Fractures as critical discontinuous structural planes in rock masses, directly govern their stability and serve as the core controlling factor in rock mechanics engineering. Existing deep learning models for fracture extraction face persistent challenges, including imbalanced integration of deep and shallow features, limited suppression of background noise, inadequate multi-scale feature representation, and large parameter sizes—making it difficult to strike a balance between detection accuracy and deployment efficiency. Focusing on the Wanshanshan quarry in Yunnan, this study first constructs a high-precision digital model using close-range photogrammetry and 3D real-scene reconstruction. A lightweight yet high-accuracy intelligent detection method, termed MSEM-Deeplabv3+, is then proposed for rock mass fracture extraction. The model adopts lightweight MobileNetV2 as the backbone network, incorporating inverted residual modules and depthwise separable convolutions, resulting in a parameter size of only 6.02 MB and FLOPs of 30.170 G—substantially reducing computational overhead. Furthermore, the proposed MAGF (Multi-Scale Attention Gated Fusion) and SCSA (Spatial-Channel Synergistic Attention) modules are integrated to enhance the representation of fracture details and semantic consistency while effectively suppressing multi-source and multi-scale background interference. Experimental results demonstrate that the proposed model achieves an mPA of 89.69%, mIoU of 83.71%, F1-Score of 90.41%, and Kappa coefficient of 80.81%, outperforming the classic Deeplabv3+ model by 5.81%, 6.18%, 4.53%, and 9.2%, respectively. It also significantly surpasses benchmark models such as U-Net and HRNet. The method accurately captures fine and continuous fracture details, preserves the spatial distribution of long-range continuous fractures, and maintains robust performance on the CFD cross-scene dataset, showcasing strong adaptability and generalization capability. This approach effectively mitigates the risks associated with manual high-altitude inspections and provides a lightweight, high-precision, non-contact intelligent solution for fracture detection in high-steep rock slopes.

To address the challenges of scarce exposed soils, strong surface heterogeneity, and insufficient synergistic monitoring using multi-source remote sensing in high-altitude Pb-Zn mining areas, this study used laboratory spectra as the reference and developed a workflow consisting of direct standardization (DS)-based correction of GF-5A bare-soil imagery, spectral enhancement, Boruta-based band selection, and XGBoost inversion. A parallel comparison with Sentinel-2B (S2B) was also conducted. By integrating SHapley Additive exPlanations (SHAP)-based feature attribution, Bootstrap uncertainty analysis, and the geographic detector to identify the sources of heavy metal contamination, an integrated analytical framework of 'inversion - reliability-source' was established. The results showed that DS tended to shift the spectra of-different land-cover types towards the 'soil spectral space' as a whole, while derivative transformation exhibited superior performance in enhancing heavy metal-sensitive bands. GF-5A provided more complete coverage over key diagnostic spectral regions, particularly within 2200-2450 nm, and achieved test-set R2 values of 0.694, 0.720, and 0.722 for Zn, Pb, and Ni, respectively, outperforming S2B overall. Although S2B yielded lower quantitative accuracy, it was more advantageous for delineating the fine details of pollution boundaries. Distance from the mining area was the dominant controlling factor for Zn and Pb contamination, whereas Ni was more strongly controlled by parent material background, and GF-5A exhibited lower overall uncertainty. This method is applicable to the remote sensing monitoring of heavy metals in bare soils of complex high-altitude mining areas and provides a basis for the integration of GF-5A and S2B.

Digital elevation models (DEMs) and vegetation cover (FVC) are critical foundational data for ecological and environmental monitoring and land-use management; however, the application of domestically produced commercial high-resolution optical satellite stereo imagery for high-precision DEM construction and vegetation parameter inversion in areas with complex topography still requires further exploration. This study focuses on the hilly area in the northeastern part of Kunming City, Yunnan Province, and utilizes tri-view stereo imagery from the Beijing-3 (BJ3-N2) satellite to conduct research on DEM reconstruction, topographic correction, image fusion, and vegetation cover inversion. DEM products were generated by matching multiple sets of panchromatic and multispectral stereo imagery, and their accuracy was verified using ICESat-2 ATL08 laser altimetry data. The results indicate that DEMs generated from panchromatic imagery exhibit higher accuracy than those from multispectral imagery; among them, the PAN-0103 DEM demonstrated the best overall performance, with a global RMSE of 3.42 m. Topographic slope and land cover type were found to have significant effects on DEM accuracy. A terrain-constrained Gram–Schmidt fusion model was constructed based on the optimal DEM to generate the fused image MP-01, which had a spectral angle (SAM) of 3.2° and a band correlation coefficient (CC) of 0.92. Further, by integrating field RTK plot data, an improved pixel binary model was constructed to perform FVC inversion. The results showed that the Mahalanobis distance classification method yielded the best classification performance, with a model inversion RMSE of 0.074—a 41.7% reduction compared to the traditional fixed-end element model. The study demonstrates that the high-resolution stereo imagery from Beijing-3 can effectively support the refined construction of DEMs and quantitative monitoring of vegetation cover in complex hilly areas and can provide a technical reference for the application of domestically produced commercial remote sensing satellites in the field of ecological monitoring.

Multimodal change detection (CD), owing to its ability to flexibly adapt to data acquired from different types of sensors, has become an important research direction in the field of remote sensing. However, existing methods generally lack feature representations with sufficient generalization capacity, leading to pronounced performance degradation when applied across diverse scenes. To address this limitation, we propose a novel FastSAM-guided adversarial learning framework (FastSAM-GAL) for unsupervised multimodal CD. The proposed framework effectively mitigates cross-modal discrepancies via adversarial learning, and constructs a dual-branch architecture comprising a FastSAM encoder and a remote sensing encoder, thereby fully exploiting the complementary information between generic visual semantic features and domain-specific remote sensing representations to substantially enhance cross-scene generalization capability. Furthermore, a progressive multi-scale perception adaptor (PMSPA) and a spatial-frequency collaborative edge augmentation (SFCEA) module are specifically designed to further improve FastSAM-GAL’s ability to perceive objects at different scales and capture change boundaries. Finally, comparative experiments on five real-world multimodal CD datasets against twelve state-of-the-art methods demonstrate that the proposed FastSAM-GAL maintains stable and reliable detection performance across multimodal remote sensing data with diverse scenes, with its κ metric significantly outperforming all competing methods.

Multimodal change detection (CD) effectively improves the flexibility and timeliness of emergency response by exploiting complementary information from different sensors. However, significant differences in imaging modalities lead to a lack of comparability across multimodal images, posing a severe challenge to change identification. Existing studies have alleviated this issue to some extent by introducing generative adversarial networks (GANs) to align multimodal images into a unified image domain. Nevertheless, these methods usually adopt a coupled encoding of semantic content and visual style, which easily leads to mutual interference and semantic drift between the two, thereby degrading the quality of modality alignment and the subsequent CD accuracy. Moreover, insufficient hierarchical feature modeling and the lack of effective background noise suppression further constrain their performance in practical applications. To this end, we propose a novel content and style decoupled GAN (CSD-GAN) for unsupervised multimodal CD. This method first employs a carefully designed explicit decoupling strategy to separate content and style representations, effectively suppressing semantic drift and significantly enhancing modality alignment quality. Subsequently, a spatial and semantic feature refinement mechanism is tailored to the distinct characteristics of shallow and deep features, thereby providing richer feature support for both modality alignment and CD. Finally, a dual-domain sparse difference perception (DSDP) module is developed to suppress background noise via cosine similarity and masked sparsity strategies, while exploiting cross-domain interactions to mine multimodal complementary information, substantially enhancing the model’s responsiveness to true changes. Systematic experiments conducted on five representative datasets demonstrate that CSD-GAN achieves competitive detection performance compared with existing methods.

Urban building rooftops represent a high-potential source of solar PV power. However, prevalent estimation methods often overlook shading from surrounding structures, causing inaccuracies. This study proposes a method to estimate rooftop PV potential by integrating open-source multi-modal spatiotemporal data. By calculating solar radiation on building rooftops under both planar and 3D conditions, we derived a solar occlusion correction ratio. Using the morphological characteristics of buildings in Kunming, we established a random forest regression model to explore the impact of shading on rooftop solar radiation. After determining the correction ratio, our method achieves second-scale estimation of solar PV potential from planar conditions to those considering 3D morphological shading, reducing the global relative error from 3.66% to 0.04% (DSM-based benchmarks; 3.66% : FLAT-based; 0.04% : our ratio-based). This research includes the creation of solar radiation maps for the main areas in Kunming, analysis of PV generation potential, energy balance, and carbon emission reduction benefits, with Wuhan further used as a test city to validate the model. The findings provide valuable insights for future policy-making on PV installations, energy consumption, and grid emissions.

Change detection (CD), as a core task in remote sensing image analysis, plays a crucial role in disaster monitoring, urban planning, and environmental assessment. Compared to traditional CD methods, multimodal CD breaks the dependency limitations on homogeneous sensor data, demonstrating higher timeliness and flexibility in disaster emergency response. However, significant imaging differences between different sensors lead to a lack of comparability in multimodal images, severely hindering their practical application. Existing multimodal CD research has effectively alleviated this issue by introducing generative adversarial networks (GANs) to transform multimodal images into the same image domain. Nevertheless, these methods generally overlook the negative impact of intrinsic change regions on image transformation quality and the enhancement of CD performance through complementary features between different modalities, resulting in limited change recognition accuracy. To address these shortcomings, we propose a novel dual-domain contrastive learning GAN (DDCL-GAN) for unsupervised multimodal CD. This network effectively suppresses interference from change regions through a carefully designed non-local spatial correlation patch contrastive learning strategy, thereby ensuring semantic consistency at the content level between transformed and original images. Additionally, we develop a multi-scale dual residual module and a semantic cross-flow alignment module to enhance the model's ability to express multi-scale semantic information and fine-grained features, respectively. Finally, we construct a dual-domain difference markov random field (DDMRF) detection model, which effectively improves detection performance by simultaneously considering dual-domain image information. To validate the effectiveness of the proposed method, we conducted comparative experiments with twelve mainstream unsupervised multimodal CD methods on nine typical datasets. Experimental results demonstrate that the proposed method achieves optimal performance across multiple average quantitative metrics, particularly in terms of mean IoU and kappa, showing improvements of at least 2.9% and 2.3%, respectively, compared to comparative methods.

Deep-learning-based methods have achieved remarkable success in hyperspectral image (HSI) classification tasks due to their promising ability. However, the high dimensionality and spectral-spatial correlations of HSIs usually lead to information redundancy and feature entanglement, limiting the classification performance. To address these issues, we propose a novel high-low frequency interaction Mamba network, called HL-Mamba, which achieves effective decoupling and interaction between global structures and edge details of HSIs in the frequency domain, thereby improving spectral-spatial representation for HSI classification. Specifically, a high-low frequency decomposition Mamba module is designed to decompose the HSI into low-frequency structural and high-frequency edge detail components, which allows the model to learn global structures and fine-grained details, enhancing classification performance. By employing two parallel Mamba branches to model long-range dependencies across different frequency components, the network achieves efficient global modeling while mitigating information redundancy. Furthermore, a cross-frequency interaction module is designed to establish complementary information flow between high- and low-frequency features through a dynamic attention mechanism. In this way, low-frequency structural features guide the aggregation of high-frequency details, whereas high-frequency textures refine global structural representations, yielding more discriminative spectral-spatial features for HSI classification. In addition, a frequency alignment loss is designed to enhance the consistency and complementarity between high- and low-frequency features, further improving classification performance. Extensive experiments on four public benchmark datasets (i.e., Indian Pines, Pavia University, WHU-Hi-HanChuan, and Houston datasets) demonstrate that the proposed HL-Mamba significantly outperforms eight comparison methods, achieving an overall accuracy of 94.07%, 93.82%, 95.28%, and 87.32%, respectively. Ablation studies further verify the effectiveness of core component within the network.

Multimodal change detection (CD) breaks the constraint that bi-temporal images must originate from the same sensor, significantly enhancing the flexibility and timeliness of disaster emergency response. However, inherent imaging differences between sensors result in multimodal images that are not directly comparable, making conventional comparison-based change extraction challenging. Existing studies have alleviated this cross-modal incomparability by employing image translation (IT) to transform multimodal images into a unified image domain. Nevertheless, these studies generally treat IT and CD as two independent sequential processes, overlooking the intrinsic correlation between them. This results in the IT process lacking guidance from change semantics, hindering effective modeling of actual change regions and thereby degrading the accuracy of CD. To address this, we propose an end-to-end multitask generative adversarial network (MTGAN) for unsupervised multimodal CD. The method innovatively integrates IT and CD into a unified learning framework, achieving bidirectional interaction between the two through a collaborative optimization strategy. The IT task provides more reliable cross-modal comparable features for CD, while the CD task reciprocally guides IT to focus on change-sensitive regions, thereby improving the overall detection performance. Furthermore, we design a high-frequency guided difference aggregation module and a multi-scale difference affinity feature refinement mechanism to enhance MTGAN's ability to capture fine-grained changes and suppress background noise, thereby further improving detection reliability. Comparative experiments on five representative multimodal datasets against twelve state-of-the-art methods demonstrate that MTGAN can identify more true changes while effectively reducing false detections, with key quantitative metrics F1 and kappa significantly outperforming the compared methods.

Due to the inherent limitations of both hyperspectral and multispectral imagery, balancing high spatial resolution with high spectral fidelity has become one of the fundamental challenges in remote sensing image processing. A prevailing strategy is to fuse these two types of data to reconstruct images that jointly preserve their respective advantages. However, existing reconstruction approaches still suffer from complex coupling between spatial and spectral information, and limited feature extraction capabilities. To address these issues, this study proposes PMSwinNet (Pyramid Multi-scale Swin Transformer Network), a novel architecture that integrates pyramid-based feature enhancement with Transformer mechanisms. The PMSwinNet incorporates multi-scale pyramid feature fusion and window-based self-attention. Through a progressive multi-stage design and three complementary components—feature extraction and reconstruction modules—the Transformer branch leverages window partitioning and shifting operations to capture long-range spatial dependencies and local contextual cues, while the pyramid features extract both global and local information across multiple spatial scales. In addition, a high-frequency branch is introduced, which employs lightweight convolutions to enhance edges, textures, and other high-frequency details, effectively suppressing blurring and artifacts during reconstruction. Experimental evaluations on multiple public hyperspectral datasets demonstrate that the PMSwinNet outperforms state-of-the-art methods, particularly in terms of detail preservation, spectral distortion suppression, and robustness.

Advancements in remote sensing and spatial information technologies have made the digitalization of geological relics a vital method for enriching geoheritage databases. However, complex topographies and intricate textures complicate the fine-detail digitization process. To accurately digitize geological relics, this study conducts qualitative and quantitative analysis of the geometric and textural information from 13 digital models. It proposes using smartphones to capture semi-circular images at geological sites, followed by digitization using ISfM-PMVS, and then interpretation and documentation of the data. In the Middle Red Beds of Central Yunnan, fault profiles were created, measuring fault length, spacing, dip direction, and dip angle. In the Thousand Turtle Mountains, digital models mapped crack morphology, with point clouds achieving a detection accuracy of 0.963. Although edge distortions exist, ISfM-PMVS produces high-fidelity 3D models, offering a fast, low-cost solution for reliable field-based data collection, supporting intelligent interpretation, digital recording, research, heritage exhibitions, and conservation.

Generative adversarial networks (GANs) possess powerful image translation capabilities. They can transform images acquired from different sensors into a unified domain, effectively mitigating the incomparability problem caused by imaging discrepancies in multimodal remote sensing change detection (CD). However, existing approaches predominantly emphasize domain unification while neglecting the loss of fine-grained features inherent in the translation process, consequently compromising both image translation quality and CD accuracy. To overcome these limitations, we propose a novel texture and structure interaction guided GAN (TSIG-GAN). This network establishes interactive guidance between image texture and structural features through a carefully designed dual-stream cross encoder-decoder architecture, enabling in-depth mining of fine-grained features and significantly improving the fidelity of translated images. Furthermore, to address the spatial scale diversity and complexity of remote sensing images, we develop a multi-scale adaptive feature pyramid (MAFP) module and a contextual semantic interaction guidance (CSIG) mechanism, aiming to further strengthen the model's robust representation of fine-grained features across multiple scales and complex scenes. Specifically, the MAFP module effectively captures spatial details of targets at different resolutions by dynamically integrating multi-scale features, thereby preventing detail loss in small objects due to scale discrepancies. The CSIG mechanism achieves deep interaction between texture and structural features at the contextual semantic level, further promoting their mutual cooperation, thereby enhancing the consistency of fine-grained features representation and semantic integrity in complex scenes. Finally, the translated fine-grained images are fed into a custom CD network to extract changes. To evaluate the effectiveness of the proposed method, we conducted systematic experiments on five representative real-world datasets and performed comparative analysis with sixteen state-of-the-art multimodal CD methods. The experimental results demonstrate that TSIG-GAN achieves significant improvements in both image translation and CD performance, exhibiting superior fine-grained restoration capability and change identification capability.

In recent years, research on image translation for multimodal remote sensing imagery in change detection (CD) has demonstrated that converting images from different sensors into a common image domain can effectively address the incomparability issues arising from imaging differences, thereby enabling traditional CD models to extract change information from diverse data sources. However, most existing studies generally overlook global contextual information during the image conversion process, resulting in an inability to fully capture the overall semantic structure of the image. This may lead to semantic confusion of similar features, thereby affecting the accuracy of CD. To address this, we propose a pseudo-siamese generative adversarial network (PS-GAN) that simultaneously considers both local and global information. Unlike conventional GANs that focus solely on local features, PS-GAN utilizes a two-branch structure in the encoder phase to separately extract global and local information. These features are then effectively fused through a carefully designed adaptive multi-scale feature fusion module, ensuring that the translated images are texturally clear and structurally intact, thus reducing the occurrence of pseudo-changes in CD. In addition, we adopt U2Net+ as the CD model and develop a gated feature modulation mechanism and an enhanced squeeze-and-excitation module to reduce the interference of redundant features while enhancing the semantic representation of change features, further improving CD accuracy. Finally, we conducted comparative experiments on four real-world datasets against eleven state-of-the-art multimodal CD methods. The experimental results clearly demonstrate the superiori of the proposed method, achieving more accurate detection outcomes.

To address the decline in self-consistency and limited spatial adaptability of traditional interpolation methods in complex terrain, this study proposes a terrain-constrained Triangulated Irregular Network (TIN) interpolation method based on UAV point clouds. The method was tested in the southern margin of the Lufeng Dinosaur National Geopark, Yunnan Province, using ground points at different sampling densities (90%, 70%, 50%, 30%, and 10%), and compared with Spline, Kriging, ANUDEM, and IDW methods. Results show that the proposed method maintains the lowest RMSE and MAE across all densities, demonstrating higher stability and self-consistency and better preserving terrain undulations. This provides technical support for high-precision DEM reconstruction from UAV point clouds in complex terrain.

Carbonate-hosted clay-type lithium deposits have emerged as strategic resources critical to the global energy transition, yet their exploration faces the dual challenges of technical complexity and environmental sustainability. Traditional methods often entail extensive land disruption, particularly in ecologically sensitive ecosystems where vegetation coverage and weathered layers hinder mineral detection. This study presents a case study of the San Dan lithium deposit in central Yunnan, where we propose a hierarchical anomaly extraction and multidimensional weighted comprehensive analysis. This comprehensive method integrates multi-source data from GF-3 QPSI SAR, GF-5B hyperspectral, and Landsat-8 OLI datasets and is structured around two core parts, as follows: (1) Hierarchical Anomaly Extraction: Utilizing principal component analysis, this part extracts hydroxyl and iron-stained alteration anomalies. It further employs the spectral hourglass technique for the precise identification of lithium-rich minerals, such as montmorillonite and illite. Additionally, concealed structures are extracted using azimuth filtering and structural detection in radar remote sensing. (2) Multidimensional Weighted Comprehensive Analysis: This module applies reclassification, kernel density analysis, and normalization preprocessing to five informational layers—hydroxyl, iron staining, minerals, lithology, and structure. Dynamic weighting, informed by expert experience and experimental adjustments using the weighted weight-of-evidence method, delineates graded target areas. Three priority target areas were identified, with field validation conducted in the most promising area revealing Li2O contents ranging from 0.10% to 0.22%. This technical system, through the collaborative interpretation of multi-source data and quantitative decision-making processes, provides robust support for exploring carbonate-clay-type lithium deposits in central Yunnan. By promoting efficient, data-driven exploration and minimizing environmental disruption, it ensures that lithium extraction meets the growing demand while preserving ecological integrity, setting a benchmark for the sustainable exploration of clay-type lithium deposits worldwide.

Metallic lithium is an important strategic mineral and a key element in new energy technologies. Carbonate clay-type lithium deposits are a promising new frontier in lithium resource development. Rapidly determining the occurrence and enrichment mechanisms of lithium is a critical issue that must be addressed for evaluation, prediction, and development. To explore the role of hyperspectral technology in the prospecting of clay-type lithium deposits and achieve efficient and environmentally friendly exploration, this study focuses on the Sandan Town region. Three core samples were collected for Li2O content and ground hyperspectral data. The study employs Fractional Order Differentiation (FOD) methods and correlation coefficients for spectral data processing and feature selection. Sensitive features of the raw hyperspectral reflectance data and their fractional-order differentiated spectra were identified. Additionally, a hyperspectral estimation model for Li2O content in carbonate clay-type lithium ores was developed using a genetic algorithm to optimize the XGBoost model. The findings reveal the following: (1) A significant inverse correlation between Li2O content and spectral reflectance, highlighted by pronounced absorption features in the 2.15-2.32 mu m range. (2) Fractional-order differentiation markedly improved the detection of subtle spectral differences and absorption features of the original spectra. Utilizing a significance level of p = 0.01, the correlation coefficient method efficiently extracted characteristic bands, thereby reducing the data dimensionality. (3) The GA-XGBoost model, enhanced through genetic algorithm optimization, exhibited superior performance. With a 1.5 order differential transformation, this model achieved an R2 of 0.96 between predicted and actual Li2O content values, indicating a high level of accuracy. This investigation validates the effectiveness of the FOD and GA-XGBoost approaches in the hyperspectral estimation of Li2O content within exploration core samples. It also delineates the optimal differential order and hyperparameter configurations necessary for accurate Li2O content inversion, presenting a novel methodological approach for the rapid evaluation of lithium ore resources.

This study presents a novel approach to identifying the spatial distribution of geological outcrops using remote sensing, with the Lufeng Dinosaur Valley in China as the focal area. We integrate GF-5 hyperspectral imagery with the automated morphological endmember extraction–pure pixel index (AMEE–PPI) algorithm to improve endmember extraction accuracy. Additionally, we apply a suite of spectral transformation techniques—log(1/R), continuum removal (CR), and first-order derivative (FD)—combined with feature band selection algorithms including the successive projection algorithm (SPA), iteratively retained informative variables (IRIVs), and competitive adaptive reweighted sampling (CARS), effectively reducing data dimensionality and enhancing model performance. For classification, we evaluate three machine learning models: random forest (RF), k-nearest neighbor (KNN), and support vector machine (SVM). Among them, the SVM model demonstrates superior accuracy and stability. The proposed original spectral–SPA–SVM (OS–SPA–SVM) model achieves high precision in identifying outcrops, offering practical value for geological exploration and assessment. Our findings contribute theoretically and methodologically to the application of hyperspectral remote sensing and machine learning in geological research, providing insights into the unique geological characteristics of the Lufeng Dinosaur Valley.

Hyperspectral reflectance provides a pathway for estimating soil iron oxide and heavy metal zinc(Zn) content. The method and process for retrieving soil physicochemical properties from soil reflectance spectra mainly include spectral preprocessing-feature wavelength selection-machine learning modeling. To find the optimal model combination, this study first applies conventional spectral transformations (Continuum Removal, CR; Standard Normal Variate, SNV; First Derivative, FD and Second Derivative, SD) to the original soil spectra, then uses competitive adaptive reweighted sampling (CARS) and the Boruta algorithm to select sensitive bands, and finally constructs four machine learning models (Partial Least Squares Regression, PLSR; Support Vector Machine, SVM; Back Propagation Neural Network, BPNN and Extreme Gradient Boosting, XGBoost). The results show that spectral transformations (CR、SNV、FD and SD) can reduce the interference of external environments on soil spectra, effectively highlighting absorption and reflection features in the spectral curve, thus improving the accuracy of feature band selection and the prediction accuracy of the model. Among the feature selection methods, CARS is more suitable for soil iron oxide, while Boruta is more suitable for heavy metal Zn. In machine learning methods, both linear and nonlinear models can well explain the relationship between soil iron oxide and spectral reflectance, while the relationship between soil heavy metal Zn and spectral reflectance is nonlinear. The best retrieval model combination for soil iron oxide is FD_CARS_SVM, with RC2 = 0.878, RMSEC = 4.395, R2V = 0.849, RMSEV = 4.478, and RPDV = 2.576. The best retrieval model combination for heavy metal zinc is FD_Boruta_XGBoost, with RC2 = 0.999, RMSEC = 0.102, R2V = 0.682, RMSEV = 2.697, and RPDV = 1.772.

Objective Soil's heavy metal contamination is a significant global environmental issue that demands immediate attention. Particularly in the karst mining regions of Southwest China, where complex topography and intensive mining threaten agricultural ecosystems and human health. Nickel (Ni) is a heavy metal characterized by its high mobility and strong bio-enrichment potential and represents a significant environmental risk. In the mining area of Huize County, Yunnan Province, nickel is an essential micronutrient for plants in trace amounts, but long-term mining and smelting operations have led to a remarkable accumulation of Ni in the surrounding regions, posing a profound and ongoing threat to the integrity of agricultural ecosystems and the well-being of human populations. It results in severe contamination. Traditional heavy metal detection methods, such as atomic absorption spectrometry, offer high precision. However, its operational complexity and high precision impede the application for large-scale and rapid monitoring and offer only discrete and point-based data, rendering them ill-suited for the rapid and large-scale monitoring required for effective environmental management. Hyperspectral remote sensing has emerged as a transformative solution, offering a rapid, non-destructive, and cost-effective technology capable of extensive area monitoring and large-area surveillance. We focused on the cultivated land surrounding the mining district of Huize County, Qujing City, aiming to develop a precise and reliable prediction model for soil's Ni concentration by leveraging advanced hyperspectral technology, thereby providing a critical tool for regional heavy metal pollution monitoring and sustainable soil environmental management. Methods This study was based on 83 topsoil samples collected from cultivated land in the vicinity of the Huize mining area. The hyperspectral reflectance data, ranging from 350 nm to 2500 nm, were acquired using an ASD Field Spec 3 spectroradiometer, while the corresponding soil's nickel content was determined by an X-ray fluorescence spectrometer. The Kennard--Stone (K--S) algorithm was employed to partition the samples into a training set and a test set at a 7 & ratio;3 ratio. For spectral preprocessing, noisy edge bands (350-449 nm and 2451-2500 nm) and regions corresponding to sensor transitions (990-1010 nm and 1820-1850 nm) were excised to eliminate non--chemical step effects. Subsequently, outliers were identified and removed, followed by the application of Savitzky-Golay (SG) smoothing. Six conventional spectral transformations, including second--order differentiation (SD) and multiplicative scatter correction (MSC), were applied to the smoothed raw spectra. These transformations were designed to resolve overlapping spectral features and correct for physical scattering effects. Feature wavelength selection was performed using methods such as competitive adaptive reweighted sampling (CARS) and the least absolute shrinkage and selection operator (Lasso). These methods were utilized to identify the most informative wavelengths, thereby reducing data redundancy and preventing model overfitting. Finally, four inversion models, including back propagation neural network (BPNN) and partial least squares regression (PLSR), were constructed. Model performance was evaluated using the coefficient of determination (R-2), root mean square error (RMSE), and the ratio of performance to deviation (RPD). Results and Discussions A key finding of this research is that the SD transformation significantly enhances the spectral sensitivity to Ni. It yields a maximum negative correlation coefficient of-0.61 at 1907 nm, indicating that this wavelength possesses exceptionally high spectral responsiveness and is the most critical and primary characteristic wavelength for Ni inversion. The application of the Lasso proves pivotal in balancing feature selection with predictive accuracy. By selecting a concise yet information--rich subset of wavelengths, Lasso markedly improves the model's predictive capability and generalization performance. A comprehensive comparison reveals that the FD-Lasso-PLSR model demonstrates outstanding performance in predicting soil's Ni content, achieving a determination coefficient R-2 of 0.968 and a root--mean-square error of 6.599 for the training set, and a determination coefficient R-2 of 0.952, a root--mean--square error of 12.218, and an RPD of 3.359 for the test set. This indicates that the model possesses remarkable accuracy and robustness, confirming the superiority of this hybrid methodology for the rapid and precise detection of soil's Ni content in mining areas. Conclusions We unequivocally demonstrate the efficacy of hyperspectral technology for the rapid and accurate detection of soil's nickel content in the cultivated soils of mining areas. The high precision and reliability of the FD-Lasso-PLSR model, as evidenced by the test set results (R-2 is 0.952, RMSE is 12.218, and RPD is 3.359), provide crucial data support for pollution prevention and control and for soil environmental management in the karst mining regions of Southwest China and significantly advance the practical application of hyperspectral remote sensing, demonstrating its utility in topographically and geochemically complex landscapes. The ability to rapidly, non-destructively, and accurately assess soil's Ni contamination over large spatial extents is of paramount importance. We promote the practical application of hyperspectral remote sensing in regions with complex geomorphology. The capacity for rapid, non-destructive, and accurate assessment of soil's Ni content over large areas is vital for identifying pollution hotspots, tracking contaminant diffusion, evaluating the effectiveness of remediation efforts, and informing land--use planning to mitigate risks to agricultural productivity and human health. Future research should focus on integrating these models with satellite--based hyperspectral imagery or deploying networks of in--situ ground--based spectral sensors. Such efforts would facilitate the transition from local--scale assessment to large-scale and dynamic monitoring, paving the way for a truly comprehensive and responsive environmental management framework.