Method for extracting alteration anomaly based on remote sensing basic model latent space background deduction

Abstract

The application discloses an altered anomaly extraction method based on remote sensing basic model latent space background deduction, and the method takes a high-dimensional embedded vector field obtained by encoding multi-source remote sensing images through a remote sensing basic model as input, solves the problem of "absolute similarity failure" in a homogeneous coverage area through double reference vectors, constructs a deposit reference vector by using the embedded vector of a known deposit center pixel, constructs a background reference vector by randomly sampling a plurality of pixels in a to-be-measured area and taking the median along each embedded dimension, calculates the cosine similarity of the to-be-measured embedded vector and the two reference vectors pixel by pixel, takes the difference to obtain a relative anomaly score, takes the maximum value of the relative anomaly scores of each deposit in the case of multiple deposits, binarizes the pixel-by-pixel anomaly by using a preset threshold, aggregates the anomaly to a downstream grid, and obtains a continuous altered anomaly layer which is spatially aligned with a traditional evidence layer.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence interdisciplinary technology, specifically to a method for extracting alteration anomalies based on latent space background subtraction from a remote sensing basic model. Background Technology

[0002] Alteration anomaly extraction is a core component of remote sensing exploration for metallic mineral deposits. In hydrothermal mineral systems, the reaction between hydrothermal fluids and surrounding rocks creates spatial zoning of different alteration mineral assemblages. The corresponding remote sensing alteration indicators include, but are not limited to, iron staining, hydroxyl groups, and chlorite.

[0003] Traditional alteration anomaly extraction mainly relies on physically interpretable spectral band ratios or principal component transformations: band ratios or Crosta principal component transformations based on multispectral images such as Landsat, ASTER, and Sentinel-2. The fundamental constraint of such methods is that they depend on prior human knowledge of spectral physical characteristics, and each type of alteration mineral requires a separately designed exponential algorithm.

[0004] In recent years, remote sensing foundation models have emerged, learning surface representations from massive amounts of multi-source remote sensing imagery through self-supervised pre-training. These models encode annual time-series images from multiple sensors into fixed-dimensional embedding vector fields, covering the entire global landmass. A representative model is AlphaEarthFoundations, released by Google DeepMind in 2025. Its embedding vectors encode multiple spatiotemporally coupled information such as land cover type, lithological differences, and alteration mineral assemblages in a highly entangled manner, and can be considered "machine-discovered surface features," breaking through the information scope covered by manually defined indices.

[0005] However, directly using vectors embedded in remote sensing basic models for alteration anomaly extraction faces the following fundamental technical obstacles: Firstly, there is the problem of "absolute similarity failure": In areas with relatively uniform land cover, the surface reflectance composition is highly similar over a large area. Known mineral deposit pixels and ordinary background pixels share a large proportion of homogeneous background components in the high-dimensional embedding space, resulting in naturally high cosine similarity. Absolute similarity measures cannot distinguish between "similar because they are both of the same type of land cover" and "similar because they both contain alteration," making queries based on absolute similarity unable to reliably distinguish alteration anomalies.

[0006] Secondly, there is the issue of reference deposit dependence: existing mineral exploration attempts based on embedding vector similarity generally require the prior identification of specific reference deposit types. Reference vectors need to be constructed separately for different alteration patterns, which is complex and has limited ability to be generalized to unknown types of deposits.

[0007] Third, there is the problem of threshold reliance on experience: the threshold for judging absolute similarity lacks geological a priori or geometric basis, and needs to be empirically adjusted for each region and each type of alteration, making it difficult to form a generalizable engineering standard.

[0008] Fourth, the problem of output scale mismatch: The original spatial resolution of the remote sensing basic model is usually much higher than the raster resolution commonly used in the downstream mineralization prediction model. The pixel-by-pixel results are difficult to directly enter the mineralization prediction workflow. A controllable spatial aggregation method is needed to convert the original resolution results to the target grid.

[0009] The aforementioned problems have prevented the full realization of the potential of remote sensing basic models in mineral exploration, and there is an urgent need for a new method to transform the basic models from "absolute similarity query tools" into "relative anomaly extractors".

[0010] To address the above problems, this invention provides an alteration anomaly extraction method based on latent space background subtraction of remote sensing basic model. This method solves the following problems in existing alteration anomaly extraction methods based on embedded vectors of remote sensing basic model in homogeneous coverage areas: (1) the absolute similarity measurement cannot distinguish between "similar surface background" and "similar alteration signal", resulting in failure; (2) the requirement to pre-calibrate specific reference deposit types, resulting in limited generalization ability; (3) the lack of geometric basis for the discrimination threshold, requiring empirical adjustment on a zone-by-zone basis; and (4) the mismatch between the original resolution and the downstream mineralization prediction grid resolution. Summary of the Invention

[0011] The purpose of this invention is to provide a method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model. The specific steps of the method are as follows: S1. Input the multi-source remote sensing images covering the area to be tested into the pre-trained remote sensing basic model encoder, and encode each native pixel into a D-dimensional embedding vector to form a D-dimensional embedding vector field that corresponds one-to-one with the spatial range of the area to be tested. S2. Parallel construction of deposit reference vector and background reference vector: Two types of reference vectors are constructed simultaneously on the embedded vector field, including deposit reference vector and background reference vector. S3. Pixel-by-pixel calculation of dual similarity and relative anomaly: nbg pixels are randomly sampled in the effective working area of ​​the region to be tested, and the median is taken along D embedding dimensions to obtain a single D-dimensional background reference vector vbg. S4. Anomaly Threshold Determination and Binarization: For each native pixel x in the area to be tested, calculate the cosine similarity St,k(x) between its embedding vector e(x) and the reference vectors {vore, k} of each ore deposit, and the cosine similarity Sb(x) with the background reference vector vbg. Define the relative anomaly score: Ar(x) = maxk = 1, ..., K ( St, k(x) − Sb(x) ); S5. Abnormal density aggregation to downstream grid: Binarize Ar pixel by pixel with a preset threshold τ to obtain an abnormal binary map of the original resolution; S6. Integrating the anomaly layer into the downstream mineralization prediction system: After dividing the anomaly binary map according to the target grid size, the proportion of anomaly pixels to valid pixels in each grid is calculated: AE-EPS = Nanomaly(g) / Nvalid(g), resulting in a continuous variable anomaly density layer spatially aligned with the downstream evidence layer.

[0012] Furthermore, the constraints satisfied by the remote sensing base model in S1 include the ability to encode multi-source spatiotemporally heterogeneous remote sensing images into embedding vectors of a unified dimension, and similar land features exhibiting semantic similarity in the embedding space.

[0013] Furthermore, the construction of the ore deposit reference vector in S2 includes selecting K (K ≥ 1) known ore deposit center pixel positions and extracting the corresponding D-dimensional embedding vectors. When K ≥ 2, the embedding vectors of the K ore deposits are retained as independent ore deposit reference vectors {vore, 1, vore, 2, ..., vore, K}, rather than averaging them into a single reference vector.

[0014] Furthermore, the background reference vector construction in S2 includes randomly sampling nbg pixels in the effective working area of ​​the region to be tested, extracting an nbg × D embedding vector sample matrix, calculating the median along each of the D embedding dimensions, and obtaining a single D-dimensional background reference vector vbg.

[0015] Furthermore, the specific steps of S3 are as follows: S31. For each pixel location x in the region to be tested, calculate the cosine similarity between the embedding vector e(x) ∈ RD and the reference vectors {vore, k} of each ore deposit and the background reference vector vbg: Cosine similarity is defined as the inner product of two vectors divided by the product of their magnitudes: The relative anomaly score Ar(x) of pixel x is defined as the difference between the target similarity and the background similarity. In the case of multiple deposits, the maximum value of Ar for each deposit is taken:

[0016] The anomaly threshold τ of S4 is directly given by the normalized geometric relation of the high-dimensional latent space rather than relying on subjective experience, and its geometric meaning is as follows: S41.1 In the normalized D-dimensional embedding space, the similarity noise level introduced by the floating-point operation precision is much smaller than τ, so that the pixels identified as anomalous have considerable real differences, while τ is not too large to miss weak anomalies. S41.2. For each pixel x, determine whether it is an abnormal pixel using the following formula: The specific value of the threshold τ is determined by the spatially normalized geometric characteristics of the remote sensing basic model, the coverage type of the area to be measured, and the typical intensity of the target alteration signal.

[0017] Furthermore, the specific steps of S5 are as follows: S51. Divide the area to be measured into several grids according to the target grid size G×G. The target grid size G should be determined according to the grid resolution of the downstream mineralization prediction model. S52. For each grid g, count the number of outlier pixels Nanomaly(g) and the number of valid pixels Nvalid(g) within that grid. S53. Calculate the proportion of anomalies to valid pixels to obtain the anomaly density score AE-EPS for the grid:

[0018] Furthermore, in S4, when the number of deposits K ≥ 2, the relative anomaly score Ar(x) in the case of multiple deposits takes the maximum value of K deposits instead of the average value.

[0019] Furthermore, cosine similarity includes target cosine similarity S_t,k(x)=cos(e(x), v_ore,k) (k=1,…,K) and background cosine similarity S_b(x)=cos(e(x), v_bg), both of which are defined as cos(u, v) = (u·v) / (‖u‖2·‖v‖2), where u·v is the inner product of the two vectors and ‖·‖2 is the L2 norm.

[0020] The present invention has the following beneficial effects: The present invention effectively solves the problem of "absolute similarity failure": by introducing background reference vector and subtraction operation, the remote sensing basic model is transformed from "absolute similarity query tool" to "relative anomaly extractor", which can still reliably identify alteration anomalies in areas with relatively uniform land cover types.

[0021] This invention eliminates the dependence on the physical meaning inversion of spectral bands: This invention does not require physically interpretable calculations (such as ratios, principal component transformations) on the original spectral bands, but performs calculations directly in the high-dimensional latent space of the remote sensing basic model. Therefore, it can simultaneously respond to multiple entangled information such as alteration mineral assemblage, lithological differences, and land cover types, and has a richer dimension of alteration index information compared to traditional single-target alteration.

[0022] The background reference vector data-driven construction of this invention automatically obtains the background reference vector from random sampling of the entire region by taking the median along D embedding dimensions. This eliminates the need for manually specifying the background region or manually calibrating the background threshold, and has extremely low dependence on human intervention.

[0023] The threshold of this invention is given by geometric relations: the outlier threshold is directly given by the normalized geometric relations of the high-dimensional latent space, which has strong portability for engineering application.

[0024] This invention supports arbitrary alteration patterns and arbitrary mineral types: only the center pixel positions of K ≥ 1 known mineral deposits are required to construct a mineral deposit reference vector. The same algorithm process can be used for different alteration patterns and different mineral types. The design of taking the maximum value of multiple mineral deposits naturally supports the anomaly extraction of the union of multiple types of mineral deposits.

[0025] The core of this invention lies in a complete operator chain of "dual reference vector construction - pixel-by-pixel dual similarity calculation - relative anomaly score background subtraction - multi-deposit maximum value - density aggregation". Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram comparing the HDBS principle of the present invention. Detailed Implementation

[0027] This invention provides a method for extracting alteration anomalies based on latent space background subtraction from a remote sensing basic model.

[0028] Example 1 uses the extraction of alteration anomalies in the East Tianshan verification area (90°30′—93°10′E, 41°40′—42°50′N, with a total area of ​​33,760 km²) of the Central Asian porphyry copper deposit metallogenic domain as an example to illustrate the specific implementation steps.

[0029] S1. Multi-source remote sensing imagery encoded into a high-dimensional embedding vector field: Using Google Earth Engine and the AlphaEarth Foundations model, the annual time-series images of the Sentinel-2, Sentinel-1, and Landsat sensors covering the eastern Tianshan prediction area were encoded into a 64-dimensional embedding vector field. AlphaEarth Foundations used an STP encoder to uniformly compress the multi-sensor images into a 64-dimensional embedding space with a native resolution of 10 m. A total of 3.376 × 10 m images were obtained for the eastern Tianshan prediction area. 9 A 64-dimensional embedding vector for each native cell; S2. Construction of Deposit Reference Vectors and Background Reference Vectors: 64-dimensional embedding vectors were extracted from the central pixels of 10 known porphyry copper deposits (Tuwu, Yandong, Yanxi, Yuhai, Dong'er District, Chihu, Fuxing, Linglong, Dananhu South, and Partagxi), and 10 deposit reference vectors were constructed accordingly. 10,000 pixels were randomly sampled from the effective working area of ​​the prediction zone (excluding water bodies and built-up areas), and a 10,000 × 64 pixel-dimensional matrix was extracted. The median was calculated along each of the 64 embedding dimensions to obtain a 64-dimensional background reference vector. S3. Double similarity calculation and relative anomaly score: For each native pixel, the target similarity and background similarity of 10 ore deposits are calculated, and the maximum value of the relative anomaly score of each ore deposit is taken. The actual test shows that the absolute cosine similarity between the known ore deposit pixels and the background pixels is generally above 0.85. However, after background subtraction, the relative anomaly score of the ore deposit pixels remains significantly positive, while the relative anomaly score of the background pixels approaches zero. Alteration and background are stably distinguished. S4. Threshold binarization: Take τ = 0.02 to binarize the relative anomaly score pixel by pixel to obtain the original 10 m resolution anomaly binary map. The value of τ is an engineering choice for the eastern Tianshan region, the AlphaEarth Foundations model used, and the porphyry copper deposit target alteration. It should be specifically determined for different remote sensing basic models, different cover types, and different target alterations. S5. Anomalous density aggregation to a 500m grid: The prediction area is divided into a grid of 500 m × 500 m. The anomalous density AE-EPS is calculated for each grid according to formula (2-6). The total number of grids in the prediction area is 246,654, of which 239,857 are effective (accounting for 97.24%). Finally, a continuous variable AE-EPS layer is obtained that is spatially aligned with other evidence layers of the downstream mineralization prediction model. S6. Layer Integration into Downstream Metallogenic Prediction System: The AE-EPS layer, as a novel evidence layer, is combined with 11 traditional evidence layers (metallogenic favorable strata, magmatic rock contact zone, Bouguer gravity anomaly, aeromagnetic anomaly, deep faults, fault density, Cu geochemical anomaly, Mo geochemical anomaly, remote sensing iron staining anomaly, remote sensing hydroxyl anomaly, and remote sensing chlorite alteration anomaly) and fed into the downstream metallogenic prediction model. The importance analysis of the downstream model's permutation features shows that the AE-EPS layer ranks among the top, complementing the traditional iron staining, hydroxyl, and chlorite three types of remote sensing alteration indices. It effectively supplements the multi-source coupling signals that traditional indices cannot capture, verifying the engineering value of the AE-EPS layer.

[0030] This invention is not limited to AlphaEarth Foundations as a remote sensing foundation model, but can be extended to any other remote sensing foundation model that can encode multi-source remote sensing images into a unified dimension embedding vector and whose similar ground features have similar semantics, including future versions of AlphaEarth Foundations or remote sensing foundation models released by other vendors; the embedding dimension D is determined by the architecture of the remote sensing foundation model adopted.

[0031] This invention is not limited to porphyry copper deposits, but can be extended to the extraction of remote sensing alteration anomalies in any metallic or non-metallic mineral (such as gold, silver, lead-zinc, rare earth minerals, etc.) with hydrothermal alteration as a key indicator. It can also be extended to non-mineral application scenarios (such as ecological surface anomaly monitoring, detection of specific land cover changes, etc.), as long as the two basic premises of "the existence of a known reference sample center pixel" and "most pixels in the area to be measured are of homogeneous background" are met.

[0032] In different application scenarios of this invention, the following implementation details should be adapted rather than using the specific values ​​of this embodiment: (1) Background sampling amount nbg: should be based on the embedding dimension D to ensure statistical robustness; (2) Anomaly threshold τ: (3) should be specifically determined based on the normalized geometric characteristics of the embedding space of the remote sensing base model used, the coverage type of the area to be measured and the typical intensity of the target alteration signal; Target grid size G: (4) should be determined based on the grid resolution of the downstream mineralization prediction model.

[0033] Although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims are still within the scope of protection of the present invention.

Claims

1. A method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model, characterized in that: The specific steps of the method are as follows: S1. Input the multi-source remote sensing images covering the area to be tested into the pre-trained remote sensing basic model encoder, and encode each native pixel into a D-dimensional embedding vector to form a D-dimensional embedding vector field that corresponds one-to-one with the spatial range of the area to be tested. S2. Parallel construction of deposit reference vector and background reference vector: Two types of reference vectors are constructed simultaneously on the embedded vector field, including deposit reference vector and background reference vector. S3. Pixel-by-pixel calculation of dual similarity and relative anomaly: nbg pixels are randomly sampled in the effective working area of ​​the region to be tested, and the median is taken along D embedding dimensions to obtain a single D-dimensional background reference vector vbg. S4. Anomaly Threshold Determination and Binarization: For each native pixel x in the area to be tested, calculate the cosine similarity St,k(x) between its embedding vector e(x) and the reference vectors {vore, k} of each ore deposit, and the cosine similarity Sb(x) with the background reference vector vbg. Define the relative anomaly score: Ar(x) = maxk = 1, ..., K ( St, k(x) − Sb(x) ); S5. Abnormal density aggregation to downstream grid: Binarize Ar pixel by pixel with a preset threshold τ to obtain an abnormal binary map of the original resolution; S6. Integrating the anomaly layer into the downstream mineralization prediction system: After dividing the anomaly binary map according to the target grid size, the proportion of anomaly pixels to valid pixels in each grid is calculated: AE-EPS = Nanomaly(g) / Nvalid(g), resulting in a continuous variable anomaly density layer spatially aligned with the downstream evidence layer.

2. The method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model as described in claim 1, characterized in that: The constraints satisfied by the remote sensing basic model in S1 include the ability to encode multi-source spatiotemporally heterogeneous remote sensing images into embedding vectors of a unified dimension, and similar land features exhibiting semantic similarity in the embedding space.

3. The method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model as described in claim 1, characterized in that: The construction of the ore deposit reference vector in S2 includes selecting K (K ≥ 1) known ore deposit center pixel positions and extracting the corresponding D-dimensional embedding vectors. When K ≥ 2, the embedding vectors of the K ore deposits are retained as independent ore deposit reference vectors {vore, 1, vore, 2, ..., vore, K}, instead of averaging them into a single reference vector.

4. The method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model as described in claim 1, characterized in that: The background reference vector construction in S2 includes randomly sampling nbg pixels in the effective working area of ​​the region to be tested, extracting an nbg × D embedding vector sample matrix, calculating the median along each of the D embedding dimensions, and obtaining a single D-dimensional background reference vector vbg.

5. The method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model as described in claim 1, characterized in that: The specific steps of S3 are as follows: S31. For each pixel location x in the region to be tested, calculate the cosine similarity between the embedding vector e(x) ∈ RD and the reference vectors {vore, k} of each ore deposit and the background reference vector vbg: Cosine similarity is defined as the inner product of two vectors divided by the product of their magnitudes: The relative anomaly score Ar(x) of pixel x is defined as the difference between the target similarity and the background similarity. In the case of multiple deposits, the maximum value of Ar for each deposit is taken:

6. The method for extracting alteration anomalies based on latent space background subtraction using a remote sensing fundamental model as described in claim 1, characterized in that: The anomaly threshold τ of S4 is directly given by the normalized geometric relation of the high-dimensional latent space rather than relying on subjective experience, and its geometric meaning is as follows: S41.1 In the normalized D-dimensional embedding space, the similarity noise level introduced by the floating-point operation precision is much smaller than τ, so that the pixels identified as anomalous have considerable real differences, while τ is not too large to miss weak anomalies. S41.

2. For each pixel x, determine whether it is an abnormal pixel using the following formula: The specific value of the threshold τ is determined by the spatially normalized geometric characteristics of the remote sensing basic model, the coverage type of the area to be measured, and the typical intensity of the target alteration signal.

7. The method for extracting alteration anomalies based on latent space background subtraction using a remote sensing fundamental model as described in claim 1, characterized in that: The specific steps of S5 are as follows: S51. Divide the area to be measured into several grids according to the target grid size G×G. The target grid size G should be determined according to the grid resolution of the downstream mineralization prediction model. S52. For each grid g, count the number of outlier pixels Nanomaly(g) and the number of valid pixels Nvalid(g) within that grid. S53. Calculate the proportion of anomalies to valid pixels to obtain the anomaly density score AE-EPS for the grid:

8. The method for extracting alteration anomalies based on latent space background subtraction from a remote sensing fundamental model as described in claim 1, characterized in that: In S4, when the number of deposits K ≥ 2, the relative anomaly score Ar(x) in the case of multiple deposits takes the maximum value of K deposits instead of the average value.

9. The method for extracting alteration anomalies based on latent space background subtraction using a remote sensing fundamental model as described in claim 1, characterized in that: The cosine similarity in S4 includes target cosine similarity S_t,k(x)=cos(e(x), v_ore,k) (k=1,…,K) and background cosine similarity S_b(x)=cos(e(x), v_bg), both of which are defined as cos(u, v) = (u·v) / (‖u‖2·‖v‖2), where u·v is the inner product of two vectors and ‖·‖2 is the L2 norm.

Images