicating that it is a good indicator of rice damage caused by heavy metal stress, with the maximum HMSI occurring in rice subjected to high pollution; and (iii) high pollution occurred on both sides of the Xiangjiang River, whereas moderate pollution mainly existed around the heavily polluted areas. Areas with non-pollution and mild pollution were distributed over most of the study area. Combining rice Cab with LAI is a feasible method to determine the distribution of rice heavy metal stress levels over a large area.

Timely and accurate prediction of winter wheat yield is crucial for national food security. In the field of crop yield prediction, deep learning techniques are playing an increasingly important role. However, many existing methods mainly utilize convolutional neural network (CNN) or long short-term memory (LSTM) network, failing to fully exploit the spatiotemporal information in remote sensing data. To address this issue, a CNN–bidirectional long short-term memory (BiLSTM)–attention model for winter wheat yield prediction was proposed using time series Sentinel-1A synthetic aperture radar images. The histogram dimension reduction technique was employed to generate the samples. The CNN was used to extract the spatial–spectral features from the image samples, and the BiLSTM network was adopted to learn the temporal features of winter wheat growth stages from the time series samples. Furthermore, an attention mechanism was introduced to make the networks learn important features more efficiently to improve the accuracy of yield prediction. The time series Sentinel-1A synthetic aperture radar images covering Weishi County, Kaifeng city, Henan province, China, were used for model training and validation. The experimental results demonstrated that the proposed model exhibited good accuracy in yield prediction for the study area, with a coefficient of determination of 0.79, a root mean square error of 583.53 kg/ha, and a mean absolute error of 458.41 kg/ha. The proposed method has a promising application in crop yield prediction and provides a useful reference for similar crop yield prediction.

With the development of drone-based aerial photography technology, unmanned aerial vehicle (UAV) images have become the main tools to create three-dimensional maps and monitor pit slopes for mining areas. To address the problem of poor feature-matching of high-resolution UAV images on pit slopes, we propose a new feature-matching algorithm that makes improvements in three main aspects: (1) A hybrid scale space is established in combination with a spatial domain scale and a frequency domain. The space of the spatial domain scale is generated with anisotropic diffusion filtering, and the space of the frequency domain scale is generated with a phase consistency filter and a weighted least square filter; (2) to distribute feature points uniformly across an image, we first partition images in the stage of feature point extraction and then extract feature points from each partition; (3) a boosted efficient binary local image descriptor is constructed for the detected interest points, thus enhancing the robustness of the descriptor and improving the matching speed. Eight groups were selected from the same UAV sequence images for feature-matching experiments, and the results were compared with classical algorithms such as scale-invariant feature transform, oriented features from accelerated segment test, rotated binary robust independent elementary features, and KAZE. It is found that our improved algorithm appropriately improves the matching speed under the condition of high matching accuracy and better matches the UAV images of pit slopes than the existing image registration methods.

The semantic imbalance of class boundary areas is a key factor in decreasing the classification accuracy of the remote sensing land cover algorithm. We propose a multi-source remote sensing image semantic segmentation network based on multi-modal collaboration and boundary-guided fusion (BGF). The BGF module uses class boundary information as a restriction condition, embeds semantic alignment strategies into the encoder, and enhances the deep semantic features of each mode. On this basis, the boundary guidance strategy is used to assign different weights to the boundary and the internal area of the category to guide the feature fusion. Furthermore, to reduce the impact of multi-modal feature heterogeneity on feature fusion, a cross-modal collaborative fusion module is constructed to associate complementary information between multi-modal features and fully explore the collaborative relationship between multi-modal images from both spatial and channel domains. The comparative experiments were conducted with representative algorithms on the WHU-OPT-SAR data set. The experimental results show that the proposed method has increased the mean intersection over union and overall accuracy indicators by 3.3% and 2.2%, respectively, compared with MCANet, especially that the road category intersection and merger ratio has increased by 10.0% compared with MCANet. We proved the effectiveness of the proposed model.

Although crop yield prediction remains challenging due to various influencing factors, remote sensing data offer a comprehensive view of crop growth dynamics, enabling the capture of spatial and temporal variations in crop development across large areas. Current researchers in this study domain are actively exploring how remotely sensed data and advanced machine learning (ML) methods can be used to improve yield predictions. We contribute to the ongoing exploration but differ in two main ways. First, unlike the current works that usually focus on developing complex deep learning (DL) models for mining the spatial and temporal information inherent in remotely sensed data, our approach rather focuses on simplifying yield prediction through an efficiently weighted ensemble method. The incorporation of effective weighting strategies into ensemble methods is currently lacking in the existing literature. Second, we streamline the integration of remote sensing data into ML methods by extracting the pixel-recorded values from the various remotely sensed images as opposed to conventionally feeding these numerous bands into a DL algorithm. Our experimental results indicate that the proposed approach is capable of outperforming current state-of-the-art yield prediction algorithms with an R2 of 0.872, mean squared error of 120.211 bu/acre, mean absolute error of 8.355, and root mean squared error of 10.637. In addition, our feature importance analysis confirmed that remotely sensed data are able to provide deep information related to crop yield over other climatic variables.

Saliency detection is a conventional computer vision technique used to identify salient regions in difference images for change detection (CD) in multi-temporal/bi-temporal remotely sensed images. However, most existing saliency-based CD methods have primarily focused on analyzing the colors and brightness of different images, often neglecting the rich spectral-spatial and context-aware features of remotely sensed multi-spectral images. Furthermore, these methods independently leverage the visual saliency features and spectral-spatial features of remote sensing images, without seamlessly integrating the two components. In response to this limitation, this paper proposes a novel CD method based on context-aware saliency-spectral-spatial features, introducing a weighted coarse-to-fine fusion strategy. The proposed method fully leverages context-aware saliency features, as well as spectral and spatial features from remotely sensed images. It summarizes the combination of these features into three principles: local low-level feature, global low-level feature, and high-level feature. Then, with the assistance of segmentation objects, the proposed method generates a single-scale salience map through global-local comparisons between an object and its K-level neighborhood objects, employing the three principles of multi-level features. Finally, in accordance with principles 2 and 3, a weighted coarse-to-fine fusion strategy at the pixel level is designed to incorporate multi-scale saliency maps. Fusion weights for pixels, from coarse-to-fine scale, are adaptively obtained by considering the heterogeneity of the object they belong to and the intensity difference between the pixel and the object. To verify the effectiveness of the proposed method, ten comparative experiments were conducted against other state-of-the-art methods The experimental results show that the proposed method better employs context-aware saliency-spectral-spatial features to preserve the details of CD map changes and reduce noise. Furthermore, experiments conducted on three datasets, featuring various region types, region sizes, and spatial resolutions of remotely sensed images, highlight the feasibility and transferability of the proposed method for change detection.

Deep learning-based dehazing of remote sensing images faces two major problems: distortion of remote sensing images and lack of large-scale real paired datasets. To solve these problems, we propose a remote sensing image dehazing method based on data mixing and Laplace network. The Laplace pyramid can divide remote sensing images into different frequency domain layers (the low-frequency layer retains global color information, and the high-frequency layer retains texture details from coarse to fine), and these features are fed into a lightweight multilayer perceptron to learn long-range dependencies. A backbone network consisting of a spatial weighted residual channel attention module can help the residual haze removal module to learn the distribution of haze in remote sensing images for effective haze removal. To address the problem of lack of large-scale real datasets, we cross-mix and restructure the synthetic dataset with the small-sample real dataset, and use the restructured mixed dataset for training, and the trained model can effectively recover the color information of real remote sensing images. After validating the effectiveness and superiority of our model on synthetic datasets, hybrid datasets, and synthetic hyperspectral datasets, we conduct generalizability experiments, and the results show the potential application of our method in advanced vision tasks.

Synthetic aperture radar (SAR) inherently encounters speckle noise as a characteristic challenge, which degrades image quality significantly. In recent years, learning-based methods have achieved remarkable success in SAR despeckling. However, these methods are often limited by complex speckle-noise patterns and the size of SAR images. To further improve despeckling quality and adapt arbitrary-size SAR images, we propose a novel method based on a conditional diffusion model coupled with a multi-scale and dual attention mechanism. During despeckling, the strong modeling capability of the diffusion model enables more accurate original clear scene restoration. Multi-scale features encompassing low-frequency large structure information and high-frequency details in SAR images are aggregated and input to position and channel dual attention mechanism, which effectively extracts global context information and establishes long-range dependencies, ensuring a broader perspective and more accurate distinction between structure textures and speckle noise. Using smooth noise estimation across overlapping patches to guide the sampling process, the proposed method is capable of handling arbitrary-size SAR images. Extensive experiment demonstrates that the proposed method has outstanding despeckling performance and flexible adaptability for arbitrary-size SAR images.

Ship detection is an important and challenging task in the field of synthetic aperture radar (SAR) image processing. Recently, deep learning technologies have yielded superior performance for object detection in remote sensing images. However, it is difficult to obtain the labels of SAR images, which limits the application of deep learning in ship detection from SAR images. To break the limitation of label information, we propose a self-supervised framework based on self-distillation for ship detection from SAR images in this paper. The framework consists of three core components: a self-supervised learning paradigm utilizing knowledge distillation, a deep residual shrinkage network (SAR-DRSN) model, and an oriented bounding boxes progressive generation model. The core of our method is a self-supervised variant of knowledge distillation, which propels the deep learning process in the absence of labeled data. The SAR-DRSN model excels in generating high-quality feature maps, significantly reducing the speckle noise. In addition, we introduce an iterative strategy for the accurate and precise delineation of ships, involving continuous refinement of oriented bounding boxes to optimize size and rotation angle for precise ship localization. Our experiments, obtained on two SAR datasets, demonstrate that the proposed method can achieve a satisfactory performance in ship detection without requiring any label information.

This article introduces a method for detecting ice floes in the Yellow River based on unmanned aerial vehicle (UAV) remote sensing images and deep learning technology. First, the article establishes a semantic segmentation dataset for Yellow River ice floes, filling the current data gap in this field. It designs a data augmentation scheme to improve the algorithm’s overfitting phenomenon. By analyzing the features of ice jam UAV remote sensing images and experimental verification, this article employs the U-Net algorithm as the foundation. By introducing the ConvNeXt-U feature extraction network, Squeeze-and-Excitation Network attention mechanism, and designing a hybrid loss function, the algorithm’s feature extraction ability is enhanced, improving the detection effect of small targets and overall improving the detection metrics of various semantic segmentation algorithms. Experimental results show that the proposed ConvNeXt-SE-U-Net increases accuracy by 16.05%, precision by 38.89%, and mean intersection over union by 30.35%. This method can accomplish high-precision ice floe semantic segmentation tasks in UAV remote sensing images, which is of great significance for Yellow River management.

Many existing multi-scale approaches for remote sensing image dehazing typically employ a unidirectional strategy, progressing from coarse to fine scales. This approach often leads to poor dehazing performance because information from subsequent scales is not fully leveraged and estimation errors at coarser scales will propagate to finer scales. To overcome the shortcomings and enhance the performance of image dehazing, bidirectional mechanisms are introduced to our proposed lightweight multi-scale bidirectional network (MSB-Net). The submitted MSB-Net comprises three intra-scale branches and two inter-scale branches. In the intra-scale branches, Transformer-based UNet integrates non-local sparse aggregation blocks to enhance global information recovery and improve modeling efficiency by learning global features at different scales. In the inter-scale branches, convolutional neural network (CNN)-based UNet augmented with bidirectional feature propagation blocks (BFPB) is employed to adaptively aggregate the multi-scale features at different scales. Although the local feature enhancement block in the CNN-based UNet is central to emphasizing local features, BFPB effectively exploits information spanning various scales for reconstruction by allowing bidirectional information propagation in both coarse-to-fine and fine-to-coarse flows. Experimental results demonstrate that MSB-Net outperforms other popular Transformer-based approaches, particularly achieving a 16% reduction in model parameters.

True orthophotos have become one of the most important spatial foundational geographic information due to their accurate geometric information and realistic texture features. The traditional method of producing true orthophotos is limited by the accuracy of digital elevation model (DEM)/digital surface model (DSM)/three-dimensional (3D) models. It inevitably has non-orthophoto issues, such as facade distortions and edge deformations, especially in areas with buildings. To address these challenges, we propose an innovative method that leverages a neural radiance field integrated with multi-resolution hash encoding. This method generates orthophotos directly from multi-view images and does not require additional data such as DEM, DSM, or 3D models. In comparison with the existing methods, our experimental results have achieved high-quality orthophotos, addressing various issues such as building facade distortions, building edge deformations, tree area deformations, and the presence of moving objects in orthophotos. In addition, evaluating the orthophoto generation method requires a ground truth dataset. We offer an open-source dataset containing multi-view images of unmanned aerial vehicle scenes and corresponding ground truth orthophotos of the area. This dataset comprises real-world and synthetic data, encompassing diverse scenes such as cities, trees, lakes, and more. It can be used for a comprehensive assessment of the quality of orthophotos produced by different methods. Our proposed method has been thoroughly evaluated using a wide range of challenging datasets. The experimental results demonstrate that our method outperforms traditional algorithms and the relatively latest method.

Small ships in synthetic aperture radar (SAR) images have small dimensions, making them susceptible to interference from sea waves, coastlines, and clutter between ships. To address the problem of low detection accuracy caused by the small scale of vessels and background noise, this paper proposes a high-precision network (Pico-Net). First, a feature selection backbone network is introduced to extract detailed information on small vessels. The background noise influence is removed through feature denoising convolution, and the feature attention spatial pyramid pool is constructed to highlight the contrast between small vessels and the background. Second, a multi-scale feature reuse dynamic fusion network (MRD-FPN) with bidirectional connections was designed to facilitate the acquisition of rich semantic information. Finally, a new loss function ZIoU is constructed by combining the advantages of CIoU, EIoU, and αIoU to effectively constrain the predicted bounding boxes. On the SAR ship detection dataset and the infrared ship detection dataset, Pico-Net achieved AP50−95 of 83.5% and 50.3%, respectively. The experimental results demonstrate that Pico-Net exhibits strong noise resistance, effectively combating background interference, and achieving more precise localization of small vessels, significantly reducing false alarms and missed detections.

Synthetic aperture radar (SAR) images and optical images have found widespread applications in various remote sensing fields. Although optical images are perceptible to the human visual system, they are susceptible to weather conditions and sunlight illumination. SAR can work all day and in all weather conditions, but SAR images suffer from geometric distortions, shadows, and speckle noise, severely affecting the human eye recognition of SAR images. SAR-to-optical (S2O) image translation can integrate the advantages of two types of images for the practical requirements in remote sensing applications. Existing methods for S2O image translation often produce low-quality translated images with notable deficiencies in color distribution and detail preservation. To address these problems, we propose a domain transfer generative adversarial network (DTGAN) based on U-Net and transformer for S2O image translation. The generator of DTGAN employs U-Net as a framework and incorporates two modules, a spatial transformer module, and a multi-scale channel transformer module. We additionally apply a multi-scale discriminator to train the generator. Extensive experiments are performed with other state-of-the-art methods, and the visual and quantitative evaluation results show that our method not only achieves the optimal metrics but also translates images that are more suitable for observation and analysis.

Extracting effective scene features is important for remote sensing image classification. Generally, the multi-view features contain information of consistency and complementarity, and efficient integration of them is helpful to enhance the performance of remote sensing image classification. Although some recent methods are able to achieve promising results, they lack analysis of the inherent relevance of multiple-view features. Thus, we present a multi-view fusion optimization method via low-rank tensor decomposition. First, the Laplacian matrix is constructed by utilizing K-nearest neighbors to generate a set of low-dimensional eigenvalues. Second, a third-order tensor is built by combining the multiple-view Laplacian features, which are factorized into many components with rank 1 using the canonical polyadic decomposition. Finally, the alternating optimization model is reconstructed by utilizing the relationship among fibers and slices of a tensor to generate optimal low-dimensional embedding features. Experiments of classification on three remote sensing image data sets AID, WHU, and UCMerced are constructed. The results of the experiments show that the new proposed method achieves better performance than others.

We report the development of a handheld, ultra-wideband impulse radar with very low size, weight, and power requirements. This radar is designed to remotely detect different layers in snow or ice that tend to crack or break under certain conditions indicative of potential avalanche hazards. First, the basic hardware design and configuration are introduced. Then, the development of a series of electromagnetics sensing models, which support the training and testing of machine learning algorithms for weak layer detection, is described. The principles and performance of these computational models are compared and validated with lab and snowpack measurements. Finally, the machine learning algorithms are applied to different snow profile measurements to generate hardness profiles for the user.

In remote sensing, accurate spectral characteristics are required to identify and describe the targets. Recently, miniature hyperspectral imagers (HySIs) have been flown on small satellites. The non-ideal shape of channels’ spectral response function (SRF) leads to a degraded spectrum, which is normally computed using the band average (BA) method. In this research, the system spectral shape factor (SSSF)-based method is proposed and demonstrated to restore the spectral shape and also to realize super spectral resolution. Computation of spectral radiance (SR) requires channel output and SSSF at that wavelength. SSSF is the convolution of the normalized input signal at a given wavelength and SRF. As the input spectrum is not known prior, coefficients required for SSSF computation are innovatively arrived at from the BA spectrum, which is coarsely similar to the input. The proposed method is tested using SRFs of Chandrayaan-1 and data of Airborne Visible/Infrared Imaging Spectrometer-Next Generation respectively simulating as sensor and input, respectively. The results confirm well matching of SSSF-based spectra with the original with very small deviations in SR values (0.6%), spectral angle map (0.5 deg), and signal information divergence (1×10−5). Slopes remained the same. This study opens up the possibility of optimizing sensor configurations and helps compute accurate spectra from miniature HySIs.

Hyperspectral imaging is a rapidly growing field that utilizes a diverse range of camera designs. It has been demonstrated that hyperspectral cameras can be constructed from commercial off-the-shelf (COTS) components, offering the potential for low-cost and widespread use. However, these COTS-based systems may face performance limitations due to optics not being optimized for hyperspectral purposes. Characterizing and comparing the performance of hyperspectral cameras is complex, a challenge recognized by the ongoing development of the IEEE P4001 standard. Specifically, the spatial coregistration among different spectral bands is crucial for the quality and integrity of the recorded spectra in each pixel. The proper evaluation of coregistration and resolution necessitates the measurement of the point spread function for each band. We use measurements of the full 2D sampling point spread function (SPSF) to compare the performance of two hyperspectral imagers (HSIs) built from COTS components: the engineering model of the hyperspectral camera onboard the Hyperspectral Cubesat for Ocean Observation-1 cubesat and an extremely compact HSI developed for drone flights. In addition, the line spread functions across and along the track, keystone distortion, and coregistration errors are derived from the SPSFs. A simplified measurement scheme is also tried and found to provide fair accuracy for the tested cameras. The results highlight the importance of measuring the spatial SPSF for characterizing and comparing different hyperspectral cameras. Results also show that good spectral integrity can be achieved through spatial binning, which of course requires a tradeoff against spatial resolution. Nonetheless, it is evident that such low-cost systems can offer useful capabilities.

Single-vehicle light detection and ranging (LiDAR) has limitations in capturing comprehensive environmental information. The advancement of vehicle-to-infrastructure (V2I) collaboration presents a potent solution to this challenge. During the collaboration, point cloud registration precisely aligns data from various LiDARs, effectively mitigating the constraints associated with data collection by a single-vehicle LiDAR. Registration furnishes autonomous vehicles with a more comprehensive and dependable environmental understanding. Currently, there are various types and performances of LiDAR in practical application scenarios. So, it is more necessary to perform heterogeneous point cloud registration, and there is still a relatively large room for improvement. Consequently, we introduce a coarse-to-fine approach to heterogeneous point cloud registration (C2F-HPCR), establishing the inaugural benchmark for point cloud registration in intricate vehicle-infrastructure collaboration contexts. C2F-HPCR acquires an initial registration matrix through its coarse registration module. Subsequently, it uses the overlap estimation module to extract overlap points between two point clouds. These identified points are inputted into the fine registration module to obtain the final registration matrix. Experiments on the DAIR-V2X-C dataset demonstrate that the recall of C2F-HPCR in heterogeneous point cloud registration is 72.99%. C2F-HPCR shows strong performance in heterogeneous point cloud registration, facilitating efficient registration of vehicle-side point clouds and infrastructure-side point clouds. The code is available at https://github.com/916718212/C2F-HPCR

Clouds are one of the major factors leading to the failure of optical remote sensing observations on Earth, and reconstructing the information loss caused by clouds in optical images has attracted significant attention. Existing methods primarily utilize multi-temporal optical images as references, which makes it challenging to reflect real-time information. Due to the cloud-penetrating capability of synthetic aperture radar (SAR) imaging, SAR images can serve as reference images to capture real-time information during the cloud removal process. We propose a cloud removal method, DS2-SO-CR, that leverages multi-source data fusion using a conditional generative adversarial network (CGAN) combined with a residual network to effectively remove clouds from multi-cloud images. We used a CGAN model to obtain SAR-to-optical images, which are then input into a fusion residual network along with SAR images and cloudy optical images. This significantly enhances the richness and accuracy of the texture in the cloud-removed images. To ensure the preservation of cloud-free regions in cloudy images during the fusion process, we added a cloud detection component to the fusion network. The proposed method was validated on the SEN12MS-CR and HN-TS datasets, achieving the following accuracy metrics: PSNR of 30.48, SSIM of 0.892, MSE of 0.038, LPIPS of 0.146, and FID of 37.53. This study provides a reference for cloud removal in remote sensing imagery of cloudy regions, as well as support for monitoring geological disasters, vegetation, and land cover changes.

Traditional binary change detection (BCD) in remote sensing focuses solely on detecting the occurrence of change, whereas semantic change detection (SCD) provides deeper insights by recognizing specific target categories before and after modifications, making it a more advanced approach for land cover SCD (LCSCD). However, current SCD approaches in remote sensing face challenges such as slight changes, irregular regions, and fuzzy boundaries in accurately detecting complex land cover changes, which limits their practical utility. Here, we propose an innovative SCD approach, DRGMS-Net, which utilizes a dual-resolution guided multiscale network for LCSCD. DRGMS-Net integrates semantic segmentation and BCD, which can efficiently extract essential features for accurately detecting changes. First, in the semantic information encoder, we utilize a dual-resolution network for extracting features at varying resolutions and propose a multiscale aggregation pyramid pooling module to extract multilayer features, providing more detailed and semantic features for LCSCD. Moreover, we introduce a strip attention module and feature interaction layer to extract features of irregular regions and fuse bitemporal features. In the semantic information decoder, we propose a change information aggregation module to better capture change features and enhance the representation of irregular objects. Finally, we design a combined loss based on PolyLoss and semantic loss to address the challenge caused by imbalanced pixels. We demonstrate the effectiveness of DRGMS-Net on two LCSCD datasets. The values of separation kappa are 20.84% and 44.21%, and the values of mean intersection over union are 72.22% and 84.39%. The results are superior to those of other state-of-the-art methods and highlight the potential of DRGMS-Net in providing detailed information about altered regions.

The multispectral and hyperspectral image fusion (MHF) technique is designed to address the challenge of integrating the spatiotemporal characteristics of multispectral images with those of hyperspectral images. In the initial stages, the image to be fused is primarily decomposed into endmembers and abundances using a matrix decomposition method. This approach, however, may disrupt the correlation of spectral data. Subsequently, tensor decomposition-based methods emerged, with the most representative being canonical polyadic decomposition and Tucker decomposition. These methods are widely applied to the MHF problem due to their good recoverability. However, they lack the ability to introduce the physical interpretation of potential factors into the framework, and it is difficult to improve the quality of the fused image by utilizing the physical properties of the endmembers. Consequently, we employ a block term tensor decomposition algorithm based on sparse regularization to estimate the optimal high spatial resolution hyperspectral image. First, the abundance information is reconstructed into a chunk matrix, and its sparsity is characterized by introducing the l2,1 norm to eliminate the scaling effect present in the model. Second, the endmembers’ matrix sparsity is facilitated by the introduction of the row sparsity of the l2,1 norm regularization, which eliminates the inverse scaling effect present in the model. Finally, the model is solved using the conjugate alternating iteration algorithm. Experiments on three standard datasets and two local datasets demonstrate that this method outperforms state-of-the-art methods.