A lithology fine classification method based on spectrum-space-texture multi-dimensional feature fusion

Abstract

The application provides a lithology fine classification method based on spectrum-space-texture multi-dimensional feature fusion, and belongs to the field of lithology analysis and geological information identification.The method realizes synchronous sampling and dimension alignment of spectral and texture features in remote sensing images through a self-designed multi-scale two-dimensional spectrum-texture entanglement coding module, dynamically weighted fusion under multi-scale is combined with a cross-gate mechanism, and lithology fine differences are accurately mined.A spatial-neighborhood adaptive attention mechanism is introduced, a global spatial context is established for lithology patches and their boundaries, an attention weight is dynamically adjusted, target-related spatial connections are enhanced, and interference responses are inhibited.A high-dimensional tensor fusion algorithm based on collaborative compensation is proposed, multi-feature unified fusion is realized through domain projection and nonlinear compensation mapping, and high-confidence discriminant information is selected through selective enhancement and sparse activation.

Description

Technical Field

[0001] This invention belongs to the field of lithological analysis and geological information identification, and more specifically relates to a lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features. Background Technology

[0002] In recent years, with the continuous improvement of remote sensing image resolution and the rapid development of geological exploration needs, how to accurately identify and automatically interpret surface lithology information based on multi-source remote sensing data has become one of the core issues in mineral resource surveys and geological environment analysis. Traditional lithology classification methods are mostly based on manual interpretation, statistical analysis, or single spectral features, which are difficult to fully capture the diversity and microscopic differences of lithology types under complex surface structures. They are also limited by noise, insufficient texture expression capabilities, and insufficient utilization of spatial context information, resulting in blurred lithological boundaries and low discrimination accuracy.

[0003] In recent years, some studies have introduced emerging technologies such as deep learning and texture feature analysis to integrate multi-dimensional features and improve classification results. However, certain technical bottlenecks remain in achieving efficient collaborative modeling of multi-modal information such as spectral, spatial, and textural data, fine feature capture, and information compensation between different modalities. This is particularly true in areas such as identifying subtle spectral differences, depicting complex structural boundaries, and improving robustness under low signal-to-noise ratio conditions. Therefore, there is an urgent need to propose a new method for fine lithological classification that can fully integrate and dynamically enhance multi-dimensional spectral, spatial, and textural features and possess fine-grained discrimination capabilities. This would overcome the limitations of existing technologies in areas such as the collaborative utilization of heterogeneous features, feature redundancy suppression, and classification accuracy in complex geological regions, thereby promoting the innovative development of intelligent remote sensing geological interpretation technology. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing lithological fine classification methods, such as insufficient utilization of heterogeneous multidimensional feature synergy, limited ability to identify subtle spectral differences and complex texture structures, and insufficient modeling of geological boundaries and spatial context information. In response to the practical challenges of diverse lithological categories, blurred boundaries, and high feature expression dimensions in remote sensing images, this invention proposes a method that can achieve efficient fusion, dynamic enhancement, and selective optimization of spectral, spatial, and textural multidimensional features.

[0005] To achieve the above objectives, the present invention employs the following technical solution: the method comprising: Local feature extraction from multi-scale two-dimensional spectral-texture entanglement coding: This paper utilizes a multi-scale two-dimensional spectral-texture entanglement coding module to perform in-depth local feature mining on remote sensing images. The spectral curve signals of lithological regions are sampled synchronously with the corresponding image texture responses. Spectral-texture feature representations at different resolutions are generated through multi-scale windows. A cross-gating mechanism is adopted to achieve weighted fusion of the interacting spectral and texture information at multiple scales, dynamically capturing subtle spectral differences and complex texture structures between lithologies. A global spatial context is constructed based on spatial-neighborhood adaptive attention. This mechanism is introduced to perform contextual modeling of lithological patches and their boundary regions in the image. Based on the spatial neighborhood distribution characteristics among different sample patches, the attention weight is autonomously adjusted to dynamically strengthen the spatial connection related to the target lithology, while weakening the interference response of irrelevant or anomalous areas. Multi-feature fusion and enhancement: A high-dimensional tensor fusion algorithm based on collaborative compensation is proposed. In the feature layer fusion stage, multidimensional features are encoded into high-order tensors, and nonlinear directional stretching is achieved by using a collaborative compensation strategy. The optimization goal is to give full play to the complementary effects of their respective characteristics. Automatic lithological type identification is achieved based on a fine classification network. On the basis of high-dimensional feature fusion, a deep classification network is used to efficiently mine the fused features and output the final fine lithological classification results.

[0006] In one scheme, the local feature extraction of the multi-scale two-dimensional spectral-texture entanglement coding includes: A self-designed multi-scale two-dimensional spectral-texture entanglement coding module is used to perform in-depth local feature mining on remote sensing images. Through a multi-scale window system, feature representations of the target region are generated at different resolutions. By synchronously sampling the spectral curve signal and the corresponding image texture response, at each scale, the spectral curve of the neighborhood region of the core pixel and the two-dimensional texture features generated by the texture operator are extracted to achieve synchronous sampling and upsizing of the spectrum and texture, so that the spectral and texture content are highly aligned at the pixel level.

[0007] In one approach, the construction of a global spatial context includes: introducing a spatial-neighborhood adaptive attention mechanism to perform context modeling on lithological patches and their boundary regions in the image; autonomously adjusting attention weights based on the spatial neighborhood distribution characteristics between different sample patches; dynamically strengthening spatial connections related to the target lithology; and reducing interference responses from irrelevant or anomalous regions. By using a self-attention mechanism, for each pixel, a neighborhood region containing explicit spatial adjacency and implicit geological connectivity is selected. The correlation strength between each neighborhood region and the target pixel is weighted to construct a global feature interaction system, enabling each pixel to integrate spatial-neighborhood context representation. This process, through multi-layer iterative superposition or expansion of the network at different depths, achieves accurate characterization of complex geological structural boundaries and lithological transition zones, enhancing the model's sensitivity and discrimination ability to subtle changes and discontinuous interfaces.

[0008] In one scheme, the multi-feature fusion and enhancement includes: for the three heterogeneous features of spectrum, texture and space, a high-dimensional tensor fusion algorithm based on collaborative compensation is adopted to encode the three features into high-order tensors respectively, and the unified high-dimensional fusion of multi-dimensional features is achieved through domain projection, nonlinear compensation mapping and selection of enhancement modules. By employing a collaborative compensation strategy, the feature contribution weights of different modes are dynamically adjusted. Through gating compensation and high-order nonlinear activation mechanisms, the discriminative features are selectively enhanced, and noise and redundant dimensions are suppressed.

[0009] In one approach, the automatic lithology type identification based on a fine-grained classification network includes: By inputting the fused features into a deep classification network, and utilizing a multi-branch channel structure and an adaptive channel attention mechanism, the feature contribution weights of different branches and channels are adjusted in an unsupervised adaptive manner to achieve aggregate discrimination of lithology categories. The network structure optimizes the representation of complex features in multiple dimensions such as spectrum, texture and space through backbone convolution modules and self-attention weighting mechanism. It focuses on the most sensitive discriminative feature combinations for the coexistence of diverse lithologies and boundary transitions in complex geological environments.

[0010] In one scheme, the local feature extraction of the multi-scale two-dimensional spectral-texture entanglement coding further includes: using cross-gating to weightedly fuse spectral information and texture information at multiple scales to dynamically capture subtle spectral differences and complex texture structures between lithologies. Multi-scale entanglement coding employs a dual-gated cross-fusion method to achieve bidirectional weighted modulation of spectrum to texture and texture to spectrum; Finally, a scale attention convergence mechanism is used to weight the representations at each scale, balance and integrate multi-scale features, and achieve fine-grained feature preservation while effectively suppressing noise within similar patches.

[0011] In one approach, the fusion tensor is further compressed and filtered by a dynamic selective enhancement and suppression module. High-confidence discrimination information is retained by using scaling exponential linear units and sparse activation selectors, while low-weight and redundant features are masked. The output is a high-dimensional directional vector that has undergone optimal collaboration, compensation and sparse filtering, providing an efficient collaborative expression for subsequent lithology discrimination.

[0012] Beneficial effects of this invention: This invention significantly improves the automatic and refined classification capability of lithological types in complex geological bodies by constructing a deep fusion network of spectral, spatial, and textural multidimensional features. Utilizing multi-scale two-dimensional spectral-texture entanglement coding and a spatial-neighborhood adaptive attention mechanism, it effectively captures and enhances subtle spectral differences and diverse textural structures among lithological samples, overcoming the problem of insufficient utilization of heterogeneous information in traditional methods.

[0013] Meanwhile, based on a feature fusion and collaborative compensation strategy using high-dimensional tensors, redundancy and noise interference can be suppressed, and high-confidence discriminative features can be dynamically screened, further improving the accuracy and stability of classification. This method performs exceptionally well in highly challenging remote sensing interpretation scenarios such as complex lithological coexistence, blurred geological boundaries, and low signal-to-noise ratios. It achieves a high degree of intelligence, automation, and refinement in lithological type identification, providing efficient and reliable technical support for geological exploration and mineral resource investigation, and promoting the development of intelligent remote sensing geological analysis technology. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention.

[0015] Figure 2 This is a flowchart of the local feature extraction process. Detailed Implementation

[0016] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0017] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0018] like Figure 1 As shown, a lithological fine classification method based on the fusion of spectral, spatial, and textural multidimensional features specifically includes: Step 1: Local Feature Extraction from Multi-Scale Two-Dimensional Spectral-Texture Entanglement Coding A self-designed multi-scale two-dimensional spectral-texture entanglement coding module was used to perform in-depth local feature mining of remote sensing images. The spectral curve signals of lithological regions and their corresponding image texture responses were sampled synchronously, and spectral-texture feature representations at different resolutions were generated through multi-scale windows. A cross-gating mechanism was employed to weightedly fuse the interacting spectral and texture information at multiple scales, thereby dynamically capturing subtle spectral differences and complex texture structures between lithologies. This entanglement coding method, while preserving fine-grained features, significantly suppressed apparent noise within similar patches, laying a solid foundation for subsequent spatial relationship modeling.

[0019] In the problem of fine lithological classification of remote sensing images, relying solely on spectral or texture information is often insufficient to accurately characterize rock types in complex geological environments, especially when high-dimensional remote sensing data is affected by noise or surface cover. The multi-scale two-dimensional spectral-texture entanglement coding module was developed to address this challenge. Its core concept lies in deeply fusing spectral and texture features—two distinct characteristics in terms of representational dimension and data properties—at the pixel level through spatial, scale, and interdependent relationships. A gating mechanism is then used to dynamically adjust the fusion, capturing as many subtle differences in lithology as possible.

[0020] like Figure 2 As shown, in specific implementation, S101, firstly, based on the multi-band spectral data of the original remote sensing image, a multi-scale window system is constructed for the target area. At each scale, the system extracts the spectral curves of the neighborhood region for the core pixel. ,in Where k represents the spatial location and k represents the scale number. This refers to the number of bands.

[0021] S102. Next, within the same spatial window, the module generates two-dimensional texture features for the corresponding region using advanced texture operators (such as Gabor filters and improved LBP operators). Here, p represents the texture feature dimension. The synchronized sampling of spectrum and texture ensures high alignment between the spectrum and texture content of each spatial unit, fully preparing for subsequent joint modeling. Spectrum and texture are normalized and their features are upscaled to enable comparability in high-dimensional spaces of the same order of magnitude. The core mathematical process of entanglement coding is embodied in the cross-gating fusion function. In terms of design, given a pair of original features at a scale k, the gating function is defined as follows: in Activated by sigmoid gating. Nonlinear stretching of spectral / texture features As a Hadamard product and It uses trainable weights and biases to achieve bidirectional weighted modulation of spectrum on texture and texture on spectrum. This dual-gated cross-fusion method can adaptively determine which type of information is more critical for lithological differentiation at the current space / scale. For example, in some mineral strata, the texture response is particularly significant, and gating will increase the contribution of the texture channel to ensure the preservation of fine structural information. In cases where other spectral features are particularly prominent, spectral gating is used to amplify relevant channels and suppress spurious differences caused by surface texture.

[0022] S103, the scale k of the multi-scale window increases from small to large, covering microscopic spots to macroscopic geological structures; entanglement coding at each scale. All representations are retained, and weights are ultimately assigned to each scale representation through a scale-attention convergence mechanism. (Scale attention) Acquired through softmax learning, effectively balancing multi-scale fusion: As trainable parameters, they represent the weighted adaptability of different scales to the current region discrimination. In this way, the algorithm can capture information in a multi-dimensional space using a weighted combination method, which not only avoids the distortion of feature representation caused by weak texture or spectral response at a single scale, but also enhances the integration and complementarity of multi-scale information.

[0023] Step 2: Constructing a global spatial context based on spatial-neighborhood adaptive attention After acquiring local multidimensional features, a spatial-neighborhood adaptive attention mechanism is introduced to perform contextual modeling of lithological patches and their boundary regions in the image. This mechanism autonomously adjusts attention weights based on the spatial neighborhood distribution characteristics between different sample patches, dynamically strengthening spatial connections related to the target lithology while weakening interference responses from irrelevant or anomalous regions. Through multi-layered spatial relationship learning, the model's sensitivity to geological structural boundaries and lithological transition zones is significantly improved, enabling precise characterization of subtle changes under complex terrain and discontinuous interfaces.

[0024] Fine-grained lithological classification involves more than just characterizing single points or local areas; it requires in-depth modeling of the spatial continuity of geological structures and their complex boundary transitions. Geological patches often exhibit non-equilibrium and heterogeneous distributions, and boundary zones are the intersection of geomorphic evolution and tectonic activity. These realities mean that relying solely on pixel-level local features is insufficient to describe global topology and contextual relationships. The introduction of a spatial-neighborhood adaptive attention mechanism becomes crucial for effectively carrying and representing global spatial semantics. This mechanism aims to overcome the insensitivity of traditional convolutional networks to long-distance dependencies by learning the dynamic response weights between any two regions in space, constructing a highly spatially extensible contextual interaction graph, providing strong support for fine-grained depiction of lithological interfaces and complex patches.

[0025] Any pixel in the image Depth spectral features It is not isolated and helpless; its true semantic discrimination ability depends on its cognitive ability of the surrounding environment. Within the spatial-neighborhood adaptive attention mechanism framework, all pixels participate in a global feature interaction system: for each pixel… A group of neighboring regions The selected elements include explicit spatial adjacency and implicit geological connectivity (e.g., adaptive expansion via superpixels or graph cuts). For the target pixel and its neighborhood, the model will consider each pair... The strength of the association between them is measured by attention scores. Here, we utilize the concept of self-attention and introduce a learned similarity kernel function. Its definition is: in For learnable linear mappings, the following is typically used: Convolutional or fully connected methods will input features Projected onto the attention query and key space. The above formula, through softmax normalization, ensures that the weighted sum of the attention coefficients of each target point with all neighboring pixels is 1. The efficient construction of the spatial attention matrix allows pixels to no longer simply "borrow" from their surrounding neighbors, but to "remotely perceive" patches or even distant regions that are semantically related to them, thus achieving strong inter-region dependency modeling.

[0026] For example, pixels in lithological transition zones often inherit both the spectral-texture properties of inner patches and are influenced by the structure of outer patches. The model learns to dynamically improve feature responses related to the main lithological category through adaptive optimization of attention scores. With "the same kind" The A value is higher for points with high A values, while the attention weight is automatically reduced for noisy points or cross-boundary interference areas to ensure the purity of the discrimination information to the greatest extent. The spatial aggregation feature obtained after the above weighting is written as: in This is a value mapping function, which can also be parameterized as convolution. That is, the representation vector that integrates spatial-neighborhood context is more original. It is more discriminative and robust. This process can be iteratively stacked in multiple layers or embedded in different depths of the network to expand its expressive power.

[0027] Unlike traditional global attention methods which risk spatial generalization, the proposed adaptive neighborhood selection combines the advantages of both local detail and wide-area connectivity. The flexible extension of this mechanism, employing clustering, segmentation, or prior guidance from geological experts, effectively overcomes the limitations of spatial heterogeneity and complex geological regular morphologies. In actual training, all attention parameters and mapping weights can be optimized through backpropagation within the end-to-end network using a global loss function. This spatial-neighborhood adaptive attention mechanism not only endows the model with sensitivity to complex geological structures but also captures structural convergence and exceptional patterns of lithological patches across different spatial scales.

[0028] The unique advantage of this mechanism lies in the fact that its spatial interaction relationship is not established by predefined rules, but dynamically formed under the joint promotion of data-driven and structural a priori conditions, making it particularly efficient in dealing with complex scenarios with undulating terrain, dense fractures, and mixed lithology.

[0029] Step 3: Multi-class feature fusion and enhancement For three types of heterogeneous features—spectral, textural, and spatial—a high-dimensional tensor fusion algorithm based on collaborative compensation is proposed. In the feature layer fusion stage, multidimensional features are encoded as high-order tensors, and a collaborative compensation strategy is used to achieve nonlinear directional stretching. The optimization objective is to fully leverage the complementary effects of each feature. The algorithm adaptively maps and compresses the original feature space through feature selective enhancement and noise suppression modules, strengthening the feature components most relevant to lithology discrimination, effectively suppressing noise and redundant information, achieving efficient collaborative expression of multidimensional and multi-domain features, and improving overall sample discrimination performance.

[0030] The fundamental challenge of lithology identification lies in the heterogeneity and high-dimensional correlation of data, especially in multi-source remote sensing scenarios. The three main categories of information—spectral, textural, and global spatial features—have independent value but also exhibit overlap, redundancy, and high-order interactions. Simply connecting or superimposing these features often leads to mutual interference and loss of expressive power, and can even affect sample differentiation due to noise amplification. To address this challenge, a collaborative compensation high-dimensional tensor fusion algorithm has emerged. The algorithm's core idea is rooted in the theoretical foundation of multimodal tensor learning. In its implementation, it is guided by lithology discrimination sensitivity, uses collaborative suppression as the main thread, and is supported by nonlinear mapping. It significantly stretches the original features towards the discrimination direction while adaptively eliminating redundancy. First, the three types of fused features output by the spectral-textural-spatial modules after preprocessing... Structurally, it is already a high-dimensional deep representation, carrying different modalities of local fine-grainedness, texture dependence, and spatial structural properties. Each type of feature is organized into a ternary high-order tensor at the cell level. ,in These represent the dimensions of the three types of features. The fusion strategy is based on the interactivity of higher-order tensors—rather than ordinary vector concatenation—which enables more complex feature complementarity and selection.

[0031] The tensor fusion process uses a path gradient of domain projection, nonlinear compensation mapping, and selection enhancement. For a single sample... The regional projection is first composed of three types of feature tensors. Through different linear mappings Projected onto a unified high-dimensional discriminant subspace, forming an inner product tensor representation. Building upon the domain-based approach, collaborative compensation is introduced. The algorithm not only strengthens interaction terms but also focuses on suppressing noise dimensions, adaptively adjusting the contributions of different modes through a gated compensation function. The compensation function structure is as follows: in For Sigmoid gated enhancers, and For learnable kernels and biases, The optional higher-order nonlinear activation (PReLU) is used to achieve targeted selective stretching of features. This mechanism essentially achieves an optimal balance between inter-domain competition and cooperation through gradient learning—that is, strengthening coupled features with strong discriminative power (such as typical lithological spatial-texture responses) while simultaneously weakening apparent spectral components that perform poorly in the current patch. All gating weights and fusion weights can be optimized through backpropagation from end to end within the network.

[0032] After high-dimensional collaborative compensation, the fused tensor Further compression is needed to reduce the expression space to a level that can be directly utilized by downstream discriminative networks. To this end, the algorithm introduces a dynamic selective enhancement and suppression module, resulting in the following transformation: Here, SELU stands for Scaling Exponential Linear Unit. This is a sparse activation selector (which can use sparse attention or learnable gating). The two outputs merge, aiming to preserve rich latent information while simultaneously "masking" low-weight, redundant, or potentially noisy feature dimensions through differentiable sparse filtering, thus achieving feature filtering and information suppression. The final output... It is a high-dimensional directional vector at the same spatial point after optimal inter-domain cooperation, compensation and sparse filtering.

[0033] Unlike the traditional simple fusion of "splicing-linear transformation-dimensionality reduction", firstly, the structural encoding of high-dimensional tensors endows the three feature domains with multi-order correlation capture power, and has stronger coupling generalization ability; secondly, the collaborative compensation strategy and gating mapping work together to ensure that the optimization objectives of different feature domains do not restrict each other and can make up for shortcomings when necessary, so as to maximize information complementarity; thirdly, the selective enhancement and suppression mechanism completely breaks away from the passive dimensionality reduction approach. When irrelevant or even harmful information emerges, the model can adaptively dilute or completely remove it, which greatly improves the noise resistance and robustness of lithology discrimination.

[0034] Step 4: Automatic identification of lithology types based on a fine-grained classification network Based on high-dimensional feature fusion, a deep classification network is employed to efficiently mine the fused features, outputting the final refined lithology classification results. This network structure is specifically optimized for expressing complex multi-dimensional features, enabling unsupervised adaptive adjustment of the contribution weights of different branch channels, thereby achieving automatic aggregation and discrimination of lithology categories. This effectively enhances the model's feature representation capabilities and its universal adaptability to complex geological environments (such as the coexistence of diverse lithologies and fine-grained patchy areas), ensuring high-confidence and accurate classification results for all types of rock.

[0035] After the collaborative fusion of multidimensional features, batches of lithological image samples possess a high-dimensional spectral-texture-spatial deep tensor representation. However, in complex geological scenes, an end-to-end fine-grained classification network is still needed to further efficiently mine and automatically discriminate these fused features. This network design uses multi-domain high-order features as input. Structurally, based on the backbone convolution, it superimposes adaptive multi-branch channels and a self-attention weighting mechanism, achieving both deep discrimination and ensuring sufficient aggregation of various lithological types. Fusion Features After convolutional transformation, normalization, and activation, a deep feature map is formed. ,in For spatial dimensions, The depth dimension after fusion.

[0036] The network employs a multi-branch structure, with each branch corresponding to further discrimination of a specific domain (such as spectral, texture, or spatial representations) or a mixture thereof. Each branch contains one to three convolutional layers (with residual blocks possible), batch normalization (BN), and an activation function. For the m-th branch, its feature transformation is denoted as... in This is a combination of branched convolution and nonlinear mapping. All branch outputs are merged and weighted, with weights determined by adaptive channel attention. The control weights are dynamically learned from input features by channel-specific attention blocks (such as Squeeze-and-Excitation or multi-head Attention blocks), specifically: in This is a learned vector. The fine features of the weighted output are... This dynamic branch weight mechanism enables the network to adaptively adjust the contribution of each feature domain to different samples and geological patches in an unsupervised manner, automatically focusing on the feature combination most sensitive to lithology discrimination. For example, in common lithological coexistence and boundary transition regions, sometimes the spectral domain dominates the discrimination, and sometimes texture or spatial relationship dominates. Both can be captured and amplified by the channel attention mechanism, thereby effectively improving the classification reliability in complex environments.

[0037] Self-attention mechanisms can also perform fine-grained modeling of spatial context. Specific implementations can include two-dimensional spatial self-attention, nonlocal blocks, or transformer mechanisms (such as Vision Transformer Patch Attention). The core idea is to model spatial location... Any two points, through the association function Adaptive weighting based on spatial relevance is performed, where $\mathbf{q}$ and $\mathbf{k}$ are trainable query and key mappings. The feature output after spatial attention weighting is: After multi-layer spatial-channel self-attention, convolution, and normalization, a discriminative feature map with intelligent aggregation is finally generated. Global average pooling (GAP) or adaptive pooling is then used for dimensionality reduction, followed by a fully connected classification layer or softmax output to obtain the lithology type probability distribution. C represents the number of rock categories. The overall loss of the classification network is defined as... in For one-hot labels of the true lithology category, the model optimizes all convolution, attention, and fusion weights end-to-end.

[0038] This network architecture is specifically designed for multi-domain, multi-scale information representation in complex scenarios: dynamic channels and spatial attention mechanisms significantly enhance the model's sensitivity and adaptability to lithological superposition, patchy mixtures, discontinuous boundaries, and fine-grained variation zones. The network can not only automatically aggregate discriminative features but also capture the structural regularities of the data itself in an unsupervised manner. Whether in geological environments with diverse rock types, complex overburden morphologies, and uneven patch sizes, or in remote sensing scenarios with significant noise interference, this network can output high-confidence, high-accuracy automatic lithology type identification results, greatly improving the scientific rigor and generalization ability of lithology identification.

[0039] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0040] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fine lithological classification based on the fusion of spectral, spatial, and textural multidimensional features, characterized in that: The method includes: Local feature extraction from multi-scale two-dimensional spectral-texture entanglement coding: This paper utilizes a multi-scale two-dimensional spectral-texture entanglement coding module to perform in-depth local feature mining on remote sensing images. The spectral curve signals of lithological regions are sampled synchronously with the corresponding image texture responses. Spectral-texture feature representations at different resolutions are generated through multi-scale windows. A cross-gating mechanism is adopted to achieve weighted fusion of the interacting spectral and texture information at multiple scales, dynamically capturing subtle spectral differences and complex texture structures between lithologies. A global spatial context is constructed based on spatial-neighborhood adaptive attention. This mechanism is introduced to perform contextual modeling of lithological patches and their boundary regions in the image. Based on the spatial neighborhood distribution characteristics between different sample patches, the attention weight is autonomously adjusted to dynamically strengthen the spatial connection related to the target lithology, while weakening the interference response of irrelevant or anomalous areas. Multi-feature fusion and enhancement: A high-dimensional tensor fusion algorithm based on collaborative compensation is proposed. In the feature layer fusion stage, multidimensional features are encoded into high-order tensors, and nonlinear directional stretching is achieved by using a collaborative compensation strategy. The optimization goal is to give full play to the complementary effects of their respective characteristics. Automatic lithological type identification is achieved based on a fine classification network. On the basis of high-dimensional feature fusion, a deep classification network is used to efficiently mine the fused features and output the final fine lithological classification results.

2. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 1, characterized in that: The local feature extraction of the multi-scale two-dimensional spectral-texture entanglement coding includes: A self-designed multi-scale two-dimensional spectral-texture entanglement coding module is used to perform in-depth local feature mining on remote sensing images. Through a multi-scale window system, feature representations of the target region are generated at different resolutions. By synchronously sampling the spectral curve signal and the corresponding image texture response, at each scale, the spectral curve of the neighborhood region of the core pixel and the two-dimensional texture features generated by the texture operator are extracted to achieve synchronous sampling and upsizing of the spectrum and texture, so that the spectral and texture content are highly aligned at the pixel level.

3. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 1, characterized in that: The construction of the global spatial context includes: introducing a spatial-neighborhood adaptive attention mechanism to perform context modeling on lithological patches and their boundary regions in the image; autonomously adjusting attention weights based on the spatial neighborhood distribution characteristics between different sample patches; dynamically strengthening spatial connections related to the target lithology; and reducing interference responses from irrelevant or abnormal regions. By using a self-attention mechanism, for each pixel, a neighborhood region containing explicit spatial adjacency and implicit geological connectivity is selected. The correlation strength between each neighborhood region and the target pixel is weighted to construct a global feature interaction system, enabling each pixel to integrate spatial-neighborhood context representation. This process, through multi-layer iterative superposition or expansion of the network at different depths, achieves accurate characterization of complex geological structural boundaries and lithological transition zones, enhancing the model's sensitivity and discrimination ability to subtle changes and discontinuous interfaces.

4. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 1, characterized in that: The multi-feature fusion and enhancement includes: for the three heterogeneous features of spectrum, texture and space, a high-dimensional tensor fusion algorithm based on collaborative compensation is adopted to encode the three features into high-order tensors respectively, and to achieve unified high-dimensional fusion of multi-dimensional features through domain projection, nonlinear compensation mapping and selection enhancement module; By employing a collaborative compensation strategy, the feature contribution weights of different modes are dynamically adjusted. Through gating compensation and high-order nonlinear activation mechanisms, the discriminative features are selectively enhanced, and noise and redundant dimensions are suppressed.

5. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 1, characterized in that: The automatic lithology type identification based on a fine-grained classification network includes: By inputting the fused features into a deep classification network, and utilizing a multi-branch channel structure and an adaptive channel attention mechanism, the feature contribution weights of different branches and channels are adjusted in an unsupervised adaptive manner to achieve aggregate discrimination of lithology categories. The network structure optimizes the representation of complex features in multiple dimensions such as spectrum, texture and space through backbone convolution modules and self-attention weighting mechanism. It focuses on the most sensitive discriminative feature combinations for the coexistence of diverse lithologies and boundary transitions in complex geological environments.

6. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 2, characterized in that: The local feature extraction of the multi-scale two-dimensional spectral-texture entanglement coding further includes: using cross-gating to weightedly fuse spectral information and texture information at multiple scales to dynamically capture subtle spectral differences and complex texture structures between lithologies. Multi-scale entanglement coding employs a dual-gated cross-fusion method to achieve bidirectional weighted modulation of spectral to texture and texture to spectral. Finally, a scale attention convergence mechanism is used to weight the representations at each scale, balance and integrate multi-scale features, and achieve fine-grained feature preservation while effectively suppressing noise within similar patches.

7. The lithological fine classification method based on the fusion of spectral-spatial-texture multidimensional features according to claim 4, characterized in that: The fusion tensor is further compressed and filtered by the dynamic selective enhancement and suppression module. It retains high-confidence discrimination information by using scaling exponential linear units and sparse activation selectors, while masking low-weight and redundant features. The output is a high-dimensional directional vector that has undergone optimal collaboration, compensation and sparse filtering, providing an efficient collaborative expression for subsequent lithology discrimination.

Images