Mineral spectral feature unmixing device using a generative adversarial network

Abstract

The application discloses a mineral spectral feature unmixing device using a generative adversarial network, comprising a multi-modal data preprocessing module for correcting and enhancing original mineral spectral data, solving small sample training problems, and providing high-quality input for subsequent unmixing, a generation network module for fusing noise and geological text description, a dynamic adversarial training module for initializing endmember features based on comparative learning of mineral paragenetic sequences, improving prior knowledge of the model on mineral combination rules, and an endmember reconstruction verification module for cyclically reconstructing a single albedo matrix using a generator network unit and a discriminator network unit; the application effectively eliminates nonlinear scattering and atmospheric noise interference through a model conversion unit and a 3D convolution kernel, combines data synthesis capacity of a conditional generative adversarial network, alleviates a small sample training bottleneck, and strengthens prior constraints of endmember feature combination rules based on comparative learning of mineral paragenetic sequences through a dynamic adversarial training mechanism.

Description

Technical Field

[0001] This invention relates to the field of geological exploration and mineral processing equipment technology, and in particular to a mineral spectral feature demixing device utilizing generative adversarial networks. Background Technology

[0002] The principle of mineral spectral feature unmixing is to decompose the mixed spectral signal into the endmember spectra and abundance distributions of different mineral components through mathematical models and algorithms, thereby realizing the quantitative identification and analysis of mineral components. Its core goal is to solve the problem of mixed spectra caused by spatial resolution limitations in remote sensing images, and to realize mineral species identification and abundance inversion. Generative adversarial networks are a deep learning model framework. Its core idea is to realize data generation and optimization through adversarial training of generators and discriminators.

[0003] In real-world exploration, mineral assemblages are complex, but the amount of labeled spectral data obtained in the field or laboratory is limited, resulting in poor generalization of traditional models. During spectral detection, factors such as multiple scattering between mineral particles and ambient light cause spectral mixing to exhibit nonlinear characteristics, leading to significant errors in traditional linear models. Therefore, this invention proposes a mineral spectral feature unmixing device using generative adversarial networks to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose a mineral spectral feature unmixing device utilizing generative adversarial networks (GANs). This GAN-based device effectively eliminates nonlinear scattering and atmospheric noise interference through model transformation units and 3D convolutional kernels, significantly improving the quality of spectral data. Combined with the data synthesis capabilities of conditional GANs, it alleviates the bottleneck of small-sample training. Through a dynamic adversarial training mechanism based on mineral co-occurrence sequence comparison learning, it strengthens the prior constraints of endmember feature combination rules, enhancing the physical interpretability of mineral unmixing. The endmember reconstruction verification module utilizes a cyclic consistency loss function to form a closed-loop feedback, ensuring high fidelity and anti-interference capabilities of endmember generation. Overall, it improves the robustness and applicability of the mineral unmixing model in complex geological scenarios.

[0005] To achieve the objectives of this invention, the following technical solution is provided: a mineral spectral feature unmixing device utilizing generative adversarial networks, comprising a multimodal data preprocessing module, a generator network module, a dynamic adversarial training module, and an endmember reconstruction and verification module. The multimodal data preprocessing module is responsible for correcting and enhancing the original mineral spectral data, solving the problem of small sample training, and providing high-quality input for subsequent unmixing. The generator network module is used to fuse noise and geological text descriptions, enhancing the ability to discriminate homogeneous regions and achieving nonlinearity. The dynamic adversarial training module is used to initialize endmember features based on comparative learning of mineral co-occurrence sequences, improving the model's prior knowledge of mineral assemblage patterns. The endmember reconstruction and verification module is used to cyclically reconstruct a single albedo matrix using generator network units and discriminator network units, and to verify the quality of endmember generation using a consistency loss function.

[0006] Further improvements are made in that: the multimodal data preprocessing module includes a model conversion unit, a noise filtering unit, and a data synthesis unit. The model conversion unit is used to convert the original reflectance into single-scattering reflectance. The noise filtering unit is used to perform cross-band noise filtering using 3D transposed convolution kernels and to eliminate atmospheric interference and sensor noise by combining spectral differential processing. The data synthesis unit is used to generate synthetic spectral data of mineral assemblages by combining conditional generative adversarial networks.

[0007] A further improvement is that the multimodal data preprocessing module uses an improved Hapke model to eliminate nonlinear scattering effects.

[0008] Further improvements are made in that the 3D convolution kernel in the noise filtering unit has a three-dimensional size of 5*5*7, and the 3D convolution kernel simultaneously extracts spatial neighborhood correlation and spectral continuity to capture the nonlinear scattering effect between mineral particles.

[0009] Further improvements are made in that: the generation network module includes a generator network unit and a discriminator network unit. The generator network unit is used to receive random noise vectors and geological text descriptions to generate mineral endmember spectral matrices and abundance distribution maps. The discriminator network unit is used to enhance the ability to distinguish notification regions by employing depth-separable convolution kernel superpixel separation constraints.

[0010] A further improvement is that the generator unit includes a transposed convolutional network and a spatial attention mechanism layer. The transposed convolutional network has a kernel size of 3*3 and a stride of 2. The spatial attention mechanism layer is used to calculate the spectral contribution weight matrix of each pixel.

[0011] A further improvement is that the discriminator network unit adopts a depth-separable convolutional structure with a kernel depth multiplier of 8.

[0012] The beneficial effects of this invention are as follows: This invention effectively eliminates nonlinear scattering and atmospheric noise interference through model transformation units and 3D convolution kernels, significantly improving the quality of spectral data. Combined with the data synthesis capability of conditional generative adversarial networks, it alleviates the bottleneck of small-sample training. Through dynamic adversarial training mechanism based on mineral coexistence sequence comparison learning, it strengthens the prior constraints of endmember feature combination rules and improves the physical interpretability of mineral unmixing. The endmember reconstruction verification module uses the cyclic consistency loss function to form a closed-loop feedback, ensuring the high fidelity and anti-interference of endmember generation. Overall, it improves the robustness and applicability of the mineral unmixing model in complex geological scenarios. Attached Figure Description

[0013] Figure 1 This is a system architecture diagram of the present invention;

[0014] Figure 2 This is an architecture diagram of the multimodal data preprocessing module of the present invention;

[0015] Figure 3 This is a diagram of the network module architecture of the present invention. Detailed Implementation

[0016] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0017] Generative Adversarial Networks (GANs) are a deep learning model framework whose core idea is to achieve data generation and optimization through adversarial training of a generator and a discriminator. The generator is input with a random noise vector and generates samples that are similar to the distribution of real data, with the goal of "deceiving" the discriminator. The discriminator receives real data and generated samples and judges their source, with the goal of improving its discrimination ability. The two play a dynamic game during training, eventually making it difficult to distinguish between generated samples and real data.

[0018] Based on this, according to Figure 1 , Figure 2 , Figure 3As shown, this embodiment provides a mineral spectral feature unmixing device using generative adversarial networks (GANs), including a multimodal data preprocessing module, a generator network module, a dynamic adversarial training module, and an endmember reconstruction and verification module. The multimodal data preprocessing module is responsible for correcting and enhancing the original mineral spectral data, solving the small sample training problem, and providing high-quality input for subsequent unmixing. The multimodal data preprocessing module adopts an improved Hapke model to eliminate nonlinear scattering effects. The generator network module is used to fuse noise and geological text descriptions, enhancing the ability to distinguish homogeneous regions and achieving nonlinearity. The dynamic adversarial training module is used to initialize endmember features based on comparative learning of mineral co-occurrence sequences, improving the model's prior knowledge of mineral assemblage patterns. The endmember reconstruction and verification module is used to cyclically reconstruct a single albedo matrix using generator network units and discriminator network units, and uses a consistency loss function to verify the quality of endmember generation.

[0019] The multimodal data preprocessing module includes a model conversion unit, a noise filtering unit, and a data synthesis unit. The model conversion unit converts the raw reflectance into single-scattering reflectance and then into single-scattering albedo, addressing the nonlinear mixing problem caused by multiple scattering between mineral particles. The noise filtering unit uses a 3D transposed convolution kernel for cross-band noise filtering, combined with spectral differential processing to eliminate atmospheric interference and sensor noise. The 3D convolution kernel in the noise filtering unit has a three-dimensional size of 5*5*7. The 3D convolution kernel simultaneously extracts spatial neighborhood correlation and spectral continuity, capturing the nonlinear scattering effect between mineral particles. It uses the 5*5*7 3D transposed convolution kernel to filter noise across bands, and combined with spectral differential processing, preserves the scattering characteristics of mineral particles and eliminates atmospheric interference. The data synthesis unit combines a conditional generative adversarial network to generate synthetic spectral data of mineral combinations, and combines geological text descriptions to generate synthetic spectral data, expanding the training samples. It also generates synthetic spectral data of specific mineral combinations through a conditional generative adversarial network, addressing the problem of scarce field exploration data and improving the model's generalization ability to small samples.

[0020] The generator network module includes a generator network unit and a discriminator network unit. The generator network unit receives random noise vectors and geological text descriptions to generate mineral endmember spectral matrices and abundance distribution maps. It simultaneously extracts spatial neighborhood correlations and spectral continuity through 3D transposed convolutional kernels to capture nonlinear scattering effects between mineral particles. The generator unit includes a transposed convolutional network and a spatial attention mechanism layer. The transposed convolutional network has a 3*3 kernel size and a stride of 2. The spatial attention mechanism layer is used to calculate the spectral contribution weight matrix of each pixel. The discriminator network unit uses depth-separable convolutional kernels to superpixel separation constraints to enhance the ability to discriminate notification regions.

[0021] The discriminator network unit adopts a depth-separable convolutional structure with a depth multiplier of 8. By using a lightweight structure with a depth multiplier of 8, the discriminator network unit reduces computational complexity and adapts to the real-time processing requirements of spaceborne equipment.

[0022] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1.A mineral spectral feature unmixing device using a generative adversarial network, characterized by: The application relates to a mineral endemism reconstruction method based on multi-modal data, which comprises a multi-modal data preprocessing module, a generation network module, a dynamic adversarial training module and an endemism reconstruction verification module, wherein the multi-modal data preprocessing module is used for correcting and enhancing original mineral spectrum data, solving a small sample training problem, providing high-quality input for subsequent unmixing, and comprises a model conversion unit, a noise filtering unit and a data synthesis unit; the model conversion unit is used for converting original reflectivity into single scattering reflectivity; the noise filtering unit is used for carrying out cross-band noise filtering by adopting a 3D transposed convolution kernel, eliminating atmospheric interference and sensor noise by combining spectral differential processing; and the data synthesis unit is used for generating synthetic spectrum data of a mineral combination by combining a conditional generative adversarial network; the generation network module is used for fusing noise and geological text description, enhancing the discrimination ability for homogeneous regions and realizing nonlinearity; the generation network module comprises a generator network unit and a discriminator network unit; the generator network unit is used for receiving a random noise vector and a geological text description, generating a mineral endemism spectrum matrix and an abundance distribution graph; the discriminator network unit is used for adopting a deep separable convolution kernel superpixel separation constraint to enhance the discrimination ability for homogeneous regions; the generator network unit comprises a transposed convolution network and a spatial attention mechanism layer; the transposed convolution network has a convolution kernel size of 3*3 and a step of 2; the spatial attention mechanism layer is used for calculating a spectrum contribution weight matrix of each pixel point; the dynamic adversarial training module is used for initializing endemism features based on comparative learning of a mineral paragenetic sequence, improving the priori cognition of the model to mineral combination rules; and the endemism reconstruction verification module is used for cyclically reconstructing a single scattering reflectivity matrix by the generator network unit and the discriminator network unit, and verifying the endemism generation quality by using a consistency loss function. 2.The mineral spectral feature unmixing device using a generative adversarial network according to claim 1, wherein: The multi-modal data preprocessing module adopts an improved Hapke model to eliminate nonlinear scattering effects. 3.The mineral spectral feature unmixing device using a generative adversarial network according to claim 1, wherein: The 3D convolution kernel in the noise filtering unit has a three-dimensional size of 5*5*7; the 3D convolution kernel synchronously extracts spatial neighborhood correlation and spectral continuity, and captures nonlinear scattering effects between mineral particles. 4.The mineral spectral feature unmixing apparatus using a generative adversarial network according to claim 1, wherein: The discriminator network unit adopts a deep separable convolution structure, and a convolution kernel depth multiplier is 8.

Images