Guest Editors Elhadi Adam, Craig Coburn, and Anthony Campbell introduce the Special Section on the 50th Anniversary of Landsat—Current Achievement and Future Directions.

The aim of this study is to map and analyze the lineament network in the Edéa, Cameroon, area using remote sensing and gravimetric data to determine their hydrogeological implications. Principal component analysis and directional filters applied to Landsat7 ETM+ and Shuttle Radar Topography Mission imagery, respectively, were used to extract remote sensing lineaments. Rose diagram of these lineaments highlights four families of lineaments along the N–S, E–W, NE–SW, and NW–SE directions. There are three major directions accounting for 74% of lineaments, including N0° to N10°, N20° to N30°, and N40° to N50°; and four minor directions (with 26% of the lineaments), including N60° N70°, N80° to N90°, N130° to N140°, and N150° to N160°. N20° to N90° directions correlate with those of major structures of the Oubanguides Complex, such as the Sanaga Fault and Central Cameroon Shear Zone. N130° to N140° direction corresponds to orientation of Shear Zones and blastomylonitic faults of Nyong Complex. Superposition of these lineaments on hydrographic network shows similarities between their directions, thus highlighting strong impact of tectonics on orientation of hydrographic network. The presence of numerous lineaments highlights strongly fractured subsoil, and their high density favors the circulation and accumulation of groundwater. Upward continuation and horizontal gradient maxima methods applied to Earth Gravitational Model 2008 data allowed the extraction of gravimetric lineaments, with a major N–S orientation, which correlates with general orientation of South Atlantic opening. Superposition of remote sensing lineaments and gravimetric lineaments highlights their parallelism, admitting that gravimetric structures are an extension in depth of surface structures defined by remote sensing.

The rapid, accurate, and intelligent extraction of vegetation areas is of great significance for conducting research on forest resource inventory, climate change, and the greenhouse effect. Currently, existing semantic segmentation models suffer from limitations such as insufficient extraction accuracy (ACC) and unbalanced positive and negative categories in datasets. Therefore, we propose the Attention U-Net model for vegetation extraction from Landsat8 operational land imager remote sensing images. By combining the convolutional block attention module, Visual Geometry Group 16 backbone network, and Dice loss, the model alleviates the phenomenon of omission and misclassification of the fragmented vegetation areas and the imbalance of positive and negative classes. In addition, to test the influence of remote sensing images with different band combinations on the ACC of vegetation extraction, we introduce near-infrared (NIR) and short-wave infrared (SWIR) spectral information to conduct band combination operations, thus forming three datasets, namely, the 432 dataset (R, G, B), 543 dataset (NIR, R, G), and 654 dataset (SWIR, NIR, R). In addition, to validate the effectiveness of the proposed model, it was compared with three classic semantic segmentation models, namely, PSP-Net, DeepLabv3+, and U-Net. Experimental results demonstrate that all models exhibit improved extraction performance on false color datasets compared with the true color dataset, particularly on the 654 dataset where vegetation extraction performance is optimal. Moreover, the proposed Attention U-Net achieves the highest overall ACC with mean intersection over union, mean pixel ACC, and ACC reaching 0.877, 0.940, and 0.946, respectively, providing substantial evidence for the effectiveness of the proposed model. Furthermore, the model demonstrates good generalizability and transferability when tested in other regions, indicating its potential for further application and promotion.

Salt marshes provide extensive ecosystem services, including habitat, recreation, coastal resilience, and carbon sequestration. The United States has the largest extent of mapped salt marshes. Therefore, it is critical to understand the ecosystem’s carbon stock and drivers in the contiguous United States (CONUS). Blue carbon ecosystems, including salt marshes, mangroves, and seagrasses, store most of their carbon within the soil; aboveground biomass (AGB) is an important ecosystem indicator. Existing AGB models in salt marshes have medium spatial resolution and limited geographic extent. To improve the spatial resolution to 10 m, we evaluated the use of Sentinel-1 and 2 data for inclusion into the AGB prediction. To incorporate these satellite observations with temporally disparate in situ samples, we evaluated the stability of training locations using the Landsat time series, finding that 71% of training data were stable from field sampling to remote sensing observation in 2020. Next, we trained a machine learning regression combining Sentinel-1, Sentinel-2, and Landsat data to predict AGB in salt marshes. We compared model performance with in situ testing data between three machine learning algorithms [support-vector machines, random forest, and extreme gradient boosting (XGBoost)], spatial scale (10, 30 m), and training data stability. The best-performing model was the 10 m XGBoost using the stable training data, which achieved a root mean square error of 301.0 and 107.33 at the plot and site scale, respectively. We created an updated 2020 salt marsh extent with Sentinel-1/2, Shuttle Radar Topography Mission, and National Elevation Dataset and estimated 3.6 (3.1 to 4.1) Tg of aboveground carbon across the CONUS. We explored salt marsh biomass drivers and found that the primary drivers of AGB are relative sea level rise, temperature, precipitation, and tidal amplitude. Our results demonstrate the need to monitor these systems to enable management, restoration, and understanding of the ecosystem’s resilience to climate change.

Remote sensing retrieval of sea ice thickness can quickly and accurately obtain the spatiotemporal distribution information of large sea areas. This has important implications for marine resource utilization and global climate change impacts. This study aims to take the Bohai Sea as an example, use Landsat 8 data, and analyze the spectral characteristics of sea ice to establish sensitive indicators for sea ice thickness remote sensing retrieval and determine the functional relationship between these indicators and sea ice thickness. This paper first uses variance analysis to select the sensitive wavebands for sea ice thickness inversion and then uses factor analysis to conduct a comparative sensitivity analysis on various combinations of sensitive wavebands. Subsequently, the relationship between sea ice thickness and sensitive zones was quantitatively analyzed through regression analysis. Finally, the accuracy of the model was verified using measured data of sea ice thickness in the Bohai Sea in 2016 and 2023, compared with the existing albedo model, and the application of the model was demonstrated. The experimental results of variance analysis and factor analysis show that the visible light band shows strong sensitivity to sea ice thickness, among which the red light band is the most sensitive. The linear-weighted combination of visible light bands (B1-coastal, B2-blue, B3-green, and B4-red) shows a significant linear correlation with sea ice thickness, with a linear regression square value as high as 0.9709. Experimental verification shows that there is a significant linear correlation between the sea ice thickness inverted by each model and the measured thickness data. Research conclusions: (1) variance analysis and factor analysis methods can effectively select sensitive bands and evaluate the inversion effect of band combinations. (2) Either a single red light band or a linear combination of visible light bands can be used as a sensitive indicator for remote sensing retrieval of sea ice thickness. (3) There is a significant linear correlation function between sensitive indicators and sea ice thickness. (4) The measured data confirms the accuracy of the model. Compared with the albedo model, the inversion model proposed in this article is more sensitive to sea ice thickness and is less affected by clouds and sediment. It can provide a reference for the inversion simulation of sea ice thickness and also provide a demonstration for the application of the temporal and spatial distribution laws of sea ice.

Soil moisture is a key factor affecting the growth of crops, and microwave remote sensing is one of the most important methods for inverse studies of soil moisture in agricultural fields in recent years. Cotton is a typical water-demanding crop in arid zones, and accurate estimation of soil moisture information in cotton fields is extremely important for optimizing irrigation management, improving water use efficiency, and increasing cotton yield. This study focuses on extracting feature sets by combining Sentinel-1 and Gaofen-6 satellite data, constructing convolutional neural network (CNN), random forest, support vector regression, and eXtreme gradient boosting model to estimate the soil moisture in cotton fields in Shihezi area of Xinjiang, and designing eight groups of experiments according to the different input data sources. The experimental results show that the accuracy of the soil moisture estimate in cotton fields in arid areas with multi-source data is significantly better than that of a single data source. Moreover, the CNN was best estimated when using multi-source data feature sets as inputs, with a coefficient of determination of 0.789, a root mean square error of 0.0249 cm3/cm3, and an average absolute error of 0.0198 cm3/cm3 for its CNN model. This result demonstrates the effectiveness of CNN in soil moisture estimation and also provides a new method for the use of multi-source remote sensing data for accurate soil moisture estimation in cotton fields in arid areas, and also explores the application of Gaofen-6 data in soil moisture.

High-resolution remote sensing images (HRRSIs) contain rich local spatial information and long-distance location dependence, which play an important role in semantic segmentation tasks and have received more and more research attention. However, HRRSIs often exhibit large intraclass variance and small interclass variance due to the diversity and complexity of ground objects, thereby bringing great challenges to a semantic segmentation task. In most networks, there are numerous small-scale object omissions and large-scale object fragmentations in the segmentation results because of insufficient local feature extraction and low global information utilization. A network cascaded by convolution neural network and global–local attention transformer is proposed called CNN-transformer cascade network. First, convolution blocks and global–local attention transformer blocks are used to extract multiscale local features and long-range location information, respectively. Then a multilevel channel attention integration block is designed to fuse geometric features and semantic features of different depths and revise the channel weights through the channel attention module to resist the interference of redundant information. Finally, the smoothness of the segmentation is improved through the implementation of upsampling using a deconvolution operation. We compare our method with several state-of-the-art methods on the ISPRS Vaihingen and Potsdam datasets. Experimental results show that our method can improve the integrity and independence of multiscale objects segmentation results.

The traditionally used rain gauge stations provide the most reliable data on the spatiotemporal distribution of precipitation; however, they are limited in space and time. As an alternative to field observations, the gridded precipitation products (GPPs) offered by remote sensing missions are widely used. On account of the uncertainties associated with the GPPs, they have to be quality-checked for confidence in application over the region of interest. Although accuracy assessment of precipitation data is a common task, there is a gap in the literature regarding a comprehensive assessment of the currently available GPPs. In this study, 14 GPPs were used to investigate their performance in catching the spatio-temporal characteristics of precipitation over geographic regions of Türkiye. According to the results, integrated multi-satellite retrievals for Global Precipitation Measurement (IMERG), multi-source weighted-ensemble precipitation (MSWEP), and Tropical Rainfall Measuring Mission (TRMM) show better performance on monthly and annual scales while on the climatology scale, CHELSA, Climate Hazards Group Infrared Precipitation with Station, ERA5, and ERA5-Land also manifest better performance. The mean monthly correlation over Aegean Region (AEG), Marmara Region (MAR), Central Anatolian Region (CAR), Mediterranean Region (MED), Black Sea Region (BSR), East Anatolian Region (EAR), and South East Anatolian Region (SEA) are 0.77, 0.81, 0.77, 0.80, 0.79, 0.77, and 0.77, respectively. The annual assessment suggests that over the MAR, CAR, MED, and SEA, the IMERG mission performs very well. While TRMM showcases its best performance in the AEG, MED, BSR, and EAR, MSWEP performs well in the BSR and SEA Region. Overall, taking the country-average results into account, it can be stated that among the used GPPs, TRMM, MSWEP, and IMERG yield the best results for Türkiye as a whole.

The mitigation of energy usage in urban areas, especially in buildings, has recently captured the attention of many city managers. Owing to the thermal images’ limited resolution, especially at the edges, creating a high-resolution (HR) surface model from them is a challenging process. This research proposes a two-phase strategy to generate an HR four-dimensional thermal surface model of building roofs. In the single-source modification phase, an enhanced thermal orthophoto is produced by retraining the enhanced deep residual super-resolution deep network, and then, using state-of-the-art structures from motion, semi-global matching, and space intersection. The final surface model’s resolution is raised by combining thermal data with visible unmanned aerial vehicle images to overcome the limitation of single-source methods in resolution increase. To this end, after visible orthophoto and digital surface model generation, buildings and their boundaries are extracted using the multi-feature semantic segmentation method. Next, in the multi-source modification phase, a fine-registered enhanced thermal orthophoto is generated, and thermal edges are identified around the boundary of the building. The visible and thermal boundaries are then matched, and any smoothness in the temperature edges is eliminated. The results show that the average difference in position between the thermal edges and building boundaries is reduced, and temperature smoothness is completely eliminated at the building edges.

Identifying coral reef ecosystem benthic substances is extremely important for protecting coral reef ecosystems and monitoring their health status. Using remote sensing images and the object-based image analysis (OBIA) method could effectively improve the identification accuracy of coral reef ecosystem benthic substances. Among them, the core of the OBIA is the selection of optimal scale parameters. In this study, we used Sentinel-2 (S2) and Gaofen-2 (GF-2) images to identify the benthic substances of coral reef ecosystems in the coastal area of Sanya City, Hainan Province. The results show that: (1) the simple benthic substances of coral reef ecosystems could be identified through the analysis of spectral separability of the benthic substances from remote sensing images, and the optimal scale parameters were determined by combining the estimation of scale parameters model method and visual identification (74 for large-scale areas and 3 for small-scale coral reef ecosystems); (2) the S2 images combined with pixel-based maximum likelihood method and support vector machine (SVM) could identify three benthic types of coral reef, coral bleaching, and sand, whose overall accuracy (OA) was 83.89% and 88.89%, respectively; (3) the GF-2 images combined with object-based K-nearest neighbors and SVM could effectively identify six benthic types of coral reef, coral bleaching, algal ridge, algae, sand, and rock, whose OA was 85.39% and 89.39%, respectively; and (4) compared with the S2 images, the GF-2 images could identify finer types of benthic substances and provide data for the analysis of the benthic substances in coral reef ecosystems, and the results of GF-2 images combined with object-based SVM showed that coral reefs in Dongmaozhou Island had the best development; coral reefs in Luhuitou Peninsula and Yulin Bay suffered severe bleaching; coral reefs in Ximaozhou Island are ordinary under serious threat from other benthic organisms. In conclusion, remote sensing technology combined with appropriate interpretation algorithms can improve the identification accuracy of coral reef ecosystem benthic substances, which can effectively maintain the health status of coral reefs and realize the long-term protection of the coral reef ecosystems.

Accurate and timely monitoring of leaf nitrogen content (LNC) is crucial for citrus tree growth control and precision fertilization. Unmanned aerial vehicle (UAV) remote sensing can provide stronger support for rapid, flexible, large-scale monitoring of LNC distribution in orchards compared with traditional LNC chemical detection methods. However, time-consuming and high-cost sample collection leads to a limited number of samples for constructing remote sensing LNC estimation models, thereby limiting the accuracy of LNC estimation. We use the semi-supervised twin neural network regression (STNNR) model to make full use of a considerable portion of unlabeled and labeled data to extend the sample data for training the model and effectively improve the accuracy of nitrogen content estimation. First, multiple spectral indices are constructed using high-resolution multispectral UAV imagery, and the relationships of these spectral indices with the LNC are analyzed to select the spectral indices sensitive to changes in the LNC. Then, the unlabeled data are used for model training under the loop consistency condition. Three different loss functions are employed for training the STNNR model on the labeled and unlabeled data. Finally, the STNNR model is evaluated through comparison experiments with other supervised learning models. Results show that the STNNR model has a high LNC estimation accuracy with a root mean square error of only 0.59 g/kg and an R2 of 0.71. The proposed method is capable of effectively obtaining LNC in the whole orchard with limited samples, thus providing support for the scientific management of citrus planting.

Investigating ground objects widely distributed in geography and large in scale is one of the primary missions for satellite sensors. On the other hand, recognizing objects from images is one of the classic tasks for convolutional neural networks (CNNs), currently the most popular computer vision technique. Data processing, such as data augmentation, channel selection, and image fusion, can be essential when applying CNNs to satellite images. With a case study of recognizing solar panels from satellite images using CNN, the related data processing issues are discussed, and an approach to embed channel fusion methods into CNN is established. As a result, the following findings are concluded from our case study: (1) not all channels in satellite images contribute to specific object recognition, and thus channel selection is necessary in applying CNN on satellite images; (2) fine-tuning the fusion method embedded in CNN improves the model stability; and (3) transfer learning is outperformed by CNN models trained with augmented data for object recognition from satellite images.

The fluorescence explorer European Space Agency mission with its payload, the fluorescence imaging spectrometer (FLORIS), aims to perform quantitative measurements of the solar induced chlorophyll fluorescence from space, with the purpose of improving the monitoring of the health of Earth vegetation. The retrieval of a faint signal, such as the one from chlorophyll fluorescence, requires very low stray-light (SL) levels. The SL reduction in FLORIS implied constraints impacting the design, by means of using low roughness optical components; the integration, through a strict contamination control; and the data processing with the need of a very accurate correction strategy making use of numerical models well correlated with experimental data, to be acquired during the on ground calibration activities. To assess and validate the correction strategy, an accurate SL characterization has been anticipated during the FLORIS optical model refurbished campaign. Different methods, such as out-of-field and out-of-band measurements, have been investigated to avoid the detector blooming affecting measurements with high input signals. By combining the results from the different approaches, it has been possible to achieve up to 9-10 orders of magnitude of explored dynamic range. The model, correlated with measurements, has finally proved to be capable of correcting SL with a reduction factor of the order of 10, down to a level of the order of 30% of the required fluorescence error (10% of a fluorescence level of about 2 mW m−2 sr−1 nm−1).

Significant advancements in ship detection technology in synthetic aperture radar (SAR) images have been made in recent years due to the rapid development of deep learning technology. Nevertheless, there are still certain issues with the current deep learning-based detection algorithms when detecting small-scale, arbitrary-direction, and densely packed ship targets in SAR images. We offer a rotating target detection network based on improved YOLOv8 as a solution to these issues. First, a new small target sampling layer is introduced in YOLOv8 to enhance the model’s ability to detect small-target objects by reducing the down-sampling rate. Second, the learnable weights are introduced to adaptively adjust the feature information at different scales and superimpose shallow features on deep features to better extract feature information when the model conducts feature fusion. Finally, the offset of the predicted bounding box concerning the true bounding box can be expressed more accurately using distribution focal loss and probabilistic intersection-over-union as the loss functions. Extensive experiments are conducted on rotated ship detection dataset in SAR images (RSDD-SAR) and SAR ship detection dataset (SSDD) to demonstrate the effectiveness of the improvement. BP-YOLOv8 achieved an F1 score of 93.01% and 97.48% on the RSDD-SAR and SSDD datasets, with an average precision of 95.87% and 98.82%, respectively.

Soil moisture (SM) is a crucial meteorological parameter affecting agricultural production and ecosystems, so accurately capturing SM is significant for farm irrigation and ecosystem protection. The delay Doppler maps data of the Cyclone Global Navigation Satellite System (CYGNSS) contains characteristic information related to SM, providing a new observation method for SM observation. First, the CYGNSS data are matched to the soil moisture active and passive dataset of 36 km×36 km grid by the nearest neighbor method. Then, a random forest regression (RFR) model is constructed for training and prediction. It is found that the RFR model based on the combination of 13 characteristic variables can accurately predict SM. On the training set, the correlation coefficient (R) of the model is 0.966, the root mean square error (RMSE) is 0.026 cm3 cm−3, and the mean error (ME) is 0.000 cm3 cm−3. On the test set, the R of the model is 0.903, the RMSE is 0.041 cm3 cm−3, and the ME is −0.001 cm3 cm−3. Simultaneously, to study the advantages of RFR prediction, this study also compared it with multiple linear regression (MLR) predictions. The results show that the RFR algorithm has higher accuracy both in the training set and the test set. Compared with MLR, R increased by 26.3% in the training set and RMSE decreased by 59.4%. R increased by 17.1% in the test set, and RMSE decreased by 35.9%. This demonstrates that the RFR model with multi-characteristic variables has good reliability, making it an efficient way to use satellite remote sensing data for SM retrieval.

The NASA Clouds and the Earth’s Radiant Energy System (CERES) project provides observed flux and cloud products to the climate science community. The CERES instruments, along with the moderate resolution imaging spectroradiometer (MODIS) and visible infrared imaging radiometer suite (VIIRS) imagers, are onboard the Terra, Aqua, NPP, and NOAA20 satellites. To produce seamless multi-platform integrated products, long-term sensor stability and inter-calibration are required. Inter-calibration between sensors within the same Sun-synchronous orbit relies on Earth invariant targets because simultaneous nadir overpasses are not possible. To facilitate inter-calibration efforts, the CERES Imager and Geostationary Calibration Group has improved the Libya-4 target characterization. Improvements include full scan angle characterization to enable daily observations, clear-sky identification using individual scan angle dynamic spatial homogeneity thresholds, and atmospheric corrections. The water vapor correction was found to be effective across the full scan, whereas the ozone and aerosol corrections were less effective. The atmospheric-corrected normalized radiance temporal fluctuations are similar across scan angles and spectral bands and between the MODIS and VIIRS imagers, suggesting that the fluctuations are a result of the natural variability of the Libya-4 surface reflectance. The Libya-4 surface variability is more than likely caused by changes in the prevailing winds that alter sand dune orientation and resulting shadows. The full-scan-imager atmosphere and angle corrected reflected solar band radiance trend standard errors are between 0.6% and 1.0%, and for near-nadir observations, they are between 0.5% and 0.8%. The advanced characterization suggests that the Libya-4 short-term surface reflectance anomalies may need to be considered for imager stability monitoring and inter-calibration efforts.

Multi-source remote sensing data fusion can provide more effective features for crop recognition, but too many feature parameters will cause data redundancy. This work takes the Honghe Hani area of Yunnan Province as the research area, relies on the Google Earth Engine cloud platform, and collaborates with Sentinel-1/2 and Shuttle Radar Topography Mission digital elevation model multi-source remote sensing data to propose a strategy for extracting paddy rice planting areas based on feature optimization and phenological information. First, four kinds of features, including spectrum, texture, polarization, and terrain, are constructed for multi-source remote sensing data, and three critical phenological periods are determined by combining normalized difference vegetation index and VH backscattering coefficient time series curves. Then, recursive feature elimination combined with random forest (RF) is used to optimize the characteristics and determine the optimal phenological period. Finally, paddy rice planting areas are extracted using an RF classifier. The results show that the optimal phenological period for paddy rice extraction is the flooding and transplanting period, the number of features decreases from 48 to 13, the overall accuracy is 98.41%, and the Kappa coefficient is 0.97. The optimized scheme not only reduces the data redundancy but also has higher extraction accuracy than the other 14 feature combination and comparison schemes.

Seaweed biomass is mainly investigated using quadrat sampling. However, the spatial limitations of quadrat sampling pose significant challenges to the thorough evaluation of seaweed resources at a larger scale. We combined multispectral photography collected using an unmanned aerial vehicle with satellite imagery and in situ hyperspectral reflectance measurement to determine the biomass of Ulva pertusa and Sargassum spp. in the intertidal zone of Gouqi Island, Zhejiang Province, China. The statistical metrics derived from aerial imagery for the biomass estimation of Ulva pertusa and Sargassum spp. achieved coefficients of determination (R2) of 0.81 and 0.74, root mean square errors (RMSE) of 96.54 and 788.64 g/m2, and mean absolute errors (MAE) of 40.71 and 327.12 g/m2, respectively. For satellite imagery, the R2 values were 0.80 and 0.68, RMSE were 97.65 and 910.45 g/m2, and MAE were 71.78 and 719.69 g/m2, respectively. The biomass of U. pertusa and Sargassum spp. fluctuated monthly, reaching their highest values in June and July, respectively. The total biomass of U. pertusa and Sargassum spp. was 1.69×107 and 9.93×107 g, and the organic carbon content was 7.98×105 and 5.04×106 g, respectively. Our research offers a new perspective and methodology, supplying critical data and a scientific rationale for the management of marine assets and the safeguarding of coastal ecologies.

In response to a community-identified need for ground-based thermodynamic (TD) profiling of the troposphere, we present the further development and validation of a differential absorption LiDAR (DIAL) technique to retrieve temperature. This paper showcases the accuracy of temperature retrievals using a perturbative technique, combining a DIAL measurement of a temperature-sensitive oxygen (O2) absorption profile with a high spectral resolution LiDAR measurement of the backscatter ratio profile near 770 nm. This study introduces three key advancements. First, the spectroscopic model used to represent the absorption of light by O2 is enhanced via a more complete physical representation, improving measurement accuracy. Second, the error estimation and masking are developed using the bootstrapping technique. Third, we present a comparison of temperature profiles from our laboratory-based instrument with collocated radiosondes, evaluating the accuracy of our updated measurements. It is essential to clarify that the instrument described in this paper does not operate as a stand-alone TD profiler, as it is not capable of measuring water vapor (WV). Instead, we focus on demonstrating the perturbative retrieval technique with temperature profiles inferred using ancillary radiosonde WV profiles. Results from a full TD profiling instrument will be presented in a future publication. The laboratory-based LiDAR instrument was operated over a 6-month period between April 21, 2022, and September 22, 2022. During this time, we launched 40 radiosondes, providing reference data to validate the accuracy of the DIAL-based temperature profiles. The results indicate that DIAL-based temperature retrievals are within ±2.5°C between 0.4 and 3 km (3.5 km) during daytime (nighttime) operation, using a 300-m range resolution and a 60-min time resolution.

Crop type classification is crucial for policymaking and precision agriculture applications. This study aimed to develop a parcel-based maize (Zea mays L.) extraction approach using Sentinel 2A-derived spectral indices and machine learning (ML) in two distinct study sites: Yeşilova and Ormankadı villages in Bursa Province, Turkey. Employing 13 widely recognized spectral indices, the investigation implemented 4 ML classifiers: support vector machines, random forest, K-nearest neighbors, and bootstrap aggregating. The training-test methodology was explored using two scenarios: Yeşilova as the training set and Ormankadı as the test set, and vice versa. The models calibrated on Yeşilova and validated on Ormankadı maintained the accuracy of the model, with an overall accuracy (OA) ranging from 79.3% to 89.9%, precision between 72.8% and 80.1%, recall between 82.1% and 84.9%, F1-score between 77.4% and 82.2%, and a Matthews correlation coefficient (MCC) ranging from 58.9% to 68.3%. Furthermore, the models consistently demonstrated good performance when Ormankadı served as the training set and Yeşilova as the test set, with commendable OA (78.7% to 84.8%), precision (85.5% to 88.0%), recall (88.0% to 91.1%), F1-score (86.2% to 89.5%), and MCC (68.2% to 76.0%). This study demonstrated the potential of using high-resolution remote sensing and ML for effective maize crop extraction using diverse datasets.

This study explores the potential of improving vision transformer, MLP-Mixer, and global filter network for cloud detection using Sentinel-2 images. A mechanism is proposed to embed K-means clustering into frameworks for cloud detection (KET-CD) to enhance the performance. A comparative analysis of these architectures for multiclass cloud detection is conducted by leveraging two standard datasets: IndiaS2 and KappaSet. Our results demonstrate that the KET-CD methods achieve remarkable accuracy by outperforming the baseline architectures and state-of-the-art cloud detection methods (Fmask and Sen2Cor), exceeding a 0.5 mean intersection over union score. By incorporating cluster label embedding, KET-CD significantly improves the ability of these models to detect clouds and their shadows effectively in complex regions. These findings validate the importance of transformer-inspired architectures in processing large-scale satellite data for cloud detection tasks and highlight KET-CD as a promising solution for accurate cloud detection in remote sensing applications.

For the land type classification from multi-source remote sensing images with deep learning models, domain adaptation is usually required to improve their generalization ability and work efficiency. We design a dual contrastive learning structure that performs separate contrasts in the encoder and decoder for domain adaptation of the model. Its core utilizes the advantages of contrast learning in classifying unlabeled data. Further optimizations are made toward the loss functions used in the network structure, including a dynamic classification weight adjustment algorithm to treat imbalanced sample numbers and a local dynamic threshold method to filter local noises. With the Potsdam dataset as the source domain and the Vaihingen and LoveDA datasets as the target domains, two sets of experiments are conducted to verify the effectiveness of the model. In addition, corresponding ablation experiments are designed to verify the necessity of the dual contrastive structure. The results show that our method achieved superior performance in building extraction when compared with state-of-the-art methods. Furthermore, it is indicated that the dual contrastive structure is more effective than a single contrastive structure in terms of building extraction.

Semantic segmentation of remote sensing images in urban scenes suffers from blurred multi-scale target boundaries, insufficient use of global context, and classification errors caused by high inter-class variance and low intra-class variance. Therefore, we propose a semantic segmentation network with edge and class guidance (ECGNet). First, ECGNet introduces multi-scale edge prior knowledge to address the problem of blurred target boundaries. Second, ECGNet applies synergistic class augmented attention to introduce class prior knowledge while retaining rich spatial dimensional localization information to alleviate the problem of classification errors caused by low intra-class variance and high inter-class variance. Finally, the multi-scale large receptive field attention in ECGNet simulates a large convolutional kernel to capture multi-scale global context information. Experiments conducted on the ISPRS Vaihingen and ISPRS Potsdam datasets show that the proposed method is competitive.

Extracting edge information is a major challenge for polarimetric synthetic aperture radar (PolSAR) image processing techniques, particularly when weak and robust edges emerge concurrently within urban areas. Despite their popularity, classic region-based PolSAR image edge detectors perform non-optimally in heterogeneous areas as the covariance matrix does not obey the Wishart distribution in such cases. To address this problem, a novel PolSAR edge detection algorithm that utilizes a change in the dominant scattering mechanism as a basis of edge detection is examined. An edge is assumed to exist whenever the dominant scattering mechanism for a given pixel is different from its neighboring pixels. Because solid edges/outer boundaries of homogeneous areas are better identified by the classical filtering-based method, its fusion with the proposed algorithm is used to retain the best edge information. The performance of the suggested technique is assessed using two real PolSAR datasets and compared with state-of-the-art techniques. The results demonstrate that the proposed method is more accurate in identifying and preserving the edges corresponding to different ground objects/structures.

Critical information about surface water bodies, particularly their dynamic behavior, is most effectively derived from water contour detection. However, the accurate detection of contours is complicated by the land–water ambiguity and the great imbalance between contour and non-contour data. A unique fully convolutional multiscale UNet-styled (MS UNet) deep network is proposed for accurate water contour detection in the visible spectrum. The MS UNet utilizes blocks of multiscale convolutional filters to improve contour detection and employs loss functions to correct the imbalance between contour and non-contour data, as well as capture the loss at both the pixel and object levels. The proposed system is shown to be more effective at detecting water contours than recent water detection systems and other popular image segmentation networks while using a fraction of the parameters.

Road traffic accidents have become a global problem. Crosswalks and intersections are small road targets in remote sensing images and, as an important part of the roadway, also have a high incidence of traffic accidents. The recognition of the crosswalks and intersections in high-resolution remote sensing images with more detailed information can provide drivers with timely information to prevent or reduce traffic accidents. Aiming at the difficulty of detecting small road targets in remote sensing images, an improved model based on YOLOv5—RS-YOLOv5 is proposed. Specifically, we designed a cross-layer feature enhancement module in the neck to capture the position information of small targets, allowing the model to locate small targets more accurately. Meanwhile, a bi-directional feature pyramid network is used as a feature fusion network to more fully obtain contextual information. Finally, to pay more attention to small targets during training and alleviate the imbalance between positive and negative samples of small targets, a loss function combining normalized Wasserstein distance loss and intersection over union loss is designed to improve the positioning loss. In the experiment, we evaluate models on two remote sensing datasets (Potsdam and Shanghai remote sensing datasets), with mAP50 reaching 69.55% and 62.9%, respectively. Experimental results demonstrate the effectiveness of the model.

Accurate segmentation of mining subsidence areas is crucial for monitoring illegal mining activities, mine remediation, and land use planning. Addressing the segmentation issue of mining subsidence in differential interferometric phase maps, we establish a differential interferometric phase map dataset specific to mining subsidence and propose an improved algorithm EMAGUNet++ based on the UNet architecture. By integrating attention gate and efficient multiscale attention mechanisms, attention weights are focused on relevant local features to enhance segmentation accuracy. Experimental results demonstrate that EMAGUNet++ achieves an intersection over the union index of 0.756 for mining subsidence segmentation, outperforming other classic segmentation models by 1.8% to 9.2%, validating the effectiveness of the model improvements. The improved model accurately segmented mining subsidence persisting over 6 years near Fugu County, and the segmentation polygon achieved an accuracy of 89.3%, showcasing excellent performance.

In recent years, transformers have been introduced into the field of remote sensing image change detection (CD) due to their excellent global context modeling capabilities. However, the global nature of the self-attention used by transformers is not sensitive to local high-frequency information, making it challenging to address complex CD problems. To address this issue, some methods have considered combining convolutional neural networks and transformers to jointly harvest local-global features. Nevertheless, these methods have not paid much attention to the interactions between the features extracted by the two components. Therefore, to address the challenges faced by existing CD methods in balancing local and global features, as well as their inadequacy in handling complex scenarios, we propose a frequency-driven transformer network (FDTNet) that improves self-attention and the overall architecture. In the overall framework, the network first extracts features to obtain primary and deep features and then utilizes the transformer encoder–decoder network to obtain context embeddings with spatiotemporal information from these primary features to guide the subsequent processing of deep features. In the transformer encoding part, we introduce a frequency-driven attention module, comprising low-frequency attention (LFA) branch, high-frequency attention (HFA) branch, and local window self-attention, where LFA captures global dependencies, HFA handles important high-frequency information, and local window self-attention supplements detailed local information loss. In the transformer decoding part, an interactive attention module is utilized to integrate context information from the transformer encoder into deep features. In addition, we propose an edge enhancement module and gate-controlled channel exchange operation, where the former enhances boundary features using the Sobel operator and the latter swaps channels to obtain richer perspective information. The experimental results show that FDTNet achieved an F1 score of 90.95% on LEVIR-CD, 82.70% on NJDS, and 79.84% on SYSU, outperforming several state-of-the-art CD methods.

Remote sensing images often contain a significant amount of clouds, which can result in substantial resource costs during transmission and storage. Cloud detection can reduce these costs. Although current cloud detection methods perform well in extracting large and thick clouds, there are still some issues, such as missed detection of small and thin clouds and false detection in non-cloud areas. Therefore, we propose a deep learning framework called DB-Net. It consists of three main modules: feature extraction module (FEM), cascaded feature enhancement module (CFEM), and feature fusion module (FFM). In the FEM, we leverage the advantages of both convolutional neural network and Transformer by utilizing two branches to reduce the loss of semantic information. To enhance the acquisition capability of multi-scale semantic information, in the CFEM, regular convolutions are replaced with deformable convolutions to adaptively capture cloud features of various sizes, and a cascaded structure is designed to enhance the interaction of information among different scales. Furthermore, to focus on small and thin cloud information and suppress non-cloud background information, we designed the FFM using attention mechanisms to enhance the target information in the features extracted by FEM and CFEM. Extensive experiments were conducted on the GF1-WHU dataset, and comparisons were made with mainstream cloud detection networks. The experimental results indicate that the proposed DB-Net method reduces cloud information omission, effectively focuses on thin clouds and small clouds, and improves overall cloud detection performance.

Modulation transfer function (MTF) is one among the key image quality indicators that characterizes the spatial response of an imaging system. There are various techniques suggested in the literature for quantification of MTF, for instance, slit-based method, point target-based method, image-structure- and fractal characteristic-based methods, and the slant-edge method. For high-resolution satellites, slant-edge-based method for on-orbit MTF estimation is the standard and widely used method due to its simplicity of implementation and practical feasibility. However, the edge-based method relies on accurate estimation of certain parameters failing, but the results may be misleading. We present the frequently occurred oversights and inadvertent miscalculations in MTF estimation using the slant-edge-based technique for high-resolution satellite sensors. We mainly address the two most common oversights in on-orbit MTF estimation process: (1) optimal edge width estimation and (2) precise empirical modeling of the slant edge. A rigorous sensitivity analysis suggests that oversights in any of these two factors can severely under or over estimate the MTF numbers. We establish accurate slant-edge-based MTF estimation procedure that can potentially avoid the aforementioned oversights.

To solve the problem of adhesion objects and data distribution deviations in few-shot scenarios, a synthetic aperture radar (SAR) object detection method based on meta-learning is proposed, which includes support feature guidance block and variational inference block. The former enhances the key features used for bounding box positioning in the query feature, so that the module can generate accurate proposals even in face of the adherent SAR objects. On this basis, to correct the deviation of the data distribution caused by the few-shot data, a variational inference block is constructed to map the supporting features to the class distribution in the hidden space. To fuse robust class-level features, meta-knowledge is used to calculate the distribution of the support feature classes of classes. The proposed algorithm uses a few-shot support set data to migrate priori knowledge to a class using the few-shot tasks and data double sampling. Moreover, a few-shot SAR object detection dataset is established to verify the effectiveness of the proposed method, and the experimental results show that our method has obvious advantages over the representative few-shot SAR object detection algorithms.

This paper proposes a novel laddering vision foundation model for change detection (CD) of remote sensing images. Current approaches have limitations in simultaneously extracting universal features and task-specific characteristics, and they cannot effectively integrate these features for detection tasks. The proposed model exploits both general features and task-specific characteristics for CD of remote sensing images. Specifically, task-agnostic characteristics are extracted from a pre-trained visual foundation model, which contains knowledge information of images. Then, the hierarchical transformer-based CD backbone is exploited to learn both long-range and local spatial information from remote sensing images. Furthermore, task-specific and universal features are integrated within the hierarchical network architecture, which can integrate heterogeneous feature maps and embedding tokens, addressing the differences in structure and content of different types of features. Three benchmark datasets are employed for comparative experiments, and experimental results certify the effectiveness and progressiveness in terms of CD of the investigated approach.

Accurate registration of point clouds acquired from different viewpoints can effectively reconstruct three-dimensional scenes, which plays a crucial role in industrial applications. Many learning-based methods rely on the extracted point cloud features to find the correspondence between two point clouds and thus achieve point cloud registration. However, repetitive geometric structures can make the local features of the learned point clouds highly similar, leading to the generation of incorrect correspondences. We propose the Two-Stage Graph Matching Point Cloud Registration Network (TSGM-Net). First, we design a dynamic graph-to-point (DGTP) module to learn the feature representation of the local graph of the point cloud to improve the recognition of local features. Then, edges are dynamically established by the Transformer and the introduced edge threshold λ, and the graph attention network is used to extract global features of the point cloud to consider the relationship among similar features in the topological structure. At the same time, scores are calculated from the three dimensions of the node itself, local and global, and summed for keypoint detection. Finally, we propose a two-stage graph matching method, where keypoints with highly similar features are divided into different point groups, and the correspondences of the point groups are established in the first-stage graph matching. The correspondence of points in corresponding point groups is established in the second-stage graph matching, which in turn reduces the impact of similar features on the accuracy of point cloud registration. Based on the benchmark dataset, we evaluate the performance of TSGM-Net in four different scenarios. Experimental results show that TSGM-Net can better handle point clouds with repetitive geometric structures and is competitive with state-of-the-art methods.

Hyperspectral and light detection and ranging (LiDAR) imaging instruments capture on-ground object information from diverse perspectives, reflecting spectral–spatial and elevation descriptions, respectively. Their complementary capturing feasibilities contribute to enhancing accurate landcover identification in multimodal data fusion tasks. However, their heterogeneous distributions always impede fusion and joint classification performance, leading to wrong classification phenomena. To solve this challenge, we proposed a deep multiscale cross-modal attention (DMSCA) network for hyperspectral and LiDAR data fusion and joint classification. Compared with existing methods, our primary motivation is to explore the intrinsic connection between these two specific remote sensing modalities and enhance their shared attributes through the implementation of various cross-modal attention mechanisms. The extracted modality features are cross-modally integrated and exchanged, thereby enhancing the overall consistency of the simulations. Specifically, these cross-modal attention mechanisms are capable of strengthening local considerable segments considering detailed hyperspectral and LiDAR geographical descriptions. The spatial-wise attention mechanism measures the contributions of neighboring samples to classification performance. The spectral-wise attention mechanism highlights the significant hyperspectral channels in terms of channel correlation. The elevation-wise attention mechanism highly connects the hyperspectral-related attention mechanisms to detailed LiDAR elevations for information fusion. Based on these mechanisms, an adaptive fusion and joint classification framework is constructed for balancing multimodal information. Multiple experiments are conducted on three widely used datasets to prove the effectiveness of DMSCA. Experimental results prove that DMSCA outperforms state-of-the-art techniques qualitatively and quantitatively.

The aim of remote sensing image scene recognition is to label a set of semantic categories based on their contents, and recognition for remote sensing images has a wide range of applications in many fields. However, it is a great challenge to extract category features with insufficiently labeled samples. We propose a Multi-scale Shift-window Cross-attention Vision Transformer (MSC-ViT) framework for remote sensing image scene recognition with limited data. Specifically, the proposed model is composed of three modules: a multi-scale feature extraction module, a shift-window transformer module, and a multi-scale cross-attention module. First, to enhance the efficiency of data utilization, we design a multi-scale module to fully extract the features of object information and spatial information contained in the image. The hierarchical transformer structure based on shifted windows, which are flexible at different scales, could match the computation of multi-scale features. The token fusion method based on the cross-attention mechanism fuses the features between multi-branch tokens and class tokens, which fully learn the information of the tokens and achieve better classification results. In addition, we integrate existing open-source datasets of remote sensing images and form a new dataset to better apply to the scene recognition task of remote sensing images with limited data. Our experimental results show that the proposed method achieves a great performance in scene classification of remote sensing images with limited data. The top-1 accuracy of the developed method is 79.84% with a 20% training ratio, 84.78% with a 40% training ratio, 89.79% with a 60% training ratio, and 91.43% with an 80% training ratio.

In recent years, graph convolutional network (GCN) has garnered increasing attention in hyperspectral image (HSI) classification. However, due to the excessive redundancy in message passing of GCN within dense graph structures, the superpixel adjacency matrices constructed by researchers tend to be sparse. This in turn leads to a scarcity of neighborhood information acquired by nodes. In addition, in complex HSI environments, the randomness of graph structures and training samples introduces interference, thereby increasing classification difficulty. To address these issues, we propose an HSI classification algorithm called robust dense graph structure based on GCN (DGGCN). The algorithm reduces computational burden by introducing a superpixel segmentation algorithm and constructs dense graph structures to enrich neighborhood information. Then, a dynamic fusion algorithm and a spatial and feature difference weighting algorithm are employed to reduce the confusion of message passing in GCN and decrease information redundancy, thus effectively addressing the aforementioned problems. Finally, we evaluate the DGGCN method on three public HSI datasets, demonstrating its superiority over other advanced classification methods in terms of overall accuracy, average accuracy, and kappa coefficient.

With the abilities of feature extraction and nonlinear representation, deep neural networks can remove spatial redundancy efficiently and perform well in terms of visible image compression. However, when it comes to multispectral images, it is necessary to consider both spatial and spectral redundancy. Based on this point, we propose an end-to-end feature domain residual coding network for multispectral image compression based on interspectral prediction. Specifically, a spatial-spectral feature extraction network and an interspectral prediction network are designed based on a pyramid structure, which can capture and fuse multi-scale features from coarse to fine. They make the best use of the spectral correlation to predict images accurately while combining with the feature domain residual coding network, which can further reduce the redundancy of spatial-spectral information. A single loss function jointly optimizes all modules in the network. There are lots of experiments with 8-band and 12-band multispectral image datasets. The experimental results demonstrate that the compression performance of the proposed method is superior to traditional compression methods (JPEG2000, 3D-SPIHT, Versatile Video Coding, PCA+JPEG2000) across evaluation indicators and even better than the advanced learned multispectral image compression algorithms.

The complexity of convolutional neural network architectures in hyperspectral image classification tasks results in long training times and high energy consumption, which severely limit the capabilities of edge computing equipment. To address the challenge of hyperspectral image classification in edge computing environments, we propose a spiking neural network (SNN) based on the hyperspectral classification method, which significantly reduces training time while maintaining high accuracy. Specifically, an approximate gradient algorithm is introduced to address the question of the non-differentiability of spike activity in spiking neurons. Subsequently, the method operates on channel and spatial dimensions, integrating a spatial–spectral attention mechanism to improve the capacity of the network to capture and simulate essential features. The proposed method is assessed using Pavia University, Indian Pines, and WHU-Hi hyperspectral datasets. The results indicate that the classification accuracy outperforms the models from spectral–spatial feature tokenization transformer (SSFTT), deep pyramidal residual networks (DPyResNet), small convolution and feature reuse (SC-FR) module, attention-based adaptive spectral–spatial kernel residual network (A2S2K-ResNet), and end-to-end spectral-spatial residual network (SSRN). It also reduces the training time by a factor of 7 for A2S2K-ResNet and 4 for SSRN. This algorithm presents an approach for deploying SNN models on edge devices, offering a fresh perspective for advancing hyperspectral image classification tasks.

The use of satellite remote sensing for rapid and accurate monitoring of water bodies can provide holistic and dynamic information on water resources, which is of great importance for the prevention and control of floods, droughts, and other disasters. Optical satellite remote sensing data can be affected by cloudy weather, which leads to the problem of large errors in water body identification. We use Sentinel-1 data to construct a synthetic aperture radar water body dataset and propose MAG-Net, a network model that fuses multi-level and multi-scale information. The model structure consists of a multi-scale residual block as the backbone network, which is designed to enhance the ability to capture different scale features of water bodies. The ability to capture detailed information is also enhanced by introducing a global information fusion block. Finally, the decoder is designed as a full-size hop-connected structure, which improves the accuracy of narrow water body recognition by combining feature maps at different scales. Experiments show that the extraction accuracy of the proposed model on Sentinel-1 data and the generalization ability on GaoFen-3 (GF-3) data outperform the existing water body extraction models. MAG-Net achieved impressive results on the Sentinel-1 dataset, with an accuracy of 97.09%, recall of 97.50%, and F1-score of 97.25%. These scores represent significant improvements of 1.89%, 2.94%, and 2.43%, respectively, compared with the worst-performing models. In addition, MAG-Net demonstrated a high accuracy rate of 95.63%, a recall rate of 94.46%, and an F1-score of 94.18% on the GF-3 verification set.

In the field of remote sensing, there are numerous challenges associated with multi-label scene classification, such as the firm correlation among labels, the presence of small-scale noisy objects or areas, cluttered backgrounds, diverse image categories in datasets, differences within classes and similarities among classes, and unbalanced class weights. Moreover, the limited representation ability of convolutional neural networks (CNNs) makes multi-label scene classification a complex endeavor. We address the CNN model's limited representation ability and aim to enhance its performance through the proposed framework. Current research has demonstrated that channel spatial attention modules can boost the representation power of CNNs. However, these modules are not yet widely utilized in the multi-label scene domain. Therefore, we have incorporated state-of-the-art channel–spatial attention modules into MobileNet_v1 to improve its representation ability for multi-label remote sensing scene classification. The proposed method employs two-level feature extraction with double channel–spatial attention residual blocks. We tested it on the UC, AID, and DFC15 multi-label datasets and evaluated its performance using various metrics. The results show that our method improves the representation power of MobileNet_v1 by 3% to 7% in terms of F1- and F2-scores, using channel spatial attention and residual connections on the UC and AID datasets. Our approach also surpasses MobileNet_v3, achieving a 3% to 8% increase in F1 and a 3% to 6% increase in F2-scores on the UC and AID datasets. On the DFC15 dataset, a significant improvement is made over the MobileNet family. Compared with the identified state-of-the-art methods, our approach achieves superior/competitive outcomes while employing fewer training parameters/floating point operations. Therefore, the proposed approach is a promising solution for multi-label remote sensing scene classification and can help overcome the challenges posed by this task in the remote sensing domain.

In light of recent advances in deep learning and high-resolution remote sensing imaging technology, there has been a growing adoption of remote sensing ship classification models that are based on deep learning methodologies. However, the efficiency of remote sensing ship classification models is affected by complex backgrounds, shooting conditions, high inter-class similarity of ship targets, and sample diversity. To tackle the challenges above, we propose a multi-scale attention-based adaptive feature fusion (AFF) network for fine-grained ship classification in remote sensing scenarios to improve the fine-grained classification ability of the model from local details. First, using the idea of information complementarity, the multi-scale feature interaction module is constructed in the multi-scale attention module. It employs bidirectional feature interaction paths to concurrently capture intricate details within both deep and shallow ship features, enhancing the interplay among different levels of information. Second, the hybrid attention module is part of the multi-scale attention module. It is designed to enhance the cross-dimensional interaction of spatial domain and channel domain information to amplify the importance of crucial feature regions and feature channels. This allows the network to pay more attention to specific areas and extract distinctive features. Finally, the AFF module is designed to automatically calibrate and fuse different levels of saliency features to obtain features with more fine-grained discrimination for model classification. In this approach, these modules synergistically collaborate and mutually reinforce each other, ultimately increasing the accuracy of ship classification tasks. We evaluated our method on three large-scale fine-grained classification benchmarks; the experimental results show that the proposed method had better fine-grained classification than other methods.

The acceleration of urbanization has increasingly exacerbated air pollution in Northwest China. However, existing studies have relatively few analyses of PM2.5 concentrations in response to land-use changes. This study quantitatively evaluated the impact of land-use changes on PM2.5 concentrations in Urumqi (2014 to 2023) using remote sensing techniques and machine learning methods. The MCD19-A2 aerosol optical depth (AOD) product, with gaps filled using a singular spectrum analysis algorithm (99.63% AOD coverage), was used to predict PM2.5 concentrations based on the light gradient boosting machine method (10-CV R2=0.93, root mean square error=17.98 μg/m3). The spatial correlation between land-use changes and PM2.5 concentrations showed that PM2.5 concentrations were highest in central urban areas but decreased by an average of 27.41 μg/m3 over the decade. Land-use type transitions (barren-grassland, grassland-barren, and grassland-cropland) were significantly negatively correlated with PM2.5, indicating these changes reduced aerosol concentrations during the research period in Urumqi. The reaction of dynamic PM2.5 to land-use and land-cover changes showed a local overlap but was not entirely consistent, as reflected by the geographically weighted regression model. Geodetector quantified the contribution of land-use change to PM2.5 reduction, particularly barren-grassland conversion, which notably reduced PM2.5 (contribution coefficient = 0.161), highlighting the importance of protecting vegetated areas for PM2.5 control in Urumqi. These findings clarify the impact of land-use change on PM2.5, supporting improvements in land management and atmospheric control strategies for sustainable development in Urumqi.

Nitrogen and phosphorus are essential elements of plants, which play important roles in representing plant growth, physiological function regulation, fruit harvest, etc. Hyperspectral technology provides a nondestructive, rapid, highly accurate, and cost-efficient method for plant leaf nutrient content estimation. There are very limited studies on nutrient diagnosis of Camellia oleifera leaves using hyperspectral technology. In this work, 160 Camellia oleifera samples were used. Hyperspectral data were obtained using a full-band spectrometer. On the basis of preprocessing, the spectral response characteristics of leaf nitrogen content (LNC) and leaf phosphorus content (LPC) were revealed by comparing different combinations of spectral indices, and the spectral variables were further selected. The optimal LNC and LPC estimation models based on three machine learning algorithms [i.e., support vector machine (SVM), random forest (RF), and back propagation neural network (BPNN)] were constructed. The results showed that the spectral sensitive regions of leaf nitrogen and phosphorus content were mainly reflected in green band, followed by red band and the long-wave direction of short-wave infrared band. Savitzky–Golay first derivative (SGFD) pretreatment method was generally better than multiplicative scatter correction. The maximum correlation coefficients of the absolute values of LNC, LPC, and spectral transformation features were 0.56 and 0.49. The optimal LNC and LPC models were both SGFD-TBNDSI-BPNN, with R2 of 0.81 and 0.79, and RMSEP of 0.55 and 0.06 g/kg, respectively. The research results can provide a reliable theoretical basis for large-scale optical remote sensing monitoring of nutrient content for Camellia oleifera.

Rice is the most important food crop in Taiwan. Early information on rice-growing conditions is thus vital for estimating rice production to guarantee national food security and grain exports. The rice-harvested area is conventionally inspected twice a year by costly interpretation of aerial photographs and intensive labor-field surveys. However, such methods of rice monitoring are inadequate for providing the government with timely information on rice-cultivated conditions. This study aims to use time series of Sentinel-1 synthetic aperture radar and Sentinel-2 multispectral data to develop a machine-learning approach for the early prediction of rice-growing fields in Taiwan. An object-based random forest (OBRF) was developed to process remotely sensed data for rice-cropping seasons from 2018 to 2021. The prediction results compared with the reference data showed that rice-growing fields could be accurately predicted before harvest, about three months for the first crop and two months for the second crop. The F-score and Kappa coefficient values achieved for the first crop were 0.87 and 0.85, and those for the second crop were 0.72 and 0.71, respectively. These findings were reaffirmed by close agreement between the official statistics and the rice area estimated from the satellite data, with the correlation coefficient of determination (R2) value greater than 0.96. A large portion of the first crop’s rice areas was abandoned or converted to upland crop cultivation in the second crop, which was confirmed by a visual interpretation of Landsat images and official statistics. Ultimately, this study proved the efficacy of using Sentinel-1/2 images and OBRF for the early prediction of rice-cultivated fields in Taiwan. Quantitative and geographical information produced from such methods was essential for the early estimation of rice production to nationally address food security concerns.

Hyperspectral band selection plays a crucial role in reducing the dimensionality of hyperspectral data and enhancing the efficiency of subsequent analysis. However, most methods measure bands using Euclidean distance without considering its limitations in high-dimensional data. In addition, researchers usually select bands based solely on information entropy, without consideration of the impact of noise. To address these challenges, this work introduces a noise-robust hyperspectral band selection model dubbed SSIM-MEMN. The proposed model leverages the structural similarity index (SSIM) to measure the similarity between hyperspectral bands. Additionally, a sorting strategy is devised to identify a representative subset of bands. Specifically, this ranking strategy incorporates both information entropy and noise level to assign scores to individual bands. Consequently, the information pertaining to ground objects is captured with greater precision, leading to enhanced classification accuracy. Extensive experiments were performed to prove the excellent performance of SSIM-MEMN in different sizes of the remaining spectral band subset, and the classification results show that this method is sufficiently robust on different public hyperspectral datasets. In brief, the SSIM-MEMN model provides an effective band selection method for the field of remote sensing image processing and analysis.

Convolution neural network (CNN) and transformer-based methods have achieved significant results in hyperspectral image (HSI) classification. These works tend to improve the effectiveness of discrimination feature extraction exclusively in the image domain. However, the excellent feature representation ability in the frequency domain is rarely utilized. Hence, in this paper, a lightweight dual-domain token learning method based on CNN-transformer, aiming to effectively utilize information from both image and frequency domains, is proposed. Specifically, first, we enhance the multi-head self-attention mechanism in the image domain to enhance the mining of spectral correlations across the adjacent spectrum. Second, considering that a single value in the frequency domain has an impact on the values of the entire frequency domain, Fourier token learning is designed to ensure the global information of the frequency domain. Finally, the lightweight network achieves state-of-the-art classification performance with an acceptable computational cost.

In recent years, the utilization of hyperspectral sensors for remote sensing has marked a profound advancement due to the success of machine learning techniques. Nevertheless, difficulties still exist, especially in locations with shadows. The heterogeneity in spectral data due to different shadow origins, such as different types of clouds and different building designs, poses a significant obstacle to the advancement of shadow-aware classification algorithms. Furthermore, precisely labeling the underlying structures in shadowed areas is a very cumbersome effort. We present a loss function-based strategy based on generative adversarial networks to address this problem. Using the context of correlated samples, our loss function combines unpaired matchings and transitive style modifications via the fusion of contrastive learning, dual learning, cycle consistency, and curriculum learning algorithms. Our work transforms the non-shadowed training instances into the shadowed counterparts for use as synthetic training samples, as opposed to the conventional method of correcting shadowed pixels to their non-shadowed counterparts. We propose learning this transformation model with unpaired data samples, which is particularly advantageous compared with the collection process of the same samples with and without shadow. Synthetic samples for shadow-obscured regions can be produced when this method is used, and these samples improve the model’s performance in classification tasks. Rigorously tested through a combination of qualitative and quantitative evaluations, the introduced data augmentation technique improves the performance of terrain classification models, especially with limited data samples.

Ocean temperature and salinity are key variables for studying the impact of the ocean on climate change. Currently, satellite remote sensing technology provides extensive sea surface information. Utilizing these data to accurately invert the subsurface thermohaline structure is crucial for understanding oceanic dynamic processes. Therefore, we introduce a method for inverting the ocean subsurface thermohaline structure based on spatiotemporal feature fusion with a dual attention mechanism. Using sea surface data such as sea surface height, sea surface temperature, sea surface salinity, and sea surface wind speed, the subsurface thermohaline structure of the ocean is inverted. New variables approximating subsurface temperature and salinity are generated from existing variables to utilize vertical distribution characteristics. In North Pacific experiments, the proposed model is comprehensively evaluated from multiple perspectives. The results demonstrate that the model effectively inverts ocean subsurface temperatures and salinity. The newly introduced variables are highly correlated with subsurface parameters, significantly enhancing the model’s performance. The overall R2 and root mean square error of subsurface temperature are 0.985 and 0.394°C, respectively, whereas those of subsurface salinity are 0.985 and 0.052 psu, respectively, throughout the year 2021. Compared with the existing models, the proposed model shows superior accuracy, stability, and generalizability.