A rare earth ore prospecting method and system fusing fern ore indication and multi-source remote sensing

Abstract

The application discloses a kind of fusion fern ore-indicating multi-source remote sensing rare earth ore prospecting method and system, belong to mineral resources exploration technical field, by multi-source remote sensing data acquisition weathering crust type rare earth ore ore-forming factor remote sensing characteristics, preliminary delineation potential area, deploy the unmanned aerial vehicle of carrying VNIR camera to obtain high-resolution image, with texture, form and spectral index comprehensive extraction fern distribution information, and fern coverage result is down sampled to with other characteristics consistent resolution, form fern coverage layer.Coupling of biological indicator (ferns) and traditional rock-mineral-structure-landform ore-indicating factors, significantly improve the positioning accuracy and exploration efficiency of weathering crust rare earth ore.

Description

Technical Field

[0001] This invention belongs to the field of mineral resource exploration technology, specifically a multi-source remote sensing method and system for rare earth mineral exploration that integrates fern-based mineral indication. It primarily targets remote sensing exploration of weathered crust-type rare earth deposits. Its core innovation lies in using rare earth-rich ferns (such as *Dictamnus dasycarpus* and *Dryopteris crassirhizoma*, where the total rare earth content can reach a certain percentage) as biomarkers for mineralization. Combined with mineralization factors such as rock, minerals, structure, and geomorphology, multi-source remote sensing information identification and intelligent algorithms are used to efficiently and accurately predict and locate potential rare earth mining areas. Background Technology

[0002] Weathering crust type rare earth deposits are mostly formed in the weathering residual layers of granite, carbonate-alkaline rocks, or acidic volcanic rocks. Rare earth elements (REEs) are often represented by Nd. 3+ Dy 3+ Y 3+ Plasmon-exchange minerals are found in altered clay minerals such as kaolinite, illite, and montmorillonite. Therefore, existing exploration methods often rely on the identification of these clay minerals to determine potential mineralization areas. However, relying on geological mapping or single remote sensing indicators is prone to misjudgment or omission due to factors such as ground cover and weak spectral density. Furthermore, due to limitations in survey scope and high costs, it is difficult to quickly screen large areas of weathered crust mineralization potential.

[0003] In recent years, studies have found that some ferns (such as *Dicranopteris dichotoma* and *Dryopteris cinnamon*) can adsorb and enrich rare earth elements (REEs) to the percentage level in weathering crust environments, exhibiting significant mineralization characteristics and providing new possibilities for biomineralization. However, bioindicators alone are insufficient to fully reveal mineralization distribution; it is necessary to simultaneously consider REE-related mineralization factors such as parent rock lithology, clay mineral alteration, topographic relief, and tectonic fracture zones. By fusing multi-source remote sensing data and incorporating the aforementioned mineralization factors along with fern distribution into predictive models using machine learning algorithms (especially deep learning), it is possible to consider local bioindicators at the regional scale, significantly improving the efficiency and accuracy of rare earth mineral exploration in weathering crusts.

[0004] Therefore, there is an urgent need to develop a multi-source remote sensing prospecting technology that integrates the biomineral characteristics of rare earth element-rich ferns with traditional rock-mineral-structure-geomorphology mineralization factors, and utilizes advanced machine learning algorithms to achieve high-precision identification and location of rare earth deposits in weathered crusts. Summary of the Invention

[0005] This invention provides a multi-source remote sensing method and system for rare earth mineral exploration that integrates fern mineralization. It utilizes remote sensing features of mineralization factors such as "rock-mineral-structure-topography" to conduct preliminary screening of weathering crust mineralization over a large area, and then carries out refined biological mineralization identification in key areas, thereby efficiently predicting and locating rare earth minerals in weathering crusts.

[0006] A multi-source remote sensing method for rare earth mineral exploration that integrates fern-based mineralization, characterized by the following steps:

[0007] (1) Acquire multi-source remote sensing data, perform atmospheric and geometric correction on the images, and extract mineralization factors, including rock-mineral, rare earth tracer elements, topography and tectonic fracture zone information;

[0008] The multi-source remote sensing data includes optical images, radar images, and DEM data;

[0009] Optical images include Sentinel-2, Landsat-8, ASTER, and GF5;

[0010] Radar imagery includes ALOS PALSAR or GF-3. ALOS PALSAR or GF-3 data are used to identify topographic and geomorphic factors and structural fracture zone factors, thereby improving the identification accuracy of weathered crust rare earth mining areas.

[0011] The extraction methods for ore-forming factors and ore-indicating factors include:

[0012] (1.1) Combining band ratio with K-means algorithm to obtain rock-mineral factors;

[0013] (1.2) The constrained energy minimization CEM algorithm is used to extract the abundance information of alteration minerals;

[0014] (1.3) Principal component analysis (PCA) was used to extract the characteristics of rare earth tracer elements;

[0015] (1.4) Analyze the slope, aspect, cutting depth and topographic variation coefficient of the DEM to obtain topographic factors;

[0016] (1.5) Construct fracture zone information using threshold detection with Sobel and Laplacian operators.

[0017] (2) After delineating the potential area of ​​the weathered crust rare earth mining area based on the mineralization factors and mineralization indicators, a drone equipped with a visible light-near infrared camera was used to fly at low altitude in the potential area to collect high-resolution images and identify the distribution of ferns by texture, morphology and spectral index.

[0018] During the low-altitude flight of the UAV, ground control points (GCPs) are deployed and orthophotos are stitched together to align the UAV images with multi-source remote sensing data in the same coordinate system.

[0019] Identifying the distribution of ferns involves the following sub-steps:

[0020] (2.1) Extract color and texture features from visible light images;

[0021] (2.2) Use spectral indices (NDVI or red edge index) to distinguish ferns from other vegetation;

[0022] (2.3) Calculate the morphological characteristics of fern leaf division degree or roundness based on object-oriented segmentation method;

[0023] (2.4) Output the distribution map of the fern plants using a machine learning classifier (SVM or RF) or a lightweight deep learning network.

[0024] (3) The obtained fern distribution is downsampled to the same resolution as other remote sensing features to form a fern coverage layer;

[0025] The downsampling step in step (3) uses neighborhood statistics, specifically including:

[0026] (3.1) Select the grid size corresponding to the main remote sensing data;

[0027] (3.2) Calculate the percentage of fern pixels in the centimeter-level image within each grid;

[0028] (3.3) Convert the percentage results into a fern coverage layer between 0 and 1.

[0029] (4) The fern coverage layer is fused with the mineralization factor to construct a multi-source feature matrix;

[0030] (5) Input the feature matrix into the deep learning model for training and prediction, output the rare earth mineralization probability map and select high potential areas.

[0031] The deep learning model is a convolutional neural network (CNN). The input of the CNN is a multi-source feature matrix: rock-mineral factors, rare earth tracer elements, topographic structural factors, and fern distribution information; the output is a mineralization probability map.

[0032] The training steps for a deep learning model include:

[0033] (5.1) Collect known mineralized points as positive samples and collect non-mineralized or background areas as negative samples;

[0034] (5.2) Register the multi-source feature matrix to match the sample labels with the pixel features;

[0035] (5.3) Set the network structure and hyperparameters, and use the cross-entropy loss function for iterative optimization;

[0036] (5.4) Evaluate the model performance based on the confusion matrix and AUC evaluation index and save the optimal model parameters.

[0037] When applying the trained deep learning model to the entire feature matrix, the mineralization probability of each pixel is output. Potential rare earth mineral areas in the weathered crust are screened based on the probability threshold, and final confirmation is made by combining ground verification.

[0038] Its core includes the following:

[0039] (1) Extraction of mineralization factors from multi-source remote sensing information

[0040] Multi-source optical remote sensing data and DEMs, including Sentinel-2, Landsat-8, ASTER, and GF5, were collected. After atmospheric and geometric corrections were completed, remote sensing characteristics of rock-mineral-structure-geomorphic mineralization factors were obtained using methods such as band ratio, principal component analysis, and alteration mineral identification, and the weathered crust rare earth mining area was preliminarily delineated.

[0041] (2) Fern monitoring by drone

[0042] Within the potential area, high-resolution imagery was acquired through low-altitude flight using drones equipped with visible-near-infrared cameras. Ferns enriched in rare earth elements (such as *Dicranopteris dichotoma* and *Dryopteris crassirhizoma*) were identified using methods including texture (gray-level co-occurrence matrix), morphology (object-oriented segmentation), and spectral indices (NDVI, RedEdgeIndex). These fern species were downsampled to a grid resolution consistent with other remote sensing features to generate a fern coverage map. The comprehensive analysis of fern distribution provided high-confidence biomineralization indicators for deep learning models.

[0043] (3) Deep learning prediction

[0044] A multi-source feature matrix is ​​constructed by combining information such as fern coverage, rock-mineral factors, alteration mineral abundance, rare earth element absorption characteristics, topographic factors, and tectonic fracture zones. A convolutional neural network (CNN) is used for training and inference to output a mineralization probability map and select high-potential areas, thereby efficiently and accurately identifying and locating rare earth deposits in weathered crusts.

[0045] A rare earth mineral exploration system for implementing the method, the system comprising:

[0046] Multi-source data processing unit: used to acquire and manage multi-source remote sensing images (optical, DEM, radar, etc.), and perform preprocessing (atmospheric correction, geometric correction, cloud cover filtering, image cropping, etc.) and extract mineralization factors and mineralization indicators;

[0047] The drone aerial photography unit includes drones equipped with visible light-near infrared cameras, ground control points (GCPs), and flight planning software for low-altitude image acquisition and fern identification in potential areas.

[0048] Feature fusion and deep learning unit: Fern cover is combined with mineralization factors such as lithology, minerals, structure, and topography to generate a multi-source feature matrix, and CNN training and inference are performed in a high-performance environment such as GPU to output a mineralization probability map;

[0049] Results Visualization and Decision Unit: Used to generate mineralization probability maps and screen high-potential areas to guide rare earth mineral exploration in weathered crusts. Using a GIS platform or WebGIS, layers such as mineralization probability, fern cover and lithological background, altered minerals, landforms, and tectonic fracture zones are displayed for interactive querying and to guide subsequent field verification and drilling deployment.

[0050] In summary, through the above methods and systems, this invention can quickly delineate potential enrichment areas of weathered crust rare earth deposits over a large area, integrate biomineralization information from ferns and mineralization factors in key areas, and improve the accuracy and efficiency of exploration by using deep learning models, thereby reducing the costs of blind ground surveys and drilling. Attached Figure Description

[0051] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0052] Figure 1 This is a schematic diagram of the process of the present invention;

[0053] Figure 2 Lithological classification and ore-bearing parent rock distribution indication map provided for the embodiments;

[0054] Figure 3 Altering mineral distribution indication map provided for the embodiments;

[0055] Figure 4 PCA analysis band indicator diagram of rare earth minerals provided for the embodiments;

[0056] Figure 5 The topographic and geomorphological features of the easily weathered crust type deposits provided in the embodiments;

[0057] Figure 6 The example provides a diagram indicating the construction of a fracture zone using convolutional kernels for edge detection.

[0058] Figure 7 A rare earth potential zone indicator map under multi-factor constraints provided for the embodiments;

[0059] Figure 8 A schematic diagram of a fern sample provided for one embodiment of the present invention;

[0060] Figure 9A schematic diagram of a rare earth mineral exploration system provided in one embodiment of the present invention. Detailed Implementation

[0061] like Figure 1 The process shown includes the following steps:

[0062] Step 1: Multi-source remote sensing data analysis (preliminary acquisition of rare earth mineralization potential areas in the weathering crust)

[0063] This step primarily utilizes multi-source remote sensing data (optical remote sensing, microwave remote sensing, thermal infrared remote sensing, hyperspectral remote sensing, etc.) and employs a series of image processing, feature extraction, classification, and multi-factor constraint methods to identify potential areas for rare earth mineralization in the weathering crust. This method systematically analyzes multiple aspects, including lithological background, alteration minerals, key rare earth tracer elements, topography, and tectonic fracture zones, to preliminarily delineate high-potential sample areas.

[0064] 1.1 Multi-source data acquisition and preprocessing

[0065] 1.1.1 Multi-source data acquisition

[0066] 1) Optical remote sensing data

[0067] Landsat-8OLI / TIRS: Data provided by the U.S. Geological Survey (USGS). The data resolution is 30m for visible-shortwave infrared (VNIR-SWIR) and 100m for thermal infrared (TIR) ​​(resampled to 30m). It is used to identify large-scale rock mass features and hydrothermal alteration zones in the study area.

[0068] ASTER: ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) is a collaborative project between the Japan Aerospace Exploration Agency (JAXA) and NASA. Launched aboard the Terra satellite on December 18, 1999, its data is available for download worldwide on NASA's EarthData platform. Covering VNIR, SWIR, and TIR bands, it can be used for lithology identification, vegetation index calculation, and mineral alteration analysis. The TIR band is sensitive to the thermal emission characteristics of different rocks, helping to differentiate between alkaline rocks, granites, and gabbro.

[0069] The Sentinel-2B MSI provides high spatial resolution for extracting spectral features of ferns and key tracer elements for REE (residue-free) extraction. The red-edge bands (705nm, 740nm, 783nm) enhance vegetation monitoring capabilities and can also monitor the absorption characteristics of the key tracer element Nd (744, 802, and 865nm bands). The SWIR bands can identify altered minerals such as kaolinite and illite. Data provided by the European Space Agency (ESA), this data has a spatial resolution of 10m in the VNIR band (443-865nm) and 20m in the SWIR band (1373-2190nm).

[0070] High-resolution AHSI hyperspectral data from GF-5 satellite provides ultra-high spectral resolution, enabling precise extraction of the rare-earth tracer element Nd. 3+ Characteristic absorption peaks. Suitable for the identification of rare earth tracer element Nd, it can be used as a supplement and verification of Sentinel-2B MSI to improve mineral exploration accuracy. Data provided by China National Space Administration, spectral resolution: visible light (5nm), shortwave infrared (10nm), spatial resolution: 30m.

[0071] 2) Radar remote sensing data

[0072] The Gaofen-3 (GF-3) SAR satellite, employing multi-polarization SAR data, can penetrate vegetation to obtain surface structural information and identify fault zones. It combines directional filters to enhance the identification of linear structures. The data was provided by the China National Space Administration. The data resolution is HH+HV polarization mode, 10m resolution, and 100km swath width.

[0073] 3) Digital Elevation Model (DEM) Data

[0074] ALOS PALSAR SAR was used to generate a DEM, with the final DEM data resolution approximately 12.5m. It is suitable for micro-topographic identification in weathered crust mining areas. Data provider: Japan Aerospace Exploration Agency (JAXA). Resolution: Single polarization (FBS): 10m, Multiple polarization (FBD): 20m, Polarization mode (PLR): 30m.

[0075] NASA DEM is used to fill data gaps and complements ALOS PALSAR SAR for calculating weathering crust development indices (slope, aspect, topographic relief, cutting depth, etc.). Combined with mineralization factors, it delineates the optimal weathering crust rare earth mineralization areas. Resolution: NASA DEM: 30m.

[0076] 1.1.2 Data Preprocessing

[0077] (1) Radiation correction

[0078] Atmospheric correction was performed using FLAASH. This removed atmospheric scattering and aerosol effects, restoring the true spectral reflectance of the Earth's surface. It also unified the spectral responses of multi-source data, improving data consistency.

[0079] (2) Geometric correction

[0080] Ensure spatial matching of multi-source remote sensing data to avoid the cumulative effect of errors on the analysis results. Use a unified coordinate system for all data, with GF5 requiring orthorectification and Landsat 8 as the reference image. Geometric calibration of GF3 data uses the ALOS PALSAR DEM for elevation.

[0081] (3) Image cropping

[0082] Image cropping is performed using the vector boundary of the study area. This removes redundant data from areas outside the study area, reducing computational load and improving processing efficiency.

[0083] (4) Cloud cover filtering

[0084] Interest region threshold filtering and masking techniques are used to remove clouds and shadows, avoiding cloud interference and ensuring data integrity.

[0085] The above steps output a high-quality, multi-source remote sensing dataset (with radiometric correction, geometric correction, cropping, and cloud cover filtering) to ensure the accuracy of subsequent analysis. DEM data, SAR data, and spectral data are cross-registered, providing a data foundation for identifying rare earth mineralization potential areas in the weathering crust.

[0086] 1.2 Lithological Background Feature Extraction Module

[0087] This module requires the use of ASTER thermal infrared band and Landsat 8 remote sensing data. By combining spectral indices, principal component analysis (PCA), and support vector machine methods, the main lithological units in the study area are identified, providing lithological background constraints for mining area identification.

[0088] 1.2.1 Band Ratio Method (BR)

[0089] Band ratio analysis can enhance the spectral characteristics of specific minerals and improve the accuracy of lithological identification. By inputting ASTER bands and calculating the resulting ratio images, the separability of different rock units can be improved.

[0090] ASTER-TIR Quartz Index (QRI): (band 10 / band 12) × (band 13 / band 12)

[0091] ASTER-TIR mafic mineral index MRI: (band 12 / band 13) × (band 13 / band 14)

[0092] ASTER-TIR mafic mineral index MRI: (band 12 / band 13) × (band 13 / band 14)

[0093] ASTER-TIR Carbonate Index (CRI): b13 / b14

[0094] ASTER-SWIR-TIR Alkali Feldspar Index (ORI): (b4×b10) / (b6×b11)

[0095] ASTER-SWIR Dolomite Index DRI: (band7 + band9) / band8

[0096] 1.2.2 Principal Component Analysis (PCA)

[0097] Principal component analysis (PCA) was performed using Landsat-8 OLI / TIRS data, selecting the full spectral bands of OLI-TIR. The PCA method in the Transform Method of ENVI software generated several principal component (PC) bands. These bands are linear combinations of the original spectral bands and are uncorrelated. The number of output principal component bands is the same as the number of input spectral bands. The first principal component (PC1) contains the largest data variance, the second principal component (PC2) contains the second largest data variance, and so on. The "Transform Parameters" were set to covariance and eigenvector. The obtained results (PC1, PC2, and PC3) were then subjected to false-color synthesis to enhance lithological characteristics.

[0098] 1.2.3 Unsupervised Classification (K-means)

[0099] (1) The bands obtained from the above band ratios, ASTER shortwave infrared bands, and principal component analysis bands are combined into a composite image containing the following lithological spectral characteristic indices and bands: QRI, MRI, FRI, CRI, ORI, DRI, B02, B04, B05, B06, B07, B08, B09, and PC1 to PC7. This is used to enhance cluster separability.

[0100] (2) Random pixel samples are extracted from the image to train the clusterer. Each sample contains pixel values ​​of all the above bands.

[0101] (3) K-Means clustering algorithm: Divide pixels into 4 lithological cluster categories. Use 50,000 randomly sampled samples from (2) as input for cluster training; apply the clustering model to the entire image to generate a classified image, customize the color of each cluster, and enhance map interpretability. The clustering formula represents finding a partitioning method among all possible clustering methods C such that the sum of the squared distances from all points to the center of their respective clusters is minimized, thus achieving a clustering effect of "high intra-cluster consistency". The formula is as follows:

[0102]

[0103] Where: argmin C : This indicates finding the optimal clustering method among all possible clustering partitions C.

[0104] K: Number of clusters, such as 4 clusters, which means it is divided into 4 lithological units.

[0105] i = 1…K: The i-th class (cluster).

[0106] C i : Belongs to the i-th cluster set (i.e., "cluster").

[0107] x∈C i : indicates that sample point x belongs to the i-th cluster.

[0108] μ i : The "mean vector" of the i-th class (i.e., the center point of all points in the cluster).

[0109] ||x-μ i || 2 : The distance from sample point x to its cluster center μ i The square of the Euclidean distance represents the "clustering error" or "density".

[0110] (4) Based on the fact that rare earth elements in weathered crusts are often enriched in parent rocks and the lithology of weathered crust types, lithological clusters related to rare earth mineralization (numbered 1 and 2) were extracted as lithological background constraints to provide a reference for subsequent rare earth mineralization prediction, such as... Figure 2 .

[0111] 1.3 Extraction of alteration mineral information

[0112] This module requires ASTER-VNIR-SWIR data as input. It extracts alteration minerals associated with rare earth mineralization in the weathering crust, including kaolinite, illite, halloysite, montmorillonite, chlorite, calcite, and hematite, to help determine the spatial distribution of rare earth mineralization in the weathering crust. Figure 3 .

[0113] 1.3.1 Spectral Matching Method

[0114] The Constrained Energy Minimization (CEM) algorithm is used for alteration mineral extraction. CEM maximizes the spectral signal of the target mineral while suppressing background noise by designing a Finite Impulse Response (FIR) filter. Its mathematical expression is as follows:

[0115] The covariance matrix of a pixel:

[0116]

[0117] x i : The spectral vector of the i-th pixel in the background.

[0118] N: Number of background sample points.

[0119] x: Background mean spectrum.

[0120] CEM filtering equation:

[0121]

[0122] Where: w is the weight vector of the CEM filter (a column vector with the same number of spectral bands);

[0123] R is the background covariance matrix, used to represent the correlation between background pixels;

[0124] d is the standard spectral vector of the target mineral (i.e., the spectral vectors of minerals such as kaolinite and illite in the spectral library);

[0125] dt: the transpose of vector d;

[0126] R -1 : The inverse of the background covariance matrix.

[0127] Target mineral pixel CEM response value calculation formula (pixel detection): Apply filter w to the spectral vector x of each pixel to obtain its spectral matching degree with the target mineral d. cem The higher the value, the higher the abundance of the target mineral.

[0128] Y cem =W T X

[0129] in:

[0130] Y cem The response value of a certain pixel represents the probability of the presence of the target mineral in that pixel;

[0131] x is the spectral vector of the pixel to be measured in ASTER or GF-5 data (i.e., its reflectance or emissivity in each band).

[0132] w T : Transpose of filter weights

[0133] Operation procedure: (1) Select standard spectral curves of kaolinite, illite, halloysite, montmorillonite, chlorite, calcite, and hematite from the USGS spectral library; (2) Perform spectral resampling to match the ASTERVNIR-SWIR band; (3) Calculate the background covariance matrix R; (4) Calculate the CEM filter weight vector w; (5) Calculate the CEM response value y of the pixel; (6) Generate the CEM grayscale image of each mineral. The higher the response value, the higher the abundance of the target mineral.

[0134] 1.3.2 Comprehensive Alteration Mineral Abundance Information

[0135] A rare earth-related alteration mineral model was constructed to obtain comprehensive alteration mineral abundance information. Based on regression coefficients, a model was constructed relating rare earth elements to Nd... 3+ The value-related alteration mineral model is as follows:

[0136] Y Nd =0.046321x1+0.019022x2+0.007328x3+0.001308x4+(-0.001450)x5+

[0137] (-0.009700)x6+(-0.040145)x7

[0138] Where Y Nd This is a comprehensive information on the abundance of altered minerals. x1, x2, x3…xn are normalized mineral characteristic values ​​(kaolinite, illite, halloysite, montmorillonite, chlorite, calcite, and hematite).

[0139] The model was used for image computation to obtain Nd 3+ Grayscale map of the overall abundance of related alteration minerals.

[0140] 1.4 REE mineral Nd 3+ Information Extraction

[0141] 1.4.1 Selection of characteristic bands

[0142] Nd 3+ Neodymium (Nd) is an important component of rare earth elements. It exhibits significant absorption characteristics in the near-infrared (VNIR) band and is a key tracer for identifying ion-adsorption type weathered crust rare earth ores. Its typical absorption band (derived from Nd) is... 3+ (Combined absorption in various altered minerals) 744nm, 802nm, and 865nm represent Nd. 3+ The main absorption peak in the VNIR band.

[0143] Using hyperspectral data (Sentinel-2B and GF5) to study Nd 3+ Absorption characteristics in the weathered layer were identified, and dimensionality was reduced by principal component analysis (PCA) to screen out areas with high abundance of direct tracer factors for rare earth mineralization in the weathered crust of the study area.

[0144] Sentinel-2B MSI (10-30m resolution): Selects VNIR bands such as B5 (703nm), B6 ​​(740nm), B7 (782nm), and B8A (865nm) to cover Nd. 3+ The main absorption characteristics.

[0145] GF-5 hyperspectral data (5-10 nm spectral resolution): Select corresponding Nd 3+ The absorbed B83 (741nm), B97 (800nm), and B112 (865nm) bands, as well as the shoulder bands B74, B90, B104, and B121, were used for feature comparison.

[0146] 1.4.2Nd 3+ Characteristic bands and principal component analysis (PCA)

[0147] Principal component analysis (PCA) is a linear dimensionality reduction method that aims to transform high-dimensional spectral data into a new set of orthogonal variables (principal components). These principal components can represent most of the information (i.e., variance) of the original data. The specific steps are as follows:

[0148] (1) Constructing the data matrix

[0149] Suppose a hyperspectral image has m pixels, and each pixel has n bands. Construct the following data matrix:

[0150]

[0151] Each row represents the spectrum of a single pixel, and each column represents a wavelength band.

[0152] (2) Centralized processing

[0153] The centering operation is to subtract this average vector from each row:

[0154]

[0155] in,

[0156] X ′ : Centralized data matrix with dimensions m×n. It is a row vector representing the average value of each band.

[0157] (3) Calculate the covariance matrix

[0158] By using the centered data of all pixels, the covariance between every two bands is calculated, thereby establishing a correlation matrix between bands and revealing the correlation between changes between bands.

[0159]

[0160] C is an n×n covariance matrix that describes the correlation between bands.

[0161] X′ is the centered data matrix with dimensions m×n.

[0162] X′ T It is the transpose of X′, with dimensions n×m.

[0163] (4) Calculate eigenvalues ​​and eigenvectors

[0164] Perform eigenvalue decomposition on the covariance matrix C to find eigenvalues ​​and eigenvectors, and extract the most useful information:

[0165] C·e i =λ i ·e i

[0166] λ i The i-th eigenvalue represents the variance contribution of the i-th principal component.

[0167] e i The corresponding eigenvector represents the weight of the original band in the linear transformation.

[0168] A total of n (λ) values ​​were found. i ,e i Yes, sort the eigenvalues ​​in descending order (λ1>λ2>...>λ). n ), for all e i According to the corresponding λ i Sort the components from largest to smallest, and select the first K principal component directions (eigenvectors) to form the transformation matrix E:

[0169] E = [e1, e2, ..., e k ]

[0170] Calculate the principal component image, and then project it onto the principal component space to reduce dimensionality and highlight the main directions of change. Project the original spectrum onto the principal component space:

[0171] Y PC =X′·E

[0172] Y PC This is the dimensionality-reduced data, a matrix of shape m×K, where each column represents the projection score on the corresponding principal component, i.e., a principal component, chosen from Nd. 3+Characteristic components whose absorption and reflection properties are consistent across the same wavelength band. For example... Figure 4 .

[0173] 1.5 Extraction of Topographic and Geomorphological Factor Information

[0174] Multiple topographic factors (slope, aspect, topographic relief, cutting depth, roughness, topographic variation coefficient, etc.) are calculated from DEM data. These factors are then combined into a comprehensive evaluation index (CSn) based on predetermined weighting coefficients. Figure 5 And a terrain mask map of “favorable weathering crust” is generated by normalization and threshold segmentation.

[0175] The detailed steps are as follows:

[0176] (1) Loading DEM data ALOS PALSAR Hi-Res Terrain Corrected

[0177] (2) Based on GIS spatial analysis technology, key terrain factors are generated and analyzed by combining DEM data and the surface analysis tools of ArcGIS software, including elevation, slope, aspect, surface cutting depth, terrain relief, roughness and terrain variation coefficient.

[0178] (3) Comprehensive Evaluation Model CSn CS n = 0.21768x1 + 0.324627x2 + 0.213371x3 + 0.326055x4 + 0.340596x5 + 0.322556x6 + 0.364317x7 where x1, x2, x3…x n These are the standardized factor values ​​(elevation, slope, aspect, surface incision depth, topographic relief, roughness, and topographic variability coefficient), CS n It is the overall score of the Nth pixel.

[0179] 1.6 Construction Information Extraction

[0180] To identify the edge features of tectonic fracture zones in remote sensing image I (SAR data), the results of Sobel and Laplacian edge detection are combined: Sobel can extract boundaries in the horizontal or vertical directions; Laplacian is sensitive to edges in all directions. Strong edge responses are selected by thresholding and considered as potential fracture-fracture zone regions, such as... Figure 6 Select ALOS / PALSAR SAR data; band: HH (horizontal transmit-horizontal receive polarization).

[0181] 1.6.1 Convolution Kernel and Edge Detection:

[0182] (1) Sobel operator

[0183] Sobel is mainly divided into Gx (detecting edges in the x-direction) and Gy (detecting edges in the y-direction), and its 3×3 convolution kernel is defined as:

[0184]

[0185] Using G x When a convolution kernel is applied to ALOS / PALSAR SAR image I, the horizontal gradient I is obtained through convolution. x Using G y When convolution kernels are applied to images, convolution yields the vertical gradient I. y :

[0186] I x =I*G x ,I y =I*G y

[0187] The overall edge strength is:

[0188]

[0189] (2) Laplacian operator

[0190] The Laplacian kernel performs second-order differentiation on the image, making it suitable for detecting edges or abrupt changes in grayscale in all directions. A 3×3 template is used as follows:

[0191]

[0192] Calculation formula: LaplacianEdge = I * L, where a value greater than 0 indicates a sudden change in grayscale, and the larger the value, the more drastic the change.

[0193] 1.6.2 Threshold Setting and Binarization

[0194] Based on experiments or experience, `sobel_threshold` and `laplacian_threshold` need to be specified to filter for significant edge pixels. In the experimental example, `sobel_threshold` = 11917; `laplacian_threshold` = 20853; regions with values ​​greater than the threshold are assigned a value of 1 and named `SobelMask` and `LapMask` respectively. If either `SobelMask` or `LapMask` is 1, it is considered a "fractured zone edge".

[0195] 1.7 Establishment of rare earth tracer element Nd under multi-factor constraints 3+ High value area

[0196] Multifactor constraint analysis (weighted superposition) to obtain Nd 3+High-value masking: This masking method incorporates lithological background, altered minerals, and Nd... 3+ The outputs from modules such as tracing, topography, and tectonic fracture zones are superimposed using multiplicative logic operations; multi-factor constraints are implemented to ultimately extract target areas with rare earth mineralization potential, thus obtaining preliminary rare earth mineralization potential zones, such as... Figure 7 .

[0197] Step 2: Use drones to precisely identify ferns and build an accurate sample library.

[0198] This step, based on the high-potential rare earth mineralization zone delineated by "multi-factor constraint analysis," deploys drones to acquire high-resolution VNIR imagery, further identifying the distribution of ferns. This imagery serves as a more precise indicator of rare earth mineralization and is then integrated with the existing feature matrix to provide the subsequent CNN model with a new feature band: "ferns."

[0199] 2.1 Precise Target Area and UAV Planning

[0200] Potential area screening: Comprehensive step 1 (lithology, alteration minerals, elemental tracer Nd) 3+ High-potential areas (such as terrain and structural fracture zones) are selected, and key areas (A, B, C, etc.) are selected for UAV flights.

[0201] Drone flight path design: Based on the area of ​​high-potential zones, terrain undulations and vegetation coverage, plan the optimal flight altitude (e.g., 100-200m) and overlap rate (forward overlap ≥70%, lateral overlap ≥60%).

[0202] Ground Marking and Coordinate Control: Ground control points (GCPs) are set up in high-potential areas to ensure the geometric accuracy of UAV imagery.

[0203] 2.2 Unmanned Aerial Vehicle Image Acquisition (VNIR)

[0204] RGB imagery: Acquire high-resolution visible light images (10cm / pixel resolution) to record the morphology, color, and texture details of vegetation. The integrated visible light image is obtained through orthographic stitching.

[0205] VNIR multispectral imagery (400–1000 nm) facilitates the extraction of vegetation indices (NDVI, PRI, etc.) and captures the spectral characteristics of ferns (red-edge shift, etc.). Atmospheric and geometric corrections are applied to finally output a spectral orthophoto image.

[0206] 2.3 Fern Identification (Texture + Morphology + Waveband Index)

[0207] In drone imagery of high-potential areas, to obtain more accurate distribution and samples of ferns, visible light RGB imagery can be combined with simple spectral indices (such as NDVI and red border), supplemented by object-level features such as texture and morphology, and then identified using machine learning methods such as Support Vector Machine (SVM). The specific steps are as follows:

[0208] (1) Feature extraction

[0209] Color characteristics: Calculate the mean values ​​of R, G, and B, as well as hue and saturation S, from the RGB or HSV space of the drone;

[0210] Texture features: Contrast, entropy, and uniformity parameters within the pixel neighborhood are extracted using GLCM (Gray-Level Co-occurrence Matrix); NDVI or RedEdgeIndex is calculated using NIR and the red edge channel;

[0211] Morphological features (object level): At high resolution (10cm / pixel), the image can be segmented into objects, and the roundness and splitting index of the region can be calculated to capture the unique feathery structure of fern leaves.

[0212] (2) Sample data collection:

[0213] Ground field survey: Select several points in the high-potential area to confirm the growth of ferns, such as... Figure 8 GPS coordinates were recorded and located in the UAV imagery; ROI labeling: On the UAV orthophoto, combined with the above coordinates and human visual identification, polygons (positive samples) were manually delineated for fern areas, and non-fern vegetation or bare land were selected as negative samples;

[0214] Generate training set: Extract the above color, texture, index, and morphological features for each labeled pixel (or object); label its category as fern (1) or non-fern (0).

[0215] (3) SVM classification

[0216] Model training: Using the labeled ROI positive and negative sample set, the classifier is trained using SVM;

[0217] Validation and parameter tuning: The Kappa coefficient was evaluated through cross-validation;

[0218] Classification output: Apply the trained SVM to the entire UAV imagery to generate a fern distribution map (1 = fern, 0 = non-fern).

[0219] (4) Downsampling and merging

[0220] Since the main feature matrix has a resolution of 10–30m, the FernMap needs to be downsampled from a centimeter-level resolution. Neighborhood statistics (majority voting or coverage calculation) can be used to create a "Fern_Cover" layer (0–1) aligned with the main features. This allows for accurate characterization of fern distribution and abundance within the potential region. Furthermore, summarizing detailed information (e.g., "Ferns account for 40% of this 10m pixel") still provides useful new information for the CNN, avoiding the need for millions of times more fragmented pixels over a large area.

[0221] (5) Results: Precise distribution and sample analysis of ferns

[0222] Fern Map: Presents the location of ferns within the potential zone in high resolution;

[0223] High-quality samples: Fern / non-fern pixels with high confidence can be exported and used as "precise samples" for CNN training in step 3;

[0224] FernCover band: compared with other features (lithology, alteration, Nd) 3+ The features (such as topography and structure) are combined to form a new feature matrix, which is used for final mineralization prediction.

[0225] 2.4 The indicative significance of ferns for rare earth mineralization

[0226] If FernMap=1 and the vegetation health is high (high VNIR vegetation index), it is easier to absorb rare earth elements, indicating greater potential.

[0227] If FernMap=0 but the multifactor mineralization index is high, it may still have moderate potential due to differences in plant species;

[0228] If FernMap=0 and the multi-factor mineralization index is low, it can be considered a low-potential exploration area.

[0229] Final output: A high-resolution "fern distribution band" from the drone, forming a new mask or coverage map for CNN model training.

[0230] Step 3: CNN Modeling and Identification of High-Potential Rare Earth Mineralization Zones

[0231] After integrating the "mineralization multi-factor features", the feature matrix adds the "Fern_Cover" band.

[0232] 3.1 Feature Matrix Integration

[0233] Existing multi-factor characteristics (lithology, altered minerals, Nd) 3+Absorption, topographic CSn, tectonic fracture zone, etc.); New fern band: Fern_Cover (0~1), including health index (NDVI_Fern).

[0234] 3.2 CNN Model Training

[0235] Training samples include positive and negative samples. Positive samples: samples from known rare earth deposits or Nd... 3+ Sampling in high-value areas and fern-covered areas; negative samples: sampling in non-mineralized areas or background areas;

[0236] The input CNN structure for each pixel treats the geographic grid data as a multi-channel image and employs 2D convolution. Output: mineralization probability (0-1). The training process sets parameters such as learning rate, batch size, and number of iterations. Finally, AUC and confusion matrix accuracy are used for evaluation.

[0237] 3.3 Application Model and High-Potential Area Extraction

[0238] The trained CNN model is input to predict the feature matrix of the entire area; the mineralization probability map is output: each pixel gets a probability value; threshold segmentation is used to determine "high-medium-low potential areas"; if the mineralization probability is high and the fern coverage is high, then ground drilling sampling is prioritized.

[0239] A rare earth mineral exploration system based on the fusion of fern and multi-source remote sensing data, such as... Figure 9 :

[0240] (a) The mineral exploration system includes:

[0241] 1. Multi-source remote sensing data fusion and initial screening: lithological background, alteration minerals, Nd 3+ By combining multiple factors such as elemental tracing, topography, and tectonic fracture zones, the potential area for rare earth mineralization has been preliminarily delineated.

[0242] 2. High-precision identification of ferns by drones: Deploy drones in the initially delineated high-potential areas to identify the distribution of ferns as a "bioindicator"; output "Fern_Cover" to provide more accurate training samples and feature bands for subsequent models.

[0243] 3. CNN Modeling and Terminal Prediction: Fern_Cover is merged with other multi-factor features to train a CNN; the final mineralization probability prediction of the study area is performed, the rare earth mineralization probability map is output, and the high mineralization potential area is finally delineated.

[0244] (II) System Functional Modules and Equipment

[0245] 1. Multi-source remote sensing data analysis module

[0246] Functions: Receives various remote sensing data such as optical, microwave, thermal infrared, hyperspectral, and DEM data, and performs preprocessing, feature extraction, and preliminary delineation of potential areas.

[0247] Main equipment / software:

[0248] (1) Data acquisition workstation

[0249] Used for downloading and managing data from official platforms such as Landsat-8, ASTER, Sentinel-2B, GF-5, ALOS PALSAR, and GF-3SAR. Storage disk ≥ 10TB for storing multi-temporal imagery.

[0250] (2) Data preprocessing server

[0251] It has atmospheric correction tools (such as ENVI / IDL+FLAASH, or Python) and image processing software (ArcGIS / QGIS).

[0252] The choice between GPU or CPU clusters depends on the project size and can process a large number of images in parallel.

[0253] (3) Feature extraction and analysis software

[0254] ENVI / IDL, ArcGIS, and Python are used to implement: band ratio (BR), PCA (principal component analysis), K-Means (unsupervised classification), CEM (alteration mineral extraction), DEM-derived topographic analysis (slope, aspect, dissection, etc.), and tectonic fracture zone extraction (Sobel + Laplacian edge detection).

[0255] (4) Output of Results

[0256] Generate a potential area raster to provide information for the next step of drone planning and deployment.

[0257] 2. High-resolution monitoring and fern identification module for unmanned aerial vehicles (UAVs)

[0258] Drones were deployed in the potential area to collect cm-level RGB / VNIR images, and ferns were identified using methods such as texture, morphology, and simplified spectroscopy. The images were then downsampled to form the Fern_Cover band, which serves as a precise indicator for CNN training.

[0259] Main equipment / software:

[0260] (1) Unmanned aerial vehicle platform: quadcopter or fixed-wing unmanned aerial vehicle, equipped with a high-resolution RGB camera and a multispectral (400-1000nm) sensor; configuration includes resolution: 10cm / pixel, flight time: ≥30 minutes; payload: multispectral camera (including visible light RGB, NIR or red edge band); with real-time image transmission or waypoint planning function.

[0261] (2) Ground control points (GCPs) and surveying equipment

[0262] RTK GPS or total station is used to calibrate ground coordinates; several GCPs are distributed to ensure that the image error after orthophoto stitching is ≤1 pixel.

[0263] UAV Ground Station (GCS) and Image Processing: Plan flight paths (altitude, overlap rate), manage flight missions, receive telemetry data, and complete Orthomosaic stitching. Output orthophotos (RGB + multispectral channels).

[0264] (3) Fern identification module

[0265] Feature extraction: color (RGB / HSV), texture (GLCM), shape (object-oriented segmentation), spectral index (NDVI, RedEdgeIndex);

[0266] Supervised learning classifier: SVM / RF (CPU environment is sufficient), generates FernMap (resolution 3~10cm), 1 = ferns, 0 = others;

[0267] Downsampling statistics: The cm-level FernMap is merged into a 10m~30m “FernCover” (0~1).

[0268] Outputs: (a) High-resolution fern distribution map (FernMap); (b) Downsampled Fern_Cover, which is then incorporated into the CNN feature matrix; (c) Precise fern sample points (e.g., high-confidence ROIs) for ground validation or CNN training.

[0269] 3. CNN Deep Learning Prediction Module

[0270] Combine "Fern_Cover" with the aforementioned multi-factors (lithology, alteration, Nd) 3+ The final feature matrix is ​​formed by fusing data from topography, structure, etc.; a CNN model is trained to output the rare earth mineralization probability distribution and identify high-potential areas.

[0271] Main equipment / software:

[0272] (1) High-performance computing equipment: GPU servers / workstations (such as NVIDIA RTX / Tesla series, 16-64GB video memory); used for deep learning training and inference.

[0273] (2) Deep learning framework: Python environment for constructing 2D convolutions; encapsulates geographic raster processing and feature matrix generator to handle all bands (lithological index, CEM mineral abundance, Nd...). 3+ PCA, terrain CSN, fracture zone mask, Fern_Cover, etc.) are aligned and stacked to form N+1 channels (10~30m grid);

[0274] (3) CNN training subsystem: collect positive and negative samples (known mineralization points vs. background); design network structure (input (N+1) channels); adjust hyperparameters such as learning rate and batch size; evaluate performance indicators such as AUC and confusion matrix; inference and result visualization, input global features into the trained CNN and output mineralization probability (0~1); extract high potential areas based on threshold (e.g., p>0.6); render and overlay fern distribution, actual geological map, etc. in GIS for expert decision-making.

[0275] 4. Other auxiliary modules and equipment

[0276] (1) Data storage management: RAID disks are used to store multi-temporal images, DEM, UAV orthophotos, feature matrices, etc.; a capacity of ≥10TB is recommended.

[0277] (2) Front-end UI and GIS visualization, WebGIS (such as Geoserver+Leaflet) or Desktop GIS (ArcGIS / QGIS), display potential area map, FernCover distribution, mineralization probability map, etc.;

[0278] (3) Interactive query, legend, threshold adjustment and other functions.

[0279] (4) Ground verification equipment: XRF analyzer, drilling equipment or surface sampling tools are used to verify the CNN prediction results in the field and form an exploration closed loop.

Claims

1. A multi-source remote sensing method for rare earth mineral exploration that integrates fern ore-indicating methods, characterized in that, The preliminary screening of weathering crust mineralization over a large area is conducted using remote sensing features of "rock-mineral-structure-topography" mineralization factors. Then, refined biomineralization identification is carried out in key areas to predict and locate rare earth minerals in the weathering crust. This includes the following steps: (1) Acquire multi-source remote sensing data, perform atmospheric and geometric correction on the images, and extract mineralization factors, including rock-mineral, rare earth tracer elements, topography and tectonic fracture zone information; The extraction methods for ore-forming and mineralization-indicating factors in step (1) include: (1.1) The rock-mineral factor is obtained by combining the band ratio with the K-means algorithm; (1.2) The constrained energy minimization CEM algorithm is used to extract the abundance information of alteration minerals; (1.3) Principal component analysis (PCA) was used to extract the characteristics of rare earth tracer elements; (1.4) Analyze the slope, aspect, cutting depth and topographic variation coefficient of the DEM to obtain topographic factors; (1.5) Constructing fracture zone information using threshold detection with Sobel and Laplacian operators; Multifactor constraint analysis to obtain Nd 3+ High-value masking: This masking method incorporates lithological background, altered minerals, and Nd... 3+ The outputs of the tracing, topography, and structural fracture zone modules are superimposed using multiplicative logic operations; multi-factor constraints are implemented to ultimately extract target areas with rare earth mineralization potential and obtain preliminary rare earth mineralization potential areas. (2) After delineating the potential area of ​​the weathered crust rare earth mining area based on the mineralization factors and mineralization indicators, a drone equipped with a visible light-near infrared camera was used to fly at low altitude in the potential area to collect high-resolution images and identify the distribution of ferns by texture, morphology and spectral index. This step is based on the high potential area of ​​rare earth mineralization in the weathered crust identified by "multi-factor constraint analysis". UAVs are deployed to acquire high-resolution VNIR images to further identify the distribution of ferns, which are used as a more accurate indicator of rare earth mineralization. Then, they are integrated with the existing feature matrix to provide new "fern" feature bands for subsequent CNN models. (3) The obtained fern distribution is downsampled to the same resolution as other remote sensing features to form a fern coverage layer; The downsampling step in step (3) uses neighborhood statistics, specifically including: (3.1) Select the grid size corresponding to the main remote sensing data; (3.2) Calculate the percentage of fern pixels in the centimeter-level image within each grid; (3.3) Convert the percentage results into a fern coverage layer between 0 and 1; (4) The fern coverage layer is fused with the mineralization factor to construct a multi-source feature matrix; (5) Input the feature matrix into the deep learning model for training and prediction, output the rare earth mineralization probability map and select high potential areas; The deep learning model is a convolutional neural network (CNN). The input of the CNN is a multi-source feature matrix: rock-mineral factors, rare earth tracer elements, topographic structural factors, and fern distribution information; the output is a mineralization probability map.

2. The multi-source remote sensing method for rare earth mineral exploration integrating fern ore indication as described in claim 1, characterized in that, The multi-source remote sensing data includes optical images, radar images, and DEM data; Optical images include Sentinel-2, Landsat-8, ASTER, and GF5; Radar imagery includes ALOS PALSAR or GF-3. ALOS PALSAR or GF-3 data are used to identify topographic and geomorphic factors and structural fracture zone factors, thereby improving the identification accuracy of weathered crust rare earth mining areas.

3. The multi-source remote sensing method for rare earth mineral exploration integrating fern ore indication as described in claim 1, characterized in that, Step (2) in identifying the distribution of ferns includes the following sub-steps: (2.1) Extract color and texture features from visible light images; (2.2) Use spectral indices to distinguish ferns from other vegetation; (2.3) Calculate the morphological characteristics of fern leaf division degree or roundness based on object-oriented segmentation method; (2.4) Output the distribution map of the fern plants using a machine learning classifier or a lightweight deep learning network.

4. The multi-source remote sensing method for rare earth mineral exploration based on fern ore indication as described in claim 1, characterized in that: In step (2), during the low-altitude flight of the UAV, ground control points (GCPs) are set up and orthophotos are stitched together to align the UAV images with the multi-source remote sensing data in the same coordinate system.

5. The multi-source remote sensing method for rare earth mineral exploration based on fern ore indication as described in claim 1, characterized in that: The training steps for the deep learning model in step (5) include: (5.1) Collect known mineralized points as positive samples and collect non-mineralized or background areas as negative samples; (5.2) Register the multi-source feature matrix to match the sample labels with the pixel features; (5.3) Set the network structure and hyperparameters, and use the cross-entropy loss function for iterative optimization; (5.4) Evaluate the model performance based on the confusion matrix and AUC evaluation index and save the optimal model parameters.

6. The multi-source remote sensing method for rare earth mineral exploration based on fern ore indication as described in claim 1, characterized in that: In step (5), when applying the trained deep learning model to the entire feature matrix, the mineralization probability of each pixel is output, potential weathered crust rare earth mining areas are screened according to the probability threshold, and final confirmation is made in combination with ground verification.

7. A multi-source remote sensing rare earth mineral exploration system integrating fern phytoreception, used to implement the multi-source remote sensing rare earth mineral exploration method integrating fern phytoreception as described in any one of claims 1 to 6, characterized in that, The system includes: A multi-source data processing unit is used to acquire and preprocess optical images, radar images, and DEM data, and extract mineralization factors and mineralization indicators. The drone aerial photography unit is used to acquire low-altitude images and identify ferns in potential areas; The feature fusion and deep learning unit constructs a multi-source feature matrix by combining fern coverage with rock-mineral factors, rare earth tracer elements, and topographic structural factors, and then trains a CNN model. Results output and visualization units are used to generate mineralization probability maps and screen high-potential areas to guide rare earth mineral exploration in weathered crusts.

Images