High vegetation coverage area surrounding rock alteration extraction and prospecting target area delineating method

Abstract

The invention belongs to the field of remote sensing geological exploration technology, remote sensing image processing and deep learning, and particularly discloses a high vegetation coverage area surrounding rock alteration extraction and prospecting target area delineating method, which comprises the following steps: preprocessing multi-source remote sensing data; the preprocessed data are input into a surrounding rock alteration extraction model, an alteration probability graph is output, and the model comprises a feature extraction module used for extracting spectral features and texture features from the preprocessed data, the spectral features comprise absorption and reflection features of iron staining and hydroxyl alteration in different wavelengths, and the texture features comprise texture features; the texture features comprise quartz vein body block-shaped textures and iron cap honeycomb-shaped textures; the feature fusion module is used for carrying out adaptive fusion on the spectral features and the texture features, enhancing alteration sensitive features through a learnable weight matrix, inhibiting vegetation interference and generating fusion features; and the alteration seed point detection module is used for outputting an alteration probability graph based on the fusion features. According to the method, the accuracy of alteration information extraction of the high vegetation coverage area can be remarkably improved.

Description

Technical Field

[0001] This application belongs to the fields of remote sensing geological exploration technology, remote sensing image processing, and deep learning. More specifically, it relates to a method for extracting the alteration of surrounding rocks in areas with high vegetation cover and delineating mineral exploration target areas. Background Technology

[0002] Wall rock alteration, as a key indicator of mineralization in hydrothermal deposits, plays a crucial role in guiding the delineation of prospecting target areas. In recent years, with the deepening of the national strategy for breakthroughs in mineral exploration, the demand for green exploration and efficient prediction of mineral resources has become increasingly urgent. Accurate extraction of wall rock alteration information has become a core aspect of mineral prediction, especially in areas with high vegetation cover. Traditional geological survey methods are costly and inefficient, necessitating the introduction of remote sensing technology to achieve rapid and large-scale alteration information identification, thereby meeting the significant strategic needs of national mineral resource security and green exploration.

[0003] Remote sensing technology, with its advantages of macroscopic scope, multispectral density, and timeliness, has become an important means of extracting alteration information. Currently, traditional methods such as principal component analysis (PCA) and band ratio methods are effective in extracting alteration in exposed areas, but they still face serious challenges in areas with high vegetation cover: First, the vegetation canopy severely interferes with surface spectral signals, obscuring alteration mineral characteristics; second, low-to-medium resolution data (such as Landsat-8) exhibits a prominent mixed-pixel effect, making it difficult to distinguish between vegetation and alteration spectral responses; third, existing vegetation suppression methods (such as NDVI thresholding) easily lead to the loss of effective alteration information; fourth, a single data source cannot simultaneously consider both spectral resolution and spatial detail; and fifth, the lack of intelligent interpretation models that integrate multi-source remote sensing data with prior geological knowledge results in a high false anomaly rate and insufficient reliability of the extraction results.

[0004] Therefore, improving the accuracy of alteration extraction of surrounding rocks in vegetated areas is an urgent problem to be solved. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this application is to provide a method for extracting the alteration of surrounding rocks and delineating mineral exploration target areas in areas with high vegetation cover, which can significantly improve the accuracy of alteration information extraction in areas with high vegetation cover.

[0006] To achieve the above objectives, in a first aspect, this application provides a multi-task collaborative method for extracting alteration of surrounding rocks in high vegetation cover areas, comprising the following steps: S10, perform radiometric calibration, atmospheric correction, spatial registration, and PAN-Sharpening fusion on the multi-source remote sensing data to generate preprocessed multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; S20, input the preprocessed multi-source remote sensing data into a pre-constructed surrounding rock alteration extraction model, and output an alteration probability map; the surrounding rock alteration extraction model includes: The feature extraction module is used to extract spectral features and texture features from the preprocessed multi-source remote sensing data. The spectral features include the absorption and reflection features of iron staining and hydroxyl alteration at different wavelengths. The texture features include quartz vein mass texture and iron cap honeycomb texture. The feature fusion module is used to adaptively fuse the spectral features and texture features, enhance the alteration-sensitive features and suppress vegetation interference through a learnable weight matrix, and generate fused features. The alteration seed point detection module is used to output an alteration probability map based on the fused features.

[0007] The multi-task collaborative method for extracting alteration of surrounding rocks in high-vegetation-covered areas provided in this application has the following advantages: Through collaborative preprocessing of multi-source remote sensing data, high-quality fused data is generated, overcoming the limitations of mixed pixel effects in medium- and low-resolution data and insufficient spectral and spatial details from single data sources; by simultaneously utilizing the spectral features of optical images and the texture features of radar images through the feature extraction module, the absorption and reflection responses of iron staining and hydroxyl alteration in specific bands, as well as the blocky and honeycomb texture patterns, can be deeply mined, avoiding the obscuring of alteration mineral features by the vegetation canopy; and through the adaptive feature fusion module… The weighted matrix enhances the alteration-sensitive bands and suppresses vegetation interference, effectively reducing the interference of vegetation background noise on spectral signals and minimizing the loss of effective alteration information caused by existing vegetation suppression methods. The alteration seed point detection module outputs a probability map based on fused features, enabling end-to-end alteration information localization and improving the recall and accuracy of alteration information extraction. The synergistic effect of these technologies, leveraging the complementary advantages of multi-source data and adaptively optimizing feature representation, achieves accurate extraction of alteration information even in high-vegetation-coverage environments, significantly improving the accuracy and robustness of alteration information extraction.

[0008] As a further preferred embodiment, the spectral characteristics include an absorption valley in the 0.45–0.51 μm band and a reflection peak in the 0.63–0.69 μm band for iron staining alteration, and a reflection peak in the 1.6 μm band and an absorption valley in the 2.2 μm band for hydroxyl alteration.

[0009] As a further preferred embodiment, the optical image data includes at least one of Landsat series images, GF-2 panchromatic image, and Sentinel-2 MSI data, and the radar image data includes at least one of GF-3 SAR image data and Sentinel-1 C-SAR data. As a further preferred embodiment, in the feature extraction module, when extracting the spectral features, a Transformer encoder is used to process the spectral sequence of the optical image data, focusing on the characteristic band responses of iron-stained minerals and hydroxyl minerals, and transforming the physical spectral mechanism into a high-dimensional feature vector; when extracting the texture features, a two-dimensional convolutional neural network is used to process the texture features of the radar image data, including the honeycomb and fragmented textures presented by the "iron cap" and the clump-like and dense shadow textures caused by silicification alteration.

[0010] As a further preferred embodiment, when the feature fusion module performs adaptive fusion, it includes: performing global average pooling on the spectral features and texture features to obtain a global descriptor for each feature channel; then, performing nonlinear interaction through a fully connected network with a bottleneck structure, and autonomously learning the weights of each feature channel by combining an adaptive weighting mechanism, enhancing the weights of alteration-sensitive bands, suppressing homogeneous texture of the vegetation canopy, and highlighting possible outcrop areas of texture abrupt changes and topographic abrupt changes in the image.

[0011] As a further preferred embodiment, the alteration seed point detection module constructs a high-dimensional feature map based on a fully connected layer and outputs a probability distribution of alteration types, including iron staining alteration and hydroxyl alteration.

[0012] As a further preferred embodiment, step S10, the preprocessing of the multi-source remote sensing data, further includes: Feature parameters are calculated from optical image data, including extracting hydroxyl mineralization and alteration anomaly information using Band5 or Band4, and extracting iron staining mineralization and alteration anomaly information using Band6 or Band5. Image masking and dataset construction were carried out, including masking irrelevant features by calculating the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Building Index (NDBI).

[0013] Secondly, this application provides a method for delineating mineral exploration target areas, comprising the following steps: (1) Perform radiometric calibration, atmospheric correction, spatial registration and PAN-Sharpening fusion on multi-source remote sensing data to generate preprocessed multi-source remote sensing data; (2) Input the preprocessed multi-source remote sensing data into a pre-constructed surrounding rock alteration extraction model and output an alteration probability map; the surrounding rock alteration extraction model includes: The feature extraction module is used to extract spectral features and texture features from the preprocessed multi-source remote sensing data. The spectral features include the absorption and reflection features of iron staining and hydroxyl alteration at different wavelengths. The texture features include quartz vein mass texture and iron cap honeycomb texture. The feature fusion module is used to adaptively fuse the spectral features and texture features, enhance the alteration-sensitive features and suppress vegetation interference through a learnable weight matrix, and generate fused features. The alteration seed point detection module is used to output an alteration probability map based on the fused features; (3) Input the alteration probability map into the pre-constructed mineral exploration target area delineation model, and determine the spatial topological relationship through encoder-decoder and jump link to delineate the mineral exploration target area.

[0014] As a further preferred embodiment, the prospecting target area delineation model adopts a U-Net structure, with the input being the fusion of the alteration probability map and deep features, and the output being a binary target area probability map.

[0015] As a further preferred embodiment, the surrounding rock alteration extraction model and the mineral exploration target area delineation model are trained using a dynamic weighted loss function. The dynamic weighted loss function includes a Dice Loss for mineral exploration target area delineation and a weighted regularization term, calculated using the following formula: L total = λ 1 L seed + λ 2 L expand + λ 3 L reg in, L seed Focal Loss for Altering Seed Point Detection; L expand Dice Loss for delineating mineral exploration target areas; L reg λ1, λ2, and λ3 are weight regularization terms; λ1, λ2, and λ3 are dynamic weights.

[0016] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-task collaborative method for extracting alteration of surrounding rocks in high vegetation cover areas provided in this application; Figure 2 This is an overall flowchart of the multi-task collaborative method for extracting alteration of surrounding rocks in high vegetation cover areas and delineating mineral exploration target areas, provided in the embodiments of this application. Figure 3 This is a schematic diagram of the construction of a network model for extracting alteration of surrounding rocks in high vegetation cover areas and delineating mineral exploration target areas through multi-task collaboration, as provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] like Figure 1 As shown, this application provides a multi-task collaborative method for extracting alteration of surrounding rocks in areas with high vegetation cover, including steps S10 to S20, which are detailed below: Step S10: Perform radiometric calibration, atmospheric correction, spatial registration, and PAN-Sharpening fusion on the multi-source remote sensing data to generate preprocessed multi-source remote sensing data.

[0020] Multi-source remote sensing data includes optical image data and radar image data.

[0021] This step preprocesses the multi-source remote sensing data, primarily to eliminate differences between data sources, ensure spatial alignment between images, and improve the spatial resolution of the images.

[0022] Step S20: Input the preprocessed multi-source remote sensing data into the pre-constructed surrounding rock alteration extraction model and output the alteration probability map.

[0023] In step S20, the surrounding rock alteration extraction model includes a feature extraction module, a feature fusion module, and an alteration seed point detection module.

[0024] The feature extraction module is used to extract spectral and texture features from the preprocessed multi-source remote sensing data. The spectral features include the absorption and reflection features of iron staining and hydroxyl alteration at different wavelengths, and the texture features include the quartz vein mass texture and the iron cap honeycomb texture.

[0025] The feature fusion module is used to adaptively fuse spectral features and texture features. It enhances alteration-sensitive features and suppresses vegetation interference through a learnable weight matrix to generate fused features. The alteration seed point detection module is used to output an alteration probability map based on fused features.

[0026] Specifically, the feature extraction module provided in this application can comprehensively utilize the spectral sequence of optical images and the texture patterns of radar images to deeply mine spectral features such as the absorption valley of iron staining alteration in the 0.45–0.51 μm band and the reflection peak in the 0.63–0.69 μm band, the reflection peak of hydroxyl alteration in the 1.6 μm band and the absorption valley in the 2.2 μm band, as well as the clumpy texture features of quartz veins and honeycomb texture features of iron caps, capturing key identifiers of alteration minerals under vegetation cover; through the adaptive weighting mechanism of the feature fusion module, global average pooling and nonlinear interaction are performed on multi-source features to learn the weights of each feature channel, enhance the weights of alteration-sensitive bands such as short-wave infrared, and suppress the homogeneous texture of the vegetation canopy, effectively separating vegetation physiological signals from the spectral response of alteration minerals; through the alteration seed point detection module, a high-dimensional feature mapping is constructed based on a fully connected layer to output the probability distribution of alteration types, which can accurately locate the iron staining and hydroxyl alteration points and reduce missed detections caused by vegetation background noise.

[0027] The multi-task collaborative method for extracting alteration of surrounding rocks in high-vegetation-covered areas provided in this application has the following advantages: Through collaborative preprocessing of multi-source remote sensing data, high-quality fused data is generated, overcoming the limitations of mixed pixel effects in medium- and low-resolution data and insufficient spectral and spatial details from single data sources; by simultaneously utilizing the spectral features of optical images and the texture features of radar images through the feature extraction module, the absorption and reflection responses of iron staining and hydroxyl alteration in specific bands, as well as the blocky and honeycomb texture patterns, can be deeply mined, avoiding the obscuring of alteration mineral features by the vegetation canopy; and through the adaptive feature fusion module… The weighted matrix enhances the alteration-sensitive bands and suppresses vegetation interference, effectively reducing the interference of vegetation background noise on spectral signals and minimizing the loss of effective alteration information caused by existing vegetation suppression methods. The alteration seed point detection module outputs a probability map based on fused features, enabling end-to-end alteration information localization and improving the recall and accuracy of alteration information extraction. The synergistic effect of these technologies, leveraging the complementary advantages of multi-source data and adaptively optimizing feature representation, achieves accurate extraction of alteration information even in high-vegetation-coverage environments, significantly improving the accuracy and robustness of alteration information extraction.

[0028] In one embodiment, the technical solution to achieve the above objectives can be as follows: This embodiment proposes a multi-task collaborative method for extracting wall rock alteration and delineating mineral exploration target areas in high vegetation cover areas. This method innovatively integrates Landsat and GF-3 SAR data, effectively separating vegetation physiological signals and alteration mineral spectral responses through vegetation-lithology spectral decoupling and suppression; it constructs a multi-task collaborative network model for extracting wall rock alteration and delineating mineral exploration target areas in high vegetation cover areas, combining geological structure and lithological prior knowledge to achieve accurate extraction and spatial orientation of alteration information; simultaneously, it introduces multi-cognitive visual filtering and distribution optimization techniques, freezes parameters, and performs lightweight adaptation to improve the model's efficiency and generalization ability in vertical applications. This embodiment can achieve rapid location of wall rock alteration and delineation of mineral exploration target areas under high vegetation cover backgrounds, providing technical methodological support and data reference for the new round of mineral exploration breakthroughs in China.

[0029] like Figure 2 As shown in this embodiment, the multi-task collaborative method for extracting wall rock alteration in high vegetation cover areas and delineating mineral exploration target areas includes the following steps: First, preprocessing multi-source remote sensing data to eliminate differences between data sources and ensure spatial alignment between images. The data involved mainly includes Landsat series images, GF-2 image panchromatic band (2m), GF-3 SAR image data, etc. After spatial alignment of the images, image fusion technology is used to improve the spatial resolution of the images. Second, a multi-task collaborative network model for extracting wall rock alteration in high vegetation cover areas is designed. This model includes the following four modules: ① Feature extraction module, which extracts spectral and textural features from different data sources. The spectral features involve the absorption and reflection characteristics of iron staining and hydroxyl alteration at different wavelengths. In terms of texture features, hydroxyl alteration often presents as quartz veins, which can be seen in nodular bioherm limestone. Iron staining alteration is prone to occur near primary sulfide ore bodies, presenting an "iron cap" feature. ② Adaptive Weight Calculation: Multi-source features are fused and analyzed to quantify weights, increasing the focus on rock outcrop alteration information and reducing interference from background information such as vegetation. ③ Alteration Extraction and Mineral Target Area Delineation Module: A dual-branch network for extracting and expanding seed points of surrounding rock alteration is constructed. On the one hand, alteration information of rock outcrop areas is directly extracted through multi-source features. On the other hand, due to the limitations of high vegetation cover areas, the determined surrounding rock alteration points are comprehensively evaluated. Based on alteration type and intensity, the correlation between fragmented alteration points is realized to delineate mineral target areas. ④ A multi-cognitive visual filter and input distribution optimization module are constructed, freezing some training model parameters to achieve efficient adaptation of a small number of parameter adjustments to a large-scale vertical application domain. Third, the constructed multi-source image dataset is labeled to build a sample dataset. Finally, the model is trained and evaluated using the sample dataset and a validation machine. The accuracy of the model is evaluated using metrics such as precision, recall, false positive rate, and mean intersection-over-union ratio.

[0030] The specific implementation steps are as follows: A. Construction of multi-source remote sensing datasets: A1. Acquisition of Multi-Source Image Data: Taking into account information such as the background of high vegetation cover areas and the time of field surveys, multi-source data was downloaded, mainly including optical images and radar images. Among them, optical image data included Landsat (8 / 9 OLI), GF-2PMS, Sentinel-2 MSI data, etc.: Multiple periods of data were acquired for the study area during the growing season (lush vegetation) and the dry season (minimum vegetation disturbance and better bedrock exposure). These data were mainly used to extract spectral alteration anomalies (iron staining, hydroxyl) and vegetation indices. The red-edge band of the aforementioned optical image data is more sensitive to vegetation monitoring, and it also has key infrared bands (such as Band 11, SWIR-1: 1.61μm; Band 12, SWIR-2: 2.20μm in Sentinel-2), which is sufficient to meet the needs of extracting spectral features of hydroxyl and iron staining alteration in high vegetation cover areas. Radar imagery data mainly includes data from Gaofen-3 (GF-3) SAR and Sentinel-1 C-SAR: acquiring data in multiple polarization modes, whether L-band or C-band data, can still effectively provide surface roughness and topographic structure information in vegetated areas. It is crucial for enhancing texture features, assisting in the identification of linear structures and blocky geological bodies, and can be used for terrain correction assistance, penetration of vegetated areas, enhancement of structural information, and characterization of surface roughness.

[0031] A2. Preprocessing of Multi-Source Image Data: Optical image preprocessing mainly includes radiometric calibration by converting the raw DN values ​​to top-level atmospheric radiance or reflectance; atmospheric correction by converting radiance to true surface reflectance using the FLAASH model; orthorectification by using DEM data to eliminate image point displacement caused by topographic undulations; and fusion and registration. The final result is a raw dataset with both high spatial and spectral resolution, identical coordinate systems, and an error of less than one pixel. Radar image preprocessing includes radiometric calibration by converting backscattering coefficients to σ° or γ° values ​​to give them true physical meaning and comparability; secondly, using filters such as Refined Lee and Gamma Map to suppress inherent speckle noise in SAR images while preserving edge and texture information as much as possible; and finally, topographic geometric correction using DEM data to convert the image from slant-range projection to ground-range projection to eliminate geometric distortion.

[0032] A3. Feature parameter calculation: For optical images, based on the input of each band, band calculation is performed in combination with the reflectivity characteristics of mineral ions such as hydroxyl and iron stain. Band5 / Band4 is used to extract hydroxyl mineralization alteration anomaly information, and Band6 / Band5 is used to extract iron stain mineralization alteration anomaly information.

[0033] A4. Image Masking and Dataset Construction: To highlight the true information of surrounding rock alteration, it is necessary to mask or suppress irrelevant features and interfering factors. By calculating the Normalized Difference Vegetation Index (NDVI), Water Index (NDWI), and Building Index (NDBI), irrelevant features and interfering factors are masked. Finally, the processed spectral features, texture features, and masked data are unified to the same spatial reference and pixel size to form a multi-dimensional, multi-scale dataset as the feature input for deep learning.

[0034] B. Construction of a Multi-task Collaborative Network Model for Extracting Rock in High Vegetation Cover Areas and Delineating Prospecting Targets: Addressing the problem of rock alteration in high vegetation cover areas, a multi-task collaborative network model for extracting rock in high vegetation cover areas and delineating prospecting targets (MCA-T-Net) was constructed, as follows: Figure 3 As shown, the specific steps are as follows.

[0035] B1. Multi-Source Feature Collaborative Extraction Module: This module employs a dual-branch parallel architecture to process multi-source remote sensing data. The spectral feature extraction branch utilizes a Transformer encoder to deeply mine the spectral sequences of Landsat-8 data, focusing on the characteristic band responses of iron-stained minerals and hydroxyl minerals, transforming the physical spectral mechanisms into high-dimensional feature vectors. Simultaneously, the spatial texture feature extraction branch uses a two-dimensional convolutional neural network (2D CNN) to process high-resolution SAR images such as Gaofen-3. By training the convolutional kernels, it adaptively captures texture patterns related to alteration, such as the honeycomb and fragmented textures exhibited by "iron caps," and the clumpy and dense shadows caused by silicification alteration. Finally, the spectral and texture feature vectors originating from the same pixel are concatenated to form a deeply fused feature representation, providing a rich data foundation containing both mineral composition and morphological structure information for subsequent processing.

[0036] B2. Multi-cognitive visual filter and optimized input distribution module: Addressing the difficulty in collecting large amounts of training data on rock alteration in high-vegetation-coverage areas, this embodiment constructs a multi-cognitive visual filter and optimized input distribution module. Through efficient parameter fine-tuning, it maintains the parameter freeze and lightweight adaptation in the pre-trained model. Combining techniques such as visual filters, sigmoid activation functions, dimensionality increase, and skip links, it can, on the one hand, simply fine-tune all parameters of the target model by minimizing the specific task loss on a given training dataset, and on the other hand, improve the performance of visual tasks.

[0037] B3. Feature Fusion Module: Global average pooling (Squeeze operation) is performed on the multi-source fused feature map output by Module 1 to obtain the global descriptor of each feature channel; then, nonlinear interaction is performed through a small fully connected network with a bottleneck structure (Excitation operation), combined with an adaptive weight mechanism (SEAttention) to autonomously learn the weights of each feature channel, enhance the weights of alteration-sensitive bands such as shortwave infrared, suppress the homogeneous texture of vegetation canopy, and highlight the possible outcrop areas of texture abrupt changes and topographic abrupt changes in the image.

[0038] B4. Dual-branch Collaborative Perception Network Module: This module constructs a multi-branch ecological information extraction network model, simultaneously optimizing the dual-branch tasks of alteration point extraction and directional expansion. In the seed point extraction branch, a high-dimensional feature mapping is built based on a fully connected layer to output the probability distribution of alteration types, accurately capturing geomorphic features such as iron staining and hydroxyl alteration. In the mineral exploration target area delineation branch, seed points are used as strong prior knowledge through an encoder-decoder and skip connections to delineate mineral exploration boundaries using pixels with similar spectra and textures around the seed points. A dynamic loss weighting technique is designed to achieve an adaptive balance between extraction and delineation tasks, constructing an end-to-end collaborative mechanism that shares underlying spatiotemporal feature representations, enabling dual-branch information interaction. This improves the spatiotemporal consistency of mineral exploration target area delineation while ensuring classification accuracy, achieving fully automated modeling from original time-series input to surrounding rock alteration extraction and mineral exploration target area delineation.

[0039] C. Train the neural network model using the training and validation sets. C1. Feature Extraction and Sample Labeling: Based on the preprocessed multi-source dataset, image patches are extracted as sample units using a multi-scale sliding window (256×256 pixels). Geological labeling is collaboratively completed, and multi-source verification is performed by combining field verification points, geological maps, and high-resolution imagery.

[0040] C2. Training Strategy and Hyperparameter Settings: A multi-task collaborative training strategy is adopted, using a dynamic weighted loss function to simultaneously optimize the seed point detection branch and the mineral exploration target area delineation branch. The training hyperparameters are set as follows: initial learning rate 1×10⁻⁶. -4 (Cosine annealing scheduling is used), the batch size is 16, the AdamW optimizer is used (weight decay of 0.01), and the number of training epochs is 200. The early stopping mechanism is triggered when the validation set loss has not decreased for 10 consecutive epochs.

[0041] C3. Loss Function Settings: The loss function consists of three parts. L total = λ 1 L seed + λ 2 L expand + λ 3 L reg . L seed Focal Loss for seed point detection branch addresses the class imbalance problem between etched and non-etched pixels; L expand To delineate branches of the mineral exploration target area using Dice Loss, the boundary fitting degree of altered connected regions is optimized; L reg The weights are L2 regularization terms to prevent overfitting; λ1, λ2, and λ3 are dynamic weights that are adaptively adjusted according to the training progress of each branch. D. Apply the trained model to extract the alteration of surrounding rocks in areas with high vegetation cover: Specific steps include: D1. Select the study area, construct a multi-source remote sensing dataset, and then input it into the trained MCA-T-Net network model to obtain anomalous areas of different levels of iron staining and hydroxyl alteration, obtain their surrounding spatial topological relationships, and delineate potential mineral exploration target areas.

[0042] D2. Evaluate the accuracy of the surrounding rock alteration extraction results using the test set. Construct a confusion matrix and evaluate the results using metrics such as precision (P), recall (R), false positive rate (MRate), and mean intersection-union ratio (MIoU).

[0043] Compared with existing technologies, this embodiment has the following advantages and beneficial effects: 1) Multi-source feature fusion effectively suppresses vegetation interference: This embodiment combines optical image spectral features and SAR data texture features simultaneously, and enhances the alteration-sensitive band through an adaptive weighting mechanism, which significantly reduces vegetation background noise and improves the recall rate of alteration information extraction.

[0044] 2) Multi-task collaborative architecture significantly improves extraction accuracy: This embodiment establishes a dual-task branch for the extraction of surrounding rock alteration and the delineation of mineral exploration target areas. Through the dual-branch network, alteration seed point detection and reconstruction of the spatial topological relationship of alteration veins are realized simultaneously, overcoming the problems of fragmented alteration information and high false negative rate in vegetation-covered areas in traditional methods, and effectively improving the prediction accuracy of mineral exploration target areas.

[0045] 3) Strong dynamic loss weight optimization and generalization ability: It adopts a dynamic weighted multi-task loss function to balance the optimization objectives of seed point localization and region expansion. Combined with multi-cognitive visual filters and optimized input distribution, it achieves efficient cross-regional adaptation and lightweight model parameter adjustment, and maintains high accuracy in unknown regions.

[0046] 4) End-to-end automated processes improve exploration efficiency: This embodiment automates the entire process from multi-source data preprocessing to target area delineation, significantly reducing manual interpretation costs and providing efficient and reliable technical support for the new round of mineral exploration breakthrough strategic actions.

[0047] This invention addresses the challenges of extracting alteration of surrounding rocks and delineating mineral exploration targets in areas with high vegetation cover. Based on multi-source remote sensing data fusion and the MCA-T-Net network model, it achieves fully automated modeling from raw time-series input to the extraction of alteration of surrounding rocks and the delineation of mineral exploration targets. It actively responds to the call of the national new round of strategic action for breakthroughs in mineral exploration and is of great significance for serving the major strategic needs of national mineral resource security and green exploration.

[0048] The following is a specific embodiment of this application. This specific embodiment provides a multi-task spatiotemporal collaborative method for monitoring complex surface ecology in hilly areas of southern China. The method of this application is implemented in the Tianlin area of ​​Baise City, Guangxi Zhuang Autonomous Region. The specific operation steps are as follows: A. Study area delineation and construction of multi-source remote sensing dataset: A1. Study Area and Data Source Selection: Tianlin County, Baise City, Guangxi (105°27′~106°15′E, 24°13′~24°45′N) was selected as the implementation area. This area belongs to the western section of the Nanling metallogenic belt, with hilly and mountainous terrain and a forest coverage rate of over 75%, making it a typical high vegetation cover area. Carlin-type gold deposits and polymetallic mineralization clues are known in the area, and limonite mineralization, iron staining, and hydroxyl alteration are relatively densely distributed. It is an ideal area for prospecting and predicting concealed to semi-concealed mineral deposits. Data includes optical imagery, radar imagery, and related auxiliary data. Optical data includes imagery acquired by GF-2 PMS in 2023 (panchromatic resolution 0.8m, multispectral resolution 3.2m), used for fine-grained land cover classification and texture feature extraction. Four Landsat-8 OLI images (dry season and peak season) were acquired, with cloud cover <10%. Used for calculating vegetation indices and extracting alteration fields. Two scenes of dual-polarization (HH+HV) strip pattern data from Gaofen-3 (GF-3) SAR were acquired to penetrate the vegetation layer and enhance structural information. Auxiliary data mainly consisted of ALOS World 3D-30m (AW3D30) DEM data for terrain correction and feature extraction, a 1:200,000 regional geological map, and field reconnaissance data verifying surrounding rock alteration points.

[0049] A2. Remote Sensing Data Preprocessing: For optical image data, the Radiometric Calibration and FLAASH modules in ENVI software were used to convert DN values ​​to surface reflectance. Furthermore, orthorectification was performed on the GF-2 imagery based on the AW3D30 DEM. All optical images were registered to the WGS84 UTM Zone 48N coordinate system, with a registration error RMSE < 0.5 pixels. Finally, the GF-2 data were fused using the Gram-Schmidt Pan Sharpening algorithm to generate a 0.8-meter resolution multispectral image. For radar image data, the SARscape module was used for radiometric calibration to generate σ° maps, and a RefinedLee filter (7x7 window size) was used to suppress speckle noise. Simultaneously, topographic geometric correction was performed.

[0050] A3. Feature parameter calculation: Based on the preprocessed Landsat-8 data, calculate the following features (Band in the formula refers to the Landsat-8 OLI band).

[0051] NDVI calculation: (float(b5)-float(b4)) / (float(b5)+float(b4)) NDBI calculation: (float(b6)-float(b5)) / (float(b6)+float(b5)) MNDWI calculation: (float(b3)-float(b6)) / (float(b3)+float(b6)) After calculation, mask thresholds were set for each parameter to reduce the influence of vegetation cover. The NDVI threshold was set to 0.5, the NDBI threshold to 0, and the MNDWI threshold to 0.1. The mask images of the three parameters were multiplied to obtain the interference mask extraction results for the study area. Furthermore, principal component analysis was performed using bands 2, 5 / 4, 6, and 7 to extract hydroxyl mineralization and alteration anomalies, and bands 2, 4, 5, and 6 / 5 were used to extract iron staining mineralization and alteration anomalies. The above bands were combined to generate the raw data for principal component extraction.

[0052] A4. Principal Component Analysis Extraction: Using the "Forward PCA Rotation Wew Stat" tool in ENVI, select the image results from step A3 to perform principal component extraction. Then, view the principal component analysis results using "view statistics File" and extract the principal component information that best matches the features.

[0053] B. Multi-task collaborative network model construction for extracting wall rock alteration in high vegetation cover areas and delineating mineral exploration target areas: B1. Multi-source Feature Collaborative Extraction Module: Spectral Branch: A 4-layer Transformer encoder (8 heads, feedforward network dimension = 512) processes the Landsat feature sequence. Input dimension is [Batch, 256*256, 12]. Texture Branch: A lightweight CNN (4 Conv-BN-ReLU blocks) processes the texture feature maps extracted from GF-3 SAR data. Feature Concatenation: The feature maps output from the two branches are concatenated along the channel dimension.

[0054] B2. Multi-cognitive visual filter and optimized input distribution module: This module employs LoRA (Low-Rank Adaptation) and mona techniques and models to efficiently fine-tune the parameters of the pre-trained visual backbone network. LoRA is set to a rank r=8 and a scaling parameter alpha=16, training only approximately 0.1% of the parameters to achieve rapid adaptation.

[0055] B3. Feature Fusion Module: Introduces a channel attention mechanism (SEBlock) to reweight the concatenated features. The compression ratio is set to 4.

[0056] B4. Dual-branch Collaborative Perception Network Module: Seed Point Detection Branch: Employs a U-Net decoder, outputting a three-class probability map for each pixel: background / iron stain / hydroxyl. Focal Loss is used with parameters α=0.25 and γ=2.0 to address class imbalance. Mineral Target Delineation Branch: Also uses a U-Net structure, but the input is a fusion of the seed point probability map and deep features, and the output is a binary target probability map (yes / no target). Dice Loss is used to optimize the boundary.

[0057] C. Model Training and Evaluation C1. Sample Labeling: Based on geological maps, field verification points, and manual interpretation, 256x256 image blocks are labeled. Seed Point Labels: Pixel-level labeling, 0-background, 1-iron staining alteration, 2-hydroxyl alteration. Target Area Labels: Based on known mineral occurrences, mineralization zones, and alteration zoning, continuous prospecting target area polygons are delineated and converted into binary masks.

[0058] C2. Training Strategy and Hyperparameters: Table 1 Network Parameter Settings

[0059] Table 2 Server Configuration

[0060] C3. The accuracy of the model is tested using four classification evaluation indicators: producer accuracy, user accuracy, overall accuracy, and Kappa coefficient. After multiple iterations of training, the model with the highest accuracy is selected for regional alteration extraction and target delineation.

[0061] D. Apply the trained model to extract regional alteration and delineate the target area: D1. Model Application: Input the entire processed multidimensional feature dataset into the trained MCA-T-Net model. The model automatically outputs a distribution map of surrounding rock alteration types (raster) and a delineation map of prospecting target areas (vector). It displays the spatial distribution of iron staining and hydroxyl alteration, and delineates Class I and Class II prospecting target areas according to the alteration distribution.

[0062] D2. Accuracy Evaluation: The accuracy of the monitoring results of the classification and regression tasks is determined by the validation set. Geochemical exploration and field survey results are used as the true label reference. Common alteration extraction methods such as principal component analysis + band ratio method, random forest method, and single task semantic segmentation method are compared. The confusion matrix is ​​calculated and evaluated by indicators such as precision (P), recall (R), false positive rate (MRate), and mean intersection-union ratio (MIoU).

[0063] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for extracting wall rock alteration in a high-vegetation-covered area in multi-task synergy, characterized in that, Includes the following steps: S10, perform radiometric calibration, atmospheric correction, spatial registration, and PAN-Sharpening fusion on the multi-source remote sensing data to generate preprocessed multi-source remote sensing data; the multi-source remote sensing data includes optical image data and radar image data; S20, input the preprocessed multi-source remote sensing data into the pre-constructed surrounding rock alteration extraction model and output the alteration probability map; The surrounding rock alteration extraction model includes: The feature extraction module is used to extract spectral features and texture features from the preprocessed multi-source remote sensing data. The spectral features include the absorption and reflection features of iron staining and hydroxyl alteration at different wavelengths. The texture features include quartz vein mass texture and iron cap honeycomb texture. The feature fusion module is used to adaptively fuse the spectral features and texture features, enhance the alteration-sensitive features and suppress vegetation interference through a learnable weight matrix, and generate fused features. The alteration seed point detection module is used to output an alteration probability map based on the fused features.

2. The method for extracting alteration of surrounding rocks in high-vegetation-coverage areas through multi-task collaborative methods as described in claim 1, characterized in that, The spectral characteristics include an absorption valley in the 0.45–0.51 μm band and a reflection peak in the 0.63–0.69 μm band for iron staining alteration, and a reflection peak in the 1.6 μm band and an absorption valley in the 2.2 μm band for hydroxyl alteration.

3. The method for extracting alteration of surrounding rocks in high vegetation cover areas through multi-task collaborative methods as described in claim 1, characterized in that, The optical image data includes at least one of Landsat series images, GF-2 panchromatic images, and Sentinel-2 MSI data, and the radar image data includes at least one of GF-3 SAR image data and Sentinel-1 C-SAR data.

4. The method for extracting alteration of surrounding rocks in high-vegetation-coverage areas through multi-task collaborative methods as described in claim 1, characterized in that, In the feature extraction module, when extracting the spectral features, a Transformer encoder is used to process the spectral sequence of the optical image data, focusing on the characteristic band responses of iron-stained minerals and hydroxyl minerals, and transforming the physical spectral mechanism into a high-dimensional feature vector; when extracting the texture features, a two-dimensional convolutional neural network is used to process the texture features of the radar image data, including the honeycomb and fragmented textures presented by the "iron cap" and the clump-like and dense shadow textures caused by silicification alteration.

5. The method for extracting alteration of surrounding rocks in high-vegetation-coverage areas through multi-task collaborative methods as described in claim 1, characterized in that, When the feature fusion module performs adaptive fusion, it includes: performing global average pooling on the spectral features and texture features to obtain a global descriptor for each feature channel; then, performing nonlinear interaction through a fully connected network with a bottleneck structure, and autonomously learning the weights of each feature channel by combining an adaptive weighting mechanism, enhancing the weights of alteration-sensitive bands, suppressing homogeneous texture of the vegetation canopy, and highlighting possible outcrop areas of texture abrupt changes and topographic abrupt changes in the image.

6. The method for extracting alteration of surrounding rocks in high-vegetation-coverage areas through multi-task collaborative methods as described in claim 1, characterized in that, The alteration seed point detection module constructs a high-dimensional feature map based on a fully connected layer and outputs the probability distribution of alteration types, including iron staining alteration and hydroxyl alteration.

7. The method for extracting alteration of surrounding rocks in high vegetation cover areas through multi-task collaborative methods as described in claim 1, characterized in that, In step S10, the preprocessing of multi-source remote sensing data further includes: Feature parameters are calculated from optical image data, including extracting hydroxyl mineralization and alteration anomaly information using Band5 or Band4, and extracting iron staining mineralization and alteration anomaly information using Band6 or Band5. Image masking and dataset construction were carried out, including masking irrelevant features by calculating the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Building Index (NDBI).

8. A method for delineating mineral exploration target areas, characterized in that, Includes the following steps: (1) Perform radiometric calibration, atmospheric correction, spatial registration and PAN-Sharpening fusion on multi-source remote sensing data to generate preprocessed multi-source remote sensing data; (2) Input the preprocessed multi-source remote sensing data into the pre-constructed surrounding rock alteration extraction model and output the alteration probability map; The surrounding rock alteration extraction model includes: The feature extraction module is used to extract spectral features and texture features from the preprocessed multi-source remote sensing data. The spectral features include the absorption and reflection features of iron staining and hydroxyl alteration at different wavelengths. The texture features include quartz vein mass texture and iron cap honeycomb texture. The feature fusion module is used to adaptively fuse the spectral features and texture features, enhance the alteration-sensitive features and suppress vegetation interference through a learnable weight matrix, and generate fused features. The alteration seed point detection module is used to output an alteration probability map based on the fused features; (3) Input the alteration probability map into the pre-constructed mineral exploration target area delineation model, and determine the spatial topological relationship through encoder-decoder and jump link to delineate the mineral exploration target area.

9. The method for delineating mineral exploration target areas as described in claim 8, characterized in that, The prospecting target area delineation model adopts a U-Net structure. The input is the fusion of the alteration probability map and deep features, and the output is a binary target area probability map.

10. The method for delineating mineral exploration target areas as described in claim 8, characterized in that, The surrounding rock alteration extraction model and the mineral exploration target area delineation model are trained using a dynamic weighted loss function. This dynamic weighted loss function includes a Dice Loss for mineral exploration target area delineation and a weighted regularization term, calculated as follows: L total = λ 1 L seed + λ 2 L expand + λ 3 L reg in, L seed Focal Loss for Altering Seed Point Detection; L expand Dice Loss for delineating mineral exploration target areas; L reg λ1, λ2, and λ3 are weight regularization terms; λ1, λ2, and λ3 are dynamic weights.

Images