Vegetation classification method and device based on space-spectrum neural network, equipment and medium
By employing information separation and gradient optimization strategies in spatial-spectral neural networks, the hyperspectral image classification network architecture is automatically adjusted, solving the problem of insufficient extraction of spectral and spatial features in existing technologies. This achieves efficient and stable vegetation classification, meeting the needs of large-scale, high-precision remote sensing monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG TIANYUAN TECHNOLOGY CO LTD
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-29
AI Technical Summary
Existing hyperspectral image classification methods still have room for improvement in terms of classification accuracy and efficiency in complex vegetation fine classification tasks, and cannot meet the needs of large-scale, high-precision remote sensing monitoring. Furthermore, existing NAS methods fail to fully consider the complex relationship between spectral and spatial dimensions, resulting in insufficient vegetation feature extraction and high computational resource consumption.
We employ a spatial-spectral neural network approach, which utilizes a spatial-spectral hybrid search space with information separation and a gradient-optimized differentiable architecture search strategy to construct an information-separating spectral transformation operator and an information-separating spatial attention deep convolution operator. This approach automatically adjusts the network architecture, extracts spectral and spatial features, reduces computational costs, and improves search efficiency.
An optimal network architecture for efficient collaborative extraction of spectral and spatial features was achieved, enabling stable and efficient parameter training, ensuring the model's classification accuracy and generalization ability, and meeting the needs of large-scale, high-precision remote sensing monitoring.
Smart Images
Figure CN122115936A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing science and technology, and in particular to a vegetation classification method, apparatus, equipment and medium based on a spatial-spectral neural network. Background Technology
[0002] Hyperspectral image classification has wide applications in various fields such as agricultural monitoring, environmental governance, and land use planning. Especially in the task of fine-grained vegetation classification, it can provide crucial technical support for crop growth assessment, vegetation resource surveys, and ecological protection. The core advantage of hyperspectral image data lies in its inclusion of a large number of continuous spectral bands. Compared to traditional RGB images, it can capture information related to the physicochemical properties of vegetation, laying the foundation for improving classification accuracy. However, hyperspectral data is characterized by high dimensionality and strong spatial heterogeneity, making its analysis and classification process challenging and requiring advanced technologies for efficient processing.
[0003] Traditional hyperspectral image classification methods rely on experts manually designing feature extraction rules and constructing vegetation feature databases for classification. This approach is not only labor-intensive but also suffers from poor feature database versatility, making it difficult to adapt to hyperspectral datasets from different sources and scenarios. With the development of deep learning technology, models such as Convolutional Neural Networks (CNNs) have been widely applied to hyperspectral image classification, automatically learning features from the data and demonstrating advantages in handling high-dimensional and complex data. However, the network structures of existing deep learning methods still require manual design, relying on researchers' experience for parameter adjustments, and the network architecture contains redundancy, leading to significant computational resource consumption. The recently emerging Neural Architecture Search (NAS) technology, by automatically exploring the optimal network structure within a predefined search space, reduces manual intervention and can adaptively adjust the model architecture for different tasks and datasets. However, existing NAS methods still have shortcomings in hyperspectral image processing: on the one hand, the search space is mostly constructed based on traditional operators such as convolution and pooling, which fails to fully consider the "image-spectrum integration" characteristic of hyperspectral images and ignores the complex relationship between spectral and spatial dimensions, resulting in insufficient extraction of full-spectrum sequence features and multi-morphological spatial features of vegetation; on the other hand, redundant search spaces increase the computational burden, reduce search efficiency, and affect the actual application effect of the model.
[0004] Therefore, there is an urgent need for a vegetation classification method based on spatial-spectral neural networks to address the technical problem that there is still room for improvement in classification accuracy and efficiency in complex vegetation fine classification tasks, which cannot meet the needs of large-scale, high-precision remote sensing monitoring. Summary of the Invention
[0005] To overcome the problems existing in related technologies, this disclosure provides a vegetation classification method, device, equipment and medium based on spatial-spectral neural networks, in order to solve the technical problem that in the task of fine classification of complex vegetation, there is still room for improvement in classification accuracy and efficiency, which cannot meet the needs of large-scale, high-precision remote sensing monitoring.
[0006] This specification provides one or more embodiments of a vegetation classification method based on a spatial-spectral neural network, comprising the following steps: Hyperspectral datasets from different platforms were divided into training, validation, and test sets, and quantitative evaluation metrics were set. Based on the information separation spatial-spectral hybrid search space, a gradient-optimized differentiable architecture search strategy is used to search for neural network architectures to obtain the target network architecture. The spatial-spectral hybrid search space includes at least an information separation spectral transform operator for extracting spectral features and an information separation spatial attention deep convolution operator for extracting spatial features. Based on the training set and validation set, the target network architecture is trained to obtain the classification model; The test set is input into the trained classification model to obtain the vegetation classification results, and the classification results are evaluated based on quantitative evaluation indicators.
[0007] Preferably, the construction of the information separation spectral transformation operator specifically includes the following steps: The input feature map is spectrally normalized to obtain the first scaling factor for each channel. The information weights of each channel are calculated based on the first scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into the spectral Transformer module for processing, and the low-information-content feature portion is subjected to lightweight convolution processing; The output of the spectral Transformer module is fused with the output of the lightweight convolution processing to obtain the output of the information separation spectral transformation operator.
[0008] Preferably, the construction of the information separation spatial attention deep convolution operator specifically includes the following steps: The input feature map is spatially normalized to obtain the second scaling factor for each channel. The information weights of each channel are calculated based on the second scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into a multi-dimensional parameterless attention deep convolution module for processing, and the low-information-content feature portion is processed by grouped convolution. The output of the multi-dimensional parameterless attention deep convolution module is fused with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator.
[0009] Preferably, the construction of the multi-dimensional parameter-free attention deep convolution module specifically includes the following steps: Calculate the one-dimensional spectral nonparametric attention weights and the two-dimensional spatial nonparametric attention weights of the input feature map, respectively. The one-dimensional spectral parameterless attention weights and the two-dimensional spatial parameterless attention weights are expanded and fused to obtain three-dimensional spatial spectral parameterless attention weights. The input feature map is weighted using the three-dimensional spatial spectrum parameter-free attention weights, and a depthwise convolution operation is performed.
[0010] Preferably, the method of using a gradient-optimized differentiable architecture search strategy to search for neural network architectures specifically includes the following steps: Optimize the weight parameters of each operator in the spatial-spectral hybrid search space on the training set; Optimize continuous structural parameters representing network architecture selection on the validation set; The above optimization process is executed iteratively until convergence, and the target network architecture is determined based on the converged continuous structural parameters.
[0011] This specification provides one or more embodiments of a vegetation classification device based on a spatial-spectral neural network, comprising: The dataset construction module is used to divide hyperspectral datasets from different platforms into training, validation, and test sets, and to set quantitative evaluation metrics. The target network architecture search module is used to search for neural network architectures based on a gradient-optimized differentiable architecture search strategy in a spatial-spectral hybrid search space to obtain the target network architecture. The spatial-spectral hybrid search space includes at least an information separation spectral transformation operator for extracting spectral features and an information separation spatial attention deep convolution operator for extracting spatial features. The classification model training module is used to train the parameters of the target network architecture based on the training set and the validation set to obtain the classification model; The classification module is used to input the test set into the trained classification model to obtain the vegetation classification results, and to evaluate the classification results based on quantitative evaluation indicators.
[0012] Preferably, the construction of the information separation spectral transformation operator specifically includes the following steps: The input feature map is spectrally normalized to obtain the first scaling factor for each channel. The information weights of each channel are calculated based on the first scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into the spectral Transformer module for processing, and the low-information-content feature portion is subjected to lightweight convolution processing; The output of the spectral Transformer module is fused with the output of the lightweight convolution processing to obtain the output of the information separation spectral transformation operator.
[0013] Preferably, the construction of the information separation spatial attention deep convolution operator specifically includes the following steps: The input feature map is spatially normalized to obtain the second scaling factor for each channel. The information weights of each channel are calculated based on the second scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into a multi-dimensional parameterless attention deep convolution module for processing, and the low-information-content feature portion is processed by grouped convolution. The output of the multi-dimensional parameterless attention deep convolution module is fused with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator.
[0014] This specification provides one or more embodiments of a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the vegetation classification method based on the spatial-spectral neural network described above.
[0015] This specification provides one or more embodiments of a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the vegetation classification method based on the spatial-spectral neural network described above.
[0016] This disclosure provides a vegetation classification method, apparatus, device, and medium based on a spatial-spectral neural network. Its advantages lie in that by independently dividing the data into training, validation, and test sets, it effectively prevents information leakage during model training and evaluation, ensuring the reliability of performance evaluation results. Pre-defined quantitative evaluation indicators provide objective and unified optimization goals and final performance metrics for the entire automated process. Through a specially designed spatial-spectral hybrid search space, the search process can automatically construct an optimal network architecture that efficiently and collaboratively extracts spectral and spatial features, addressing the problem of poor adaptability of fixed architectures. By combining an improved differentiable search strategy, high-performance structures can be found quickly with lower computational costs, realizing the transformation from manual design to automatic customization. Since the network topology has been optimized for the current task, the parameter training process can converge more stably and efficiently, allowing the model weights to fully learn the complex mapping from data to vegetation categories, thereby releasing the full performance potential of the customized architecture and obtaining a strong predictive model that can be directly used for classification. By using completely independent test sets for evaluation, the generalization ability and actual classification accuracy of the final model can be objectively and unbiasedly quantified, thus empirically demonstrating the comprehensive effectiveness of the entire automated process and proving that the produced model has practical application value. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating a vegetation classification method based on a spatial-spectral neural network provided for one or more embodiments of this specification; Figure 2 A schematic diagram of the vegetation classification method based on spatial-spectral neural network provided in one or more embodiments of this specification; Figure 3 A schematic diagram of the structure of a vegetation classification device based on a spatial-spectral neural network provided for one or more embodiments of this specification; Figure 4 This is a schematic diagram of the structure of a computer device provided for one or more embodiments of this specification. Detailed Implementation
[0019] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this invention.
[0020] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.
[0021] Method Implementation Examples According to embodiments of the present invention, a vegetation classification method based on a spatial-spectral neural network is provided, such as... Figure 1 The diagram shown is a flowchart illustrating the vegetation classification method based on a spatial-spectral neural network provided in this embodiment. The vegetation classification method based on a spatial-spectral neural network according to this embodiment includes the following steps: S110. Divide the hyperspectral datasets from different platforms into training, validation, and test sets according to a sample split ratio of 7:2.9:0.1, and set up three evaluation metrics—overall accuracy (OA), average accuracy (AA), and Kappa coefficient—to quantitatively evaluate the final classification model.
[0022] S120. Based on the information separation-based spatial-spectral hybrid search space, a gradient-optimized differentiable architecture search strategy is used to search for a neural network architecture based on unit structure to obtain the target network architecture. In existing neural network architectures, spectral features and spatial features are usually extracted together, ignoring their uniqueness in hyperspectral images. This invention introduces an information separation strategy by designing an information-separating spectral Transformer (ISTransformer) and an information-separating spatial NPA-DConv (ISNPA-DConv) module to process spectral and spatial features respectively. These two modules can effectively extract full-spectrum sequence features and multi-morphological spatial features from hyperspectral images, improving the model's feature extraction capability and reducing information redundancy. The spatial-spectral hybrid search space includes at least an information-separating spectral transform operator for extracting spectral features and an information-separating spatial attention deep convolution operator for extracting spatial features. Simultaneously, to improve the efficiency of model structure search, this invention improves the traditional DARTS (Differentiable Architecture Search) method. By optimizing the gradient calculation order and reducing redundant calculations, the computational burden in the model search process is significantly reduced, and the search speed is improved. This improved DARTS search strategy can identify the optimal network structure more quickly, reducing training time and computational resource consumption. Through gradient-based optimization, ISSS-NAS can automatically search and determine suitable network structures within a large search space.
[0023] S130. Based on the training and validation sets, the target network architecture is trained to obtain the classification model. Specifically, during the search process, the loss function of the training and validation sets is used to optimize the network structure parameters, and the optimal network structure is finally selected. After obtaining the optimal network structure, the model parameters are further optimized using the training set, and the classification accuracy is evaluated using the test set to ensure that the model can achieve the best classification performance.
[0024] S140. Input the test set into the trained classification model to obtain vegetation classification results, and evaluate the classification results based on quantitative evaluation indicators. The ISSS-NAS method of this invention can automatically adjust the network architecture according to the characteristics of different datasets to ensure optimal classification results for each dataset. By introducing an information separation strategy and an improved DARTS method, ISSS-NAS demonstrates excellent adaptability in high-precision hyperspectral vegetation classification tasks, and can process hyperspectral image datasets from different sources, meeting the needs of large-scale, high-precision remote sensing monitoring.
[0025] The method provided in this embodiment effectively prevents information leakage during model training and evaluation by independently dividing the data into training, validation, and test sets, ensuring the reliability of performance evaluation results. Pre-defined quantitative evaluation metrics provide an objective and unified optimization objective and final performance measurement standard for the entire automated process. A specially designed spatial-spectral hybrid search space enables the search process to automatically construct an optimal network architecture that efficiently and collaboratively extracts spectral and spatial features, addressing the problem of poor adaptability of fixed architectures. Combined with an improved differentiable search strategy, a high-performance structure can be quickly found with lower computational cost, realizing a shift from manual design to automatic customization. Since the network topology is optimized for the current task, the parameter training process converges more stably and efficiently, allowing the model weights to fully learn the complex mapping from data to vegetation categories, thereby releasing the full performance potential of the customized architecture and obtaining a strong predictive model that can be directly used for classification. Evaluation using completely independent test sets allows for objective and unbiased quantification of the final model's generalization ability and actual classification accuracy, thus empirically demonstrating the comprehensive effectiveness of the entire automated process and proving that the produced model has practical application value.
[0026] The search unit is a core component of ISSS-NAS, primarily consisting of spectral pooling, regular units, dimensionality reduction units, fully connected layers, and the full-spectrum feature extraction operation ISTransformer and the multi-morphological spatial feature extraction operation ISNPA-DConv. Each unit takes the outputs of the previous two layers as input, and each node takes the outputs of all previous nodes in the unit as input. The output of the i-th node in the l-th layer can be represented as:
[0027]
[0028] in Indicates the first The input set of each node. and This indicates the output of the first two units. This represents the set of operations for the search operator. It is a search operator The weight, which is related to the current layer. Regarding this, different types of elements have different structural parameters, while elements of the same type share the same set of continuous variables. . Corresponding to One candidate operation. The output is the concatenation of the outputs of all nodes, which can be formulated as:
[0029] As the search process progresses, ISSS-NAS will select different operations and search spaces.
[0030] In one embodiment, to reduce the computational cost of the model and reduce information redundancy, an information separation approach is introduced to improve the spectral Transformer. The purpose of the separation operation is to process information-rich feature maps separately from their corresponding less information-rich feature maps. This allows for targeted processing of high-information-content feature maps and simplified processing of low-information-content features, thereby saving computational costs. The scaling factor in the GroupNorm (GN) layer is used to evaluate the information content of different feature maps. The construction of the information separation spectral transform operator specifically includes the following steps: For the input feature map ,in It is the spectral channel dimension. and The dimensions are height and width. First, they are standardized by subtracting the mean μ and dividing by the standard deviation σ to obtain the first scaling factor for each channel, as shown below:
[0031] Where μ and σ are the mean and standard deviation of X, ε is a low-valued constant added for division stability, and γ and β are parameters of the trainable affine transformation. It is worth noting that the trainable parameters in the GN layer... As a method to measure the variance of pixels in each feature map, richer information reflects more variations in the pixels, resulting in a larger γ.
[0032] Based on the first scaling factor, the information weights of each channel are calculated, and according to a preset separation ratio, the input feature map is separated into a high-information-content feature part and a low-information-content feature part. Specifically, to achieve information separation, a normalization method is used to characterize the importance of the feature map:
[0033] Then, through The weights of the reweighted feature map are obtained through Function mapping to range The data is then ranked by importance, and the high-information-content portions are selected by setting thresholds or separation ratios. Or low information content weight (The separation ratio was set to 0.5 in the experiment). The input will be based on the obtained weights. Separated as well as And multiply it by the weight to obtain features containing different amounts of information:
[0034]
[0035] The high information content feature part The input is fed into the Spectral Transformer module for feature extraction. The low-information feature part is processed by pointwise convolution (PWC), that is, a two-dimensional convolution with a kernel size of 1×1, to extract spectral dimension features.
[0036] The spectral Transformer used in this chapter employs convolution operations instead of linear mappings to compute the Query, Key, and Value. It uses one-dimensional convolution to perform spectral-dimensional convolution on the self-attention-applied features, ultimately extracting features from the entire spectral sequence. The formula for the spectral Transformer in this chapter can be expressed as follows:
[0037]
[0038] in This represents the output of the spectral Transformer. This indicates the high-information portion. This represents the positional embedding. The calculation of the low-information portion can be formulated as follows:
[0039] The output of the spectral Transformer module is fused and concatenated with the output of the lightweight convolution processing to obtain the output of the information separation spectral transform operator:
[0040] In one embodiment, NPA-DConv based on information content separation uses the same method as ISTransformer to separate information to obtain the high information content portion. and low information content NPA-Dconv is an extension of one-dimensional nonparametric attention (1D NPA), two-dimensional nonparametric attention (2D NPA), and three-dimensional nonparametric attention (3D NPA), and specifically includes the following steps: Calculate the one-dimensional spectral nonparametric attention weights and the two-dimensional spatial nonparametric attention weights of the input feature map, respectively.
[0041] The one-dimensional spectral parameterless attention weights and the two-dimensional spatial parameterless attention weights are expanded and fused to obtain three-dimensional spatial spectral parameterless attention weights.
[0042] The input feature map is weighted using the three-dimensional spatial spectrum parameter-free attention weights, and a depthwise convolution operation is performed.
[0043] Specifically, first, define Target neuron The energy function is used to describe its relationship with other neurons. The linear separability is as follows:
[0044] in, and This represents the linear transformation process of a neuron. and This represents the two extreme states of the target neuron and other neurons being activated and inhibited. It can be seen that the energy function... The smaller the value, the stronger the target neuron. Other neurons The greater the linear separability between them, the better the target neuron. The more important it is. When the target neuron and other neurons are activated and compressed to extreme states, that is... and When, the energy function representing linear separability. This reaches the minimum value. To simplify the calculation, set... and .Will Substituting into a linear transformation, the energy function becomes... It can be represented as:
[0045] In the above formula It is a regularization term, where the parameter Used to control the linear variation coefficient The sparsity of the energy function. To obtain the minimum value of the energy function, let... For the coefficients of a linear transformation and bias The partial derivative is zero. Through calculation, we can obtain... and Closed-form solution:
[0046] and Representing vectors Except The mean and variance of all neurons except those in the original dataset were calculated. To conserve computational resources, the mean and variance of the original dataset were calculated. and Replace with the mean and variance calculated using all neurons. and Thus, the mean and variance only need to be calculated once for any target neuron. Therefore, the minimum energy function can be calculated as:
[0047] lower Indicates the target neuron The high linear separability from surrounding neurons also indicates that the target neuron... It possesses high importance. Therefore, it can be used... to indicate The attention weights, where the final one-dimensional spectral parametric attention 1D NPA can be described as:
[0048] in, Represents the spectral vector Energy function of all target neurons This forms an energy function vector. The Sigmoid function normalizes the attention weights. Notably, x represents a hyperspectral image patch. Any spectral vector in, This indicates a one-dimensional spectrum without parameter attention.
[0049]
[0050]
[0051] This indicates that the two-dimensional spectrum has no parameter attention. This represents the input hyperspectral image patch. It is the spectral vector of the center pixel. express The spectral vector of any pixel in the array.
[0052] Since 1D NPA and 2D NPA have already been obtained based on the characteristics of their respective spectral and spatial dimensions, 3D NPA is a combination of the two. Specifically, it involves extending the one-dimensional spectral attention and two-dimensional spatial attention to match their sizes to the hyperspectral image patches. The size is calculated, and then each element is multiplied. The formula can be expressed as:
[0053] The construction of the information separation spatial attention deep convolution operator specifically includes the following steps: Spatial dimension normalization is performed on the input feature map to obtain the second scaling factor for each channel.
[0054] Based on the second scaling factor, the information weights of each channel are calculated, and according to a preset separation ratio, the input feature map is separated into high-information-content feature parts. Low information content feature part .
[0055] The high information content feature part The input is processed by a multi-dimensional, parameter-free attention-based deep convolution module to extract multi-shaped spatial features, thereby reducing the low-information feature portions. Groupwise convolution (GWC) is used to achieve lightweight spatial dimension feature extraction.
[0056] The output of the multi-dimensional parameterless attention deep convolution module is fused and concatenated with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator:
[0057] In one embodiment, a gradient-optimized differentiable architecture search strategy is used to search for neural network architectures, specifically including the following steps: Input: Training set samples Validation set samples Structural parameters Operator weight parameters
[0058] Processing procedure: Perform this before the model converges: In the training set Calculate gradient To optimize and update the operation weight parameters of each operator in the spatial spectrum mixing search space. .
[0059] Calculate gradients on the training and validation sets. Optimize and update the continuous structural parameters representing the network architecture selection. .
[0060] The above optimization process is executed iteratively until convergence, and the converged continuous structure parameters are then used as the basis for the optimization. Determine the target network architecture.
[0061] In the network architecture search step, a search is performed on the above-mentioned structures. When the loss on the validation set no longer decreases, the optimal network architecture is determined based on the structural parameter α, which is the architecture of regular units and dimensionality reduction units. Model optimization is to train the neural network model from scratch on the training dataset in the network architecture of the optimal network architecture search.
[0062] Because the datasets have varying spectral and spatial resolutions and include different vegetation types, network architecture searches were performed separately for each dataset, and the model's classification accuracy was evaluated on the test dataset. Figure 2 The diagram shown is a schematic flowchart of the vegetation classification method based on the spatial-spectral neural network provided in this embodiment.
[0063] Device Examples According to embodiments of the present invention, a vegetation classification device based on a spatial-spectral neural network is provided, such as... Figure 3 The diagram shown is a structural schematic of the vegetation classification device based on a spatial-spectral neural network provided in this embodiment. The vegetation classification device based on a spatial-spectral neural network according to this embodiment includes: The 31 Dataset Construction Module is used to divide hyperspectral datasets from different platforms into training, validation, and test sets, and to set quantitative evaluation metrics.
[0064] The target network architecture search module 32 is used to search for neural network architectures based on a gradient-optimized differentiable architecture search strategy in a spatial-spectral hybrid search space to obtain the target network architecture. The spatial-spectral hybrid search space includes at least an information separation spectral transform operator for extracting spectral features and an information separation spatial attention deep convolution operator for extracting spatial features.
[0065] The 33 classification model training module is used to train the parameters of the target network architecture based on the training set and validation set to obtain the classification model.
[0066] The 34 classification module is used to input the test set into the trained classification model to obtain the vegetation classification results, and to evaluate the classification results based on quantitative evaluation indicators.
[0067] The device provided in this embodiment, through the 31 dataset construction module, independently divides the data into training, validation, and test sets, effectively preventing information leakage during model training and evaluation, and ensuring the authenticity and reliability of performance evaluation results. Pre-defined quantitative evaluation indicators provide objective and unified optimization goals and final performance measurement standards for the entire automated process. The 32 target network architecture search module, with its specially designed spatial-spectral hybrid search space, enables the search process to automatically construct the optimal network architecture for efficiently and collaboratively extracting spectral and spatial features, addressing the problem of poor adaptability of fixed architectures. By combining an improved differentiable search strategy, high-performance structures can be found quickly with lower computational costs, realizing the transformation from manual design to automatic customization. The 33-classification model training module, with its network topology optimized for the current task, allows for more stable and efficient parameter training, enabling the model weights to fully learn the complex mapping from data to vegetation categories. This unlocks the full performance potential of the customized architecture, resulting in a strong predictive model that can be directly used for classification. The 34-classification module, evaluated using a completely independent test set, objectively and unbiasedly quantifies the generalization ability and actual classification accuracy of the final model, thus empirically demonstrating the comprehensive effectiveness of the entire automated process and proving the practical application value of the generated model.
[0068] In one embodiment, the construction of the information separation spectral transformation operator specifically includes the following steps: The input feature map is normalized in terms of spectral dimension to obtain the first scaling factor for each channel.
[0069] Based on the first scaling factor, the information weight of each channel is calculated, and according to the preset separation ratio, the input feature map is separated into a high-information-content feature part and a low-information-content feature part.
[0070] The high-information-content feature portion is input into the spectral Transformer module for processing, and the low-information-content feature portion is subjected to lightweight convolution processing.
[0071] The output of the spectral Transformer module is fused with the output of the lightweight convolution processing to obtain the output of the information separation spectral transformation operator.
[0072] In one embodiment, the construction of the information separation spatial attention deep convolution operator specifically includes the following steps: Spatial dimension normalization is performed on the input feature map to obtain the second scaling factor for each channel.
[0073] The information weights of each channel are calculated based on the second scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio.
[0074] The high-information-content feature portion is input into a multi-dimensional parameterless attention deep convolution module for processing, and the low-information-content feature portion is processed by grouped convolution.
[0075] The output of the multi-dimensional parameterless attention deep convolution module is fused with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator.
[0076] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operations of each module processing step can be understood with reference to the description of the method embodiments, and will not be repeated here.
[0077] like Figure 4 As shown, the present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the vegetation classification method based on the spatial-spectral neural network in the above embodiments, or when the computer program is executed by a processor, it implements the vegetation classification method based on the spatial-spectral neural network in the above embodiments.
[0078] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0079] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and the contents not described in detail in the specification of the present invention are known to those skilled in the art.
Claims
1. A vegetation classification method based on a spatial-spectral neural network, characterized in that, Includes the following steps: Hyperspectral datasets from different platforms were divided into training, validation, and test sets, and quantitative evaluation metrics were set. Based on the information separation spatial-spectral hybrid search space, a gradient-optimized differentiable architecture search strategy is used to search for neural network architectures to obtain the target network architecture. The spatial-spectral hybrid search space includes at least an information separation spectral transform operator for extracting spectral features and an information separation spatial attention deep convolution operator for extracting spatial features. Based on the training set and validation set, the target network architecture is trained to obtain the classification model; The test set is input into the trained classification model to obtain the vegetation classification results, and the classification results are evaluated based on quantitative evaluation indicators.
2. The vegetation classification method based on spatial-spectral neural network as described in claim 1, characterized in that, The construction of the information separation spectral transformation operator specifically includes the following steps: The input feature map is spectrally normalized to obtain the first scaling factor for each channel. The information weights of each channel are calculated based on the first scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into the spectral Transformer module for processing, and the low-information-content feature portion is subjected to lightweight convolution processing; The output of the spectral Transformer module is fused with the output of the lightweight convolution processing to obtain the output of the information separation spectral transformation operator.
3. The vegetation classification method based on spatial-spectral neural network as described in claim 1, characterized in that, The construction of the information separation spatial attention deep convolution operator specifically includes the following steps: The input feature map is spatially normalized to obtain the second scaling factor for each channel. The information weights of each channel are calculated based on the second scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into a multi-dimensional parameterless attention deep convolution module for processing, and the low-information-content feature portion is processed by grouped convolution. The output of the multi-dimensional parameterless attention deep convolution module is fused with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator.
4. The vegetation classification method based on spatial-spectral neural network as described in claim 3, characterized in that, The construction of the multi-dimensional parameterless attention deep convolution module specifically includes the following steps: Calculate the one-dimensional spectral nonparametric attention weights and the two-dimensional spatial nonparametric attention weights of the input feature map, respectively. The one-dimensional spectral parameterless attention weights and the two-dimensional spatial parameterless attention weights are expanded and fused to obtain three-dimensional spatial spectral parameterless attention weights. The input feature map is weighted using the three-dimensional spatial spectrum parameter-free attention weights, and a depthwise convolution operation is performed.
5. The vegetation classification method based on spatial-spectral neural network as described in claim 1, characterized in that, The method of using a gradient-optimized differentiable architecture search strategy to search for neural network architectures specifically includes the following steps: Optimize the weight parameters of each operator in the spatial-spectral hybrid search space on the training set; Optimize continuous structural parameters representing network architecture selection on the validation set; The above optimization process is executed iteratively until convergence, and the target network architecture is determined based on the converged continuous structural parameters.
6. A vegetation classification device based on a spatial-spectral neural network, characterized in that, include: The dataset construction module is used to divide hyperspectral datasets from different platforms into training, validation, and test sets, and to set quantitative evaluation metrics. The target network architecture search module is used to search for neural network architectures based on a gradient-optimized differentiable architecture search strategy in a spatial-spectral hybrid search space to obtain the target network architecture. The spatial-spectral hybrid search space includes at least an information separation spectral transformation operator for extracting spectral features and an information separation spatial attention deep convolution operator for extracting spatial features. The classification model training module is used to train the parameters of the target network architecture based on the training set and the validation set to obtain the classification model; The classification module is used to input the test set into the trained classification model to obtain the vegetation classification results, and to evaluate the classification results based on quantitative evaluation indicators.
7. The vegetation classification device based on spatial-spectral neural network as described in claim 6, characterized in that, The construction of the information separation spectral transformation operator specifically includes the following steps: The input feature map is spectrally normalized to obtain the first scaling factor for each channel. The information weights of each channel are calculated based on the first scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into the spectral Transformer module for processing, and the low-information-content feature portion is subjected to lightweight convolution processing; The output of the spectral Transformer module is fused with the output of the lightweight convolution processing to obtain the output of the information separation spectral transformation operator.
8. The vegetation classification device based on spatial-spectral neural network as described in claim 5, characterized in that, The construction of the information separation spatial attention deep convolution operator specifically includes the following steps: The input feature map is spatially normalized to obtain the second scaling factor for each channel. The information weights of each channel are calculated based on the second scaling factor, and the input feature map is separated into a high-information-content feature part and a low-information-content feature part according to the preset separation ratio. The high-information-content feature portion is input into a multi-dimensional parameterless attention deep convolution module for processing, and the low-information-content feature portion is processed by grouped convolution. The output of the multi-dimensional parameterless attention deep convolution module is fused with the output of the grouped convolution processing to obtain the output of the information separation spatial attention deep convolution operator.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the vegetation classification method based on the spatial-spectral neural network as described in any one of claims 1 to 5.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the vegetation classification method based on the spatial-spectral neural network as described in any one of claims 1 to 5.