Hyperspectral image segmentation method based on spectrum-spatial feature combination

Abstract

The invention relates to the technical field of image target detection, and discloses a hyperspectral image segmentation method based on spectrum-spatial feature combination, comprising the following steps: 1) acquiring a hyperspectral image data set; 2) performing hyperspectral image segmentation based on multi-scale space-spectral feature combination of deep learning; and 3) evaluating a network segmentation effect. The hyperspectral image is segmented by using a deep learning method, the spectral features and the spatial features of the hyperspectral image are effectively extracted and fused by providing a spectral-spatial feature combined segmentation method, and meanwhile, the prediction accuracy of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image target detection, and in particular to a hyperspectral image segmentation method based on spectral-spatial feature combination. Background Art

[0002] As an important aircraft, drones have the advantages of low cost, high flexibility, fast image acquisition speed, and high spatial resolution compared to satellites or large aircraft. UAV platforms can carry a range of remote sensing instruments, including visible light, multispectral and thermal infrared remote sensors. The rapid development of remote sensing using drones has greatly simplified research and practical applications in various fields. With the rapid advancement of artificial intelligence technology, drones can achieve autonomous target recognition through AI technology, thereby achieving more intelligent flight and mission execution. Therefore, they have been widely used in agriculture, environmental monitoring, security and other fields. The confrontation between reconnaissance and camouflage in modern warfare has become more intense with the rapid development of aerospace technology, optoelectronic technology, and material technology. Camouflage, as an important means of survival for military targets, has become the focus of attention of the military forces of all countries in the world. The camouflage methods currently used include camouflage uniforms, paint, and camouflage nets. Traditional reconnaissance methods such as visible light and infrared have various limitations in the identification of camouflaged targets and cannot effectively identify camouflaged targets. With the development of hyperspectral remote sensing technology, the spectral resolution has been continuously improved, and it has become possible to conduct reconnaissance on targets in hundreds of bands. At the same time, the continuous spectral information of each pixel in the image is well preserved, and contains the spatial characteristics of the target, which contains a lot of deep dimensional information. These characteristics have greatly improved the information richness of hyperspectral data compared to ordinary color images, and have more advantages than other detection methods in the detection, identification, and classification of camouflaged targets.

[0003] Therefore, in order to meet the highly complex battlefield conditions in the future, cope with the needs of high real-time response and high secrecy in wars, and meet the needs for agile detection of camouflaged targets, we should take advantage of the miniaturization and intelligence of drones in modern warfare, and use the rich spectral information and high spatial resolution of spectral imaging to capture more detailed surface information and improve the detection capability of camouflaged targets, which is of great significance to the current intelligent detection methods of camouflaged targets.

[0004] A large number of deep learning models, represented by convolutional neural networks, have been applied to the intelligent information extraction task of remote sensing images. However, compared with visible light remote sensing images, hyperspectral remote sensing images contain richer spectral and spatial information. When these methods are directly applied to hyperspectral images, the spectral and spatial information in the hyperspectral images cannot be fully and effectively utilized, resulting in suboptimal segmentation accuracy. Secondly, there are large scale variations of different types of targets in hyperspectral images. How to effectively extract target features of different scales is also one of the difficulties in improving segmentation accuracy. Finally, since the camouflaged target has similar characteristics to the background environment and the contour boundary is more blurred, the difficulty of identifying the camouflaged target will also increase. Summary of the invention

[0005] To solve these problems, this paper proposes a hyperspectral image segmentation method based on the combination of spectral and spatial features.

[0006] The technical solution of the present invention is as follows:

[0007] A hyperspectral image segmentation method based on spectral-spatial feature combination includes the following steps:

[0008] 1) Obtain hyperspectral image dataset:

[0009] The dataset is collected by a hyperspectral video camera, and the experimental scene consists of green grass and green painted camouflaged vehicles.

[0010] 2) Hyperspectral image segmentation based on multi-scale spatial-spectral features:

[0011] Construct a multi-scale spatial-spectral feature fusion network model:

[0012] These include:

[0013] a) A multi-scale convolution window and spectral attention module are proposed to fully integrate the spectral and spatial information contained in the hyperspectral image in the downsampling stage, thereby improving the semantic segmentation accuracy of the hyperspectral image;

[0014] b) UAV remote sensing images have different object scales. By adding a new multi-scale feature extraction module, the segmentation performance of ground objects at different scales is improved;

[0015] c) The semantic information between different levels is integrated, and the deep semantic information is used as a guide for the shallow semantic information, so as to strengthen the connection between categories, realize the complementarity of spatial and geometric information, refine the segmentation edges, and thus enhance the spatial detail information and the recognition ability of camouflaged targets.

[0016] 3) Network segmentation effect evaluation

[0017] The dataset is divided into a training set and a test set. The test images are input into SegNet, PspNet, DeepLabV3+ and a multi-scale spatial-spectral feature fusion network model for comparative experiments. In addition to changing the network model, the other experimental conditions of the model comparison experiment remain the same, including the dataset, data enhancement method, etc. The effect of this method is verified by comparing the pixel accuracy PA, mean intersection-over-union ratio Miou and recognition result graph with other segmentation methods.

[0018] Further, the process of step 2) is:

[0019] Step 2.1: Multi-scale convolutional windows:

[0020] In order to effectively extract spectral features in hyperspectral images and combine them with spatial features in the downsampling stage of the backbone network, a multi-scale convolution window is proposed as the first convolution layer of the hyperspectral image input.

[0021] The sizes of convolution kernels are 1×1, 3×3, and 5×5. 1×1 convolution is used to extract the spectral features of hyperspectral images, while 3×3 and 5×5 convolutions are used to extract local spatial features of images. The feature maps extracted by convolution kernels of different scales are combined to form a joint spatial-spectral feature map, which is then subjected to 3×3 maximum pooling as the input of subsequent convolution layers.

[0022] Step 2.2: Spectral Attention Module:

[0023] The channel attention mechanism can generate additional weight information on each channel of the feature. Therefore, the channel attention is introduced after the residual structure of Res2Net to extract key band features and suppress non-key band data.

[0024] Design the spectral channel attention module: where X∈R C×H×W is the input feature map of the channel attention module, R is a real number set; H, W, C represent the height, width and number of channels of the feature map respectively. sq F represents a dimension compression operation, which uses global average pooling to compress the spatial dimension of the hyperspectral image to obtain a feature map of size 1×1×C. This method can better capture the correlation between different bands. ex For the excitation operation, it consists of two 1×1 convolutional layers, a ReLU activation function, and a Sigmoid activation function. First, the dimension is reduced to the original 1 / r, and then the dimension is restored to C channels through the next fully connected layer, where r represents the compression multiple. Through these two convolutional layers, the network can learn the band characteristics, so that the output weights can reflect the importance differences of different bands. Finally, F scale operate, The symbol represents channel-by-channel multiplication. This step is similar to the gating mechanism, which adaptively assigns corresponding weights to each channel of the input feature map.

[0025] Step 2.3: Multi-scale dense feature fusion module:

[0026] Multi-scale dense feature fusion module: The dilated convolution can obtain a larger receptive field than the ordinary convolution kernel. For a dilated convolution with a dilation rate of d and a convolution kernel size of K, the equivalent receptive field size can be expressed as:

[0027] R=(d-1)×(K-1)+K (1)

[0028] The different scale features of the input features are captured by dilated convolutions with different dilation rates. The dilated convolution layers with different dilation rates are combined in a cascade manner. The input and output features of each layer are concatenated with all outputs from the lower layers. Each dilation layer can be expressed as follows:

[0029]

[0030] where d l represents the expansion rate of a layer; [...] represents the concatenation operation between feature layers; [X l-1 ,X l-2 ,...,X 0 ] represents the output of all layers before concatenation. This dense connection can form a denser feature pyramid and a larger receptive field. At the same time, adaptive average pooling can integrate extensive spatial information and prevent overfitting. Therefore, global average pooling is added to this module, and these six branches are densely connected backwards.

[0031] Step 2.4: Feature fusion module:

[0032] Feature fusion module, which realizes the effective fusion of low-level semantic features and deep-level semantic features;

[0033] First, in order to make accurate use of spatial detail information and semantic information, the two features are weighted and fused as the input of the feature fusion module; the weighted fusion operation adaptively adjusts the weight according to the contribution of the two features, which can be specifically expressed as:

[0034] F=α·F Low +(1-α)·F High (3)

[0035] Among them, F Low Represents the shallow features of the output of ResBlock1, F High represents the deep features after the multi-scale fusion module, and F represents the fused features.

[0036] In order to further enhance the representation of spectral, multi-scale and spatial features and improve the accuracy of edge detail segmentation of camouflaged targets, three branches are used to analyze the input feature X∈R C×H×W is applied. Specifically, the spectral channel attention is the same as that in the encoder, and global average pooling is first used to generate a 1×1×C channel attention map. The channel size is reduced by 1×1 convolution and restored to the original channel size after the ReLu activation function and 1×1 convolution. Then, the Sigmoid activation function is used to obtain the channel weight distribution, which is multiplied with the original input element to obtain a feature map containing channel attention. The multi-scale branch first uses 1×1 convolution to reduce the channel dimension, and then extracts multi-scale information through 3 parallel separable dilated convolutions, where the convolution kernel size is 3×3 and the dilation rate is R. i (i=1,2,3) are 7, 9, and 11 respectively. After being restored to the original channel dimension by 1×1 convolution, the three parallel results are then fused by a sum operation. The output of the multi-scale branch is obtained by summing the 1×1 convolution with a residual connection. The spatial branch first generates a feature map using maximum pooling and average pooling, and then connects in the channel dimension to generate a spatial attention map of size 2×H×W. The spatial weight distribution is obtained by 7×7 convolution and Sigmoid activation function, which is multiplied with the original input element to obtain a feature map containing spatial attention. Finally, the results of the three branches are fused by a sum operation to produce the final feature map.

[0037] Further, the process of step 3) is:

[0038] Step 3.1: During training, fix the pixel size of all images to 360×480, set the training batch size to 8, train for 150 cycles, train the 2310 training sets in batches of 200, and verify the network on the validation set. Save the network weights after training. The optimizer uses the Adam optimizer, with an initial learning rate of 0.001 and a learning rate decay of 10% every ten batches. The loss function uses the cross entropy function.

[0039] In order to evaluate the accuracy of model segmentation, we use the average intersection over union (Miou), pixel accuracy (Pixelaccuracy), and F1 score (F1 Score) to measure the segmentation performance, as shown in formula (4-6). TP stands for true positive, FP stands for false positive, TN stands for true negative, and FN stands for false negative.

[0040]

[0041] In the formula, Precision represents the precision, Recall represents the recall, and the specific calculation is shown in formula (7-8).

[0042]

[0043] In order to illustrate the effectiveness of the proposed algorithm, the test images were input into SegNet, PspNet, DeepLabV3+ and the multi-scale spatial-spectral feature fusion network model for comparative experiments, and several predicted images were randomly selected for visualization.

[0044] The beneficial effects of the present invention are mainly manifested in: the present invention uses a deep learning method to segment hyperspectral images, and proposes a segmentation method that combines spectral and spatial features to effectively extract and fuse the spectral features and spatial features of the hyperspectral images at the encoder stage, and extracts multi-scale spatial features through a multi-scale dense feature fusion module. At the same time, at the decoder stage, deep semantic features are used to guide shallow spatial details to improve detail segmentation capabilities and improve the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the overall structure diagram of the network proposed by the present invention;

[0046] Figure 2 It is a multi-scale convolution window structure diagram proposed by the present invention;

[0047] Figure 3 It is a spectral channel attention structure diagram proposed by the present invention;

[0048] Figure 4 It is a structural diagram of the multi-scale dense feature fusion module proposed in the present invention;

[0049] Figure 5 It is a structural diagram of the feature fusion module proposed in the present invention;

[0050] Figure 6 This is a comparison chart of the recognition results proposed by the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings.

[0052] Reference Figure 1 to Figure 6 , a hyperspectral image segmentation method based on spectral-spatial feature combination, comprising the following steps:

[0053] Step 1) Obtain a hyperspectral image dataset of camouflaged vehicles:

[0054] The dataset was collected by Lingyunguang HV-125 series hyperspectral video camera. The experimental scene consists of green grass and green painted camouflaged vehicles. The image size is 1387 pixels × 1829 pixels. It has 128 spectral response bands in the wavelength range of 450-950 nanometers and a spectral resolution of 4 nanometers.

[0055] Step 2) Hyperspectral image segmentation based on multi-scale spatial-spectral feature combination

[0056] The spatial-spectral feature fusion network of the scale includes the following three aspects:

[0057] 1) A multi-scale convolution window and spectral attention module are proposed to fully integrate the spectral and spatial information contained in the hyperspectral image in the downsampling stage, thereby improving the semantic segmentation accuracy of the hyperspectral image; 2) There are different target scales in UAV remote sensing images. By adding a new multi-scale feature extraction module, the segmentation performance of ground objects at different scales is improved; 3) The semantic information between different levels is fused, and the deep semantic information is used as a guide for the shallow semantic information to strengthen the connection between categories, realize the complementarity of spatial and geometric information, and refine the segmentation edges, thereby enhancing the spatial detail information and the ability to recognize camouflaged targets.

[0058] Step 2.1: In order to effectively extract the spectral features in the hyperspectral image and combine them with the spatial features in the downsampling stage of the backbone network, a multi-scale convolution window is proposed as the first convolution layer of the hyperspectral image input;

[0059] The sizes of convolution kernels are 1×1, 3×3, and 5×5. 1×1 convolution is used to extract the spectral features of hyperspectral images, while 3×3 and 5×5 convolutions are used to extract local spatial features of images. The feature maps extracted by convolution kernels of different scales are combined to form a joint spatial-spectral feature map, which is then subjected to 3×3 maximum pooling as the input of subsequent convolution layers.

[0060] Step 2.2: The channel attention mechanism can generate additional weight information on each channel of the feature. Therefore, the channel attention is introduced after the residual structure of Res2Net to extract key band features and suppress non-key band data.

[0061] Inspired by SENet, this paper designs a spectral channel attention. C×H×W is the input feature map of the channel attention module, R is a real number set; H, W, C represent the height, width and number of channels of the feature map respectively. sqF represents a dimension compression operation, which uses global average pooling to compress the spatial dimension of the hyperspectral image to obtain a feature map of size 1×1×C. This method can better capture the correlation between different bands. ex For the excitation operation, it consists of two 1×1 convolutional layers, a ReLU activation function, and a Sigmoid activation function. First, the dimension is reduced to the original 1 / r, and then the dimension is restored to C channels through the next fully connected layer, where r represents the compression multiple. Through these two convolutional layers, the network can learn the band characteristics, so that the output weights can reflect the importance differences of different bands. Finally, F scale operate, The symbol represents channel-by-channel multiplication. This step is similar to the gating mechanism, which adaptively assigns corresponding weights to each channel of the input feature map.

[0062] Step 2.3: Multi-scale dense feature fusion module

[0063] Multi-scale dense feature fusion module, the internal structure of this module is as follows Figure 4 As shown in the figure, the hole convolution can obtain a larger receptive field than the ordinary convolution kernel. The hole convolution with different expansion rates is used to capture the different scale features of the input features. The hole convolution layers with expansion rates of (3, 5, 7, 9) are combined in a cascade manner. The input and output features of each layer are concatenated with all outputs from the lower layers. This dense connection can form a denser feature pyramid and a larger receptive field, and can capture more sampling points for fewer categories and smaller target objects. At the same time, adaptive average pooling is added to this module, which can integrate a wide range of spatial information and prevent overfitting. The six branches are tightly connected backwards.

[0064] Step 2.4: Feature Fusion Module

[0065] Feature fusion module, see the attached for its specific structure Figure 5 , to achieve the effective fusion of low-level semantic features and deep semantic features.

[0066] First, in order to make accurate use of spatial detail information and semantic information, the two features are weighted and fused as the input of the feature fusion module; the weighted fusion operation adaptively adjusts the weight according to the contribution of the two features, which can be specifically expressed as:

[0067] F=α·F Low +(1-α)·F High (3)

[0068] Among them, F Low Represents the shallow features of the output of ResBlock1, F Highrepresents the deep features after the multi-scale fusion module, and F represents the fused features.

[0069] In order to further enhance the representation of spectral, multi-scale and spatial features and improve the accuracy of edge detail segmentation of camouflaged targets, three branches are used to analyze the input feature X∈R C×H×W is applied. Specifically, the spectral channel attention is the same as that in the encoder, and global average pooling is first used to generate a 1×1×C channel attention map. The channel size is reduced by 1×1 convolution and restored to the original channel size after the ReLu activation function and 1×1 convolution. Then, the Sigmoid activation function is used to obtain the channel weight distribution, which is multiplied with the original input element to obtain a feature map containing channel attention. The multi-scale branch first uses 1×1 convolution to reduce the channel dimension, and then extracts multi-scale information through 3 parallel separable dilated convolutions, where the convolution kernel size is 3×3 and the dilation rate is R. i (i=1,2,3) are 7, 9, and 11 respectively. After being restored to the original channel dimension by 1×1 convolution, the three parallel results are then fused by a sum operation. The output of the multi-scale branch is obtained by summing the 1×1 convolution with a residual connection. The spatial branch first generates a feature map using maximum pooling and average pooling, and then connects in the channel dimension to generate a spatial attention map of size 2×H×W. The spatial weight distribution is obtained by 7×7 convolution and Sigmoid activation function, which is multiplied with the original input element to obtain a feature map containing spatial attention. Finally, the results of the three branches are fused by a sum operation to produce the final feature map.

[0070] Step 3) Network segmentation effect evaluation

[0071] The dataset is divided into a training set and a test set. The test images are input into SegNet, PspNet, DeepLabV3+ and the model proposed in this paper for comparative experiments. In addition to changing the network model, the other experimental conditions remain the same, including the dataset, data enhancement method, etc. The effect of this method is verified by comparing the pixel accuracy PA, mean intersection-over-union ratio Miou and recognition result graph with other segmentation methods.

[0072] Step 3.1: During training, fix the pixel size of all images to 360×480, set the training Batch Size to 8, train the 2310 training sets in batches of 200, and verify the network on the validation set. Save the network weights after training. Use the Adam optimizer as the optimizer, with an initial learning rate of 5e-6 and a learning rate decay of 10% every ten batches. Use the cross entropy function as the loss function.

[0073] In order to evaluate the accuracy of model segmentation, we use the average intersection over union (Miou), pixel accuracy (Pixelaccuracy), and F1 score (F1 Score) to measure the segmentation performance, as shown in formula (4-6). TP stands for true positive, FP stands for false positive, TN stands for true negative, and FN stands for false negative.

[0074]

[0075] In the formula, Precision represents the precision, Recall represents the recall, and the specific calculation is shown in formula (7-8).

[0076]

[0077] In order to illustrate the effectiveness of the proposed algorithm, this paper inputs the test images into SegNet, PspNet, DeepLabV3+ and the model proposed in this paper for comparative experiments, and then selects several predicted images for visualization. Figure 6 Shown is the comparison between the current network and the network prediction results; Figure 6 (a) is the original image, (b) is the true value image of the original image, (c) is the recognition result image of DeepLabV3+, (d) is the recognition result image of PspNet, (e) is the recognition result image of SegNet, and (f) is the recognition result image of the present invention.

[0078] In the model comparison experiment, except for changing the network model, the other experimental conditions remain the same, including the data set, data enhancement method, etc. The specific experimental results are shown in Table 1

[0079] Table 1 Comparative experiment

[0080]

[0081] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A hyperspectral image segmentation method based on spectral-spatial feature combination, characterized in that: The following steps are involved: 1) Obtain hyperspectral image dataset: Set up the experimental scene and use a hyperspectral video camera to acquire a hyperspectral image dataset; 2) Hyperspectral image segmentation based on multi-scale spatial-spectral features: Construct a multi-scale spatial-spectral feature fusion network, including: Multi-scale convolution window: It can extract channel and spatial features in the image at the same time, and fully integrate the spectral and spatial information contained in the hyperspectral image in the downsampling stage; Spectral Attention Module: The spectral attention module adaptively recalibrates the weight of each channel by generating attention masks on all feature channels, improving the semantic segmentation accuracy of hyperspectral images; The multi-scale dense feature fusion module improves the segmentation performance of ground objects at different scales by fusing and extracting multi-scale context features at the encoder stage; The feature fusion module fuses semantic features at different levels and uses deep semantic features as a guide for shallow semantic features, strengthening the connection between categories, achieving the complementarity of spatial and geometric information, and refining segmentation edges, thereby enhancing spatial detail information and the ability to recognize camouflaged targets; 3) Network segmentation effect evaluation: The dataset is divided into a training set and a test set. The test set images are input into SegNet, PspNet, DeepLabV3+ and a multi-scale spatial-spectral feature fusion network model for comparative experiments.

2. The hyperspectral image segmentation method based on spectral-spatial feature combination as claimed in claim 1, characterized in that: The process of step 2) is: Step 2.1: A multi-scale convolution window is used as the first convolution layer of the hyperspectral image input, and the sizes of the convolution kernels are 1×1, 3×3, and 5×5 respectively; the 1×1 convolution is used to extract the spectral features of the hyperspectral image, and the 3×3 and 5×5 convolutions are used to extract the local spatial features of the image; The feature maps extracted by convolution kernels of different scales are combined to form a joint spatial-spectral feature map, and then a 3×3 maximum pooling is performed as the input of the subsequent convolution layer; Step 2.2: The backbone network uses ResNet50, and the channel attention mechanism is introduced after the residual structure of the backbone network Res2Net to extract the feature bands, while suppressing the non-feature band data; Spectral channel attention module: X∈R C×H×W is the input feature map of the channel attention mechanism, R is a set of real numbers; H, W, and C represent the height, width, and number of channels of the feature map respectively; Dimensionality compression operation F sq , which uses global average pooling to compress the spatial dimension of the hyperspectral image and obtain a channel attention map, which is a feature map of size 1×1×C; Stimulus Operation F ex , which consists of two 1×1 convolutional layers, a ReLU activation function and a Sigmoid activation function; first, the channel dimension is reduced to the original 1 / r through a 1×1 convolutional layer, where r represents the compression multiple; then a ReLU activation function and a 1×1 convolutional layer are used to restore the dimension to C channels; and then the Sigmoid activation function is used to obtain the weight distribution of different channels; Multiplication operation F scale Operation, multiply the weight distribution by the original input element by element, adaptively assign corresponding weights to each channel of the input feature map, and obtain features containing channel attention; Step 2.3: Multi-scale dense feature fusion module: Capture input features X∈R using dilated convolutions with different dilation rates C×H×W For different scale features of , for a dilation rate d and a convolution kernel size K, the equivalent receptive field size is expressed as: R=(d-1)×(K-1)+K (1) The atrous convolutional layers with different dilation rates are combined in a cascaded manner, and the input and output features of each layer are concatenated with all outputs from the previous layer. Each dilation layer is represented as follows: where d l represents the expansion rate of a layer; [...] represents the concatenation operation between feature layers; [X l-1 ,X l-2 ,...,X0] represents the output of all layers before concatenation; Step 2.4: Feature fusion module: The shallow semantic features and deep semantic features are weighted and fused as the input of the feature fusion module. The weighted fusion operation adaptively adjusts the weight according to the contribution of the two features, which is expressed as: F=α·F Low +(1-α)·F High (3) Among them, F Low Represents the shallow features of the output of ResBlock1, F High represents the deep features after the multi-scale fusion module, and F represents the fused features; Use global average pooling to generate a 1×1×C channel attention map; the channel size is reduced by 1×1 convolution and restored to the original channel size after ReLu activation function and 1×1 convolution; then, use Sigmoid activation function to obtain channel weight distribution, which is multiplied with the original input element to obtain a feature map containing channel attention; The multi-scale branch first uses 1×1 convolution to reduce the channel dimension, and then extracts multi-scale information through three parallel separable dilated convolutions, where the convolution kernel size is 3×3 and the dilation rate is R. i (i=1,2,3) are 7, 9, and 11 respectively; after being restored to the original channel dimension by 1×1 convolution, the three parallel results are then fused by summation operation; the output of the multi-scale branch is obtained by summing the 1×1 convolution and residual connection; The spatial branch first generates a feature map using maximum pooling and average pooling, and then concatenates in the channel dimension to generate a spatial attention map of size 2×H×W; the spatial weight distribution is obtained through a 7×7 convolution and a Sigmoid activation function, which is multiplied with the original input element to obtain a feature map containing spatial attention; finally, the results of the three branches are fused through a sum operation to produce the final feature map.

3. The hyperspectral image segmentation method based on spectral-spatial feature combination as claimed in claim 1, characterized in that: The specific process of step 3) is as follows: Step 3.1: During training, the training set is enhanced by random rotation, random flipping and random spatial affine transformation. The pixel size of all images is fixed, the batch size and cycle of training are set, and the training set is trained in batches. The network is verified on the validation set, and the network weights are saved after training. The Adam optimizer is used, the initial learning rate is set to 0.001, the cosine decay strategy is used to adjust the learning rate, and the cross entropy function is used as the loss function. The segmentation performance is measured using the average intersection-over-union ratio (Miou), pixel accuracy, and F1 score (F1 Score), as shown in formulas (3)-(5); Where TP stands for true positive, FP stands for false positive, TN stands for true negative, and FN stands for false negative; In the formula, Precision represents the precision rate, Recall represents the recall rate, and the specific calculation is shown in formulas (6)-(7); The test set images are input into SegNet, PspNet, DeepLabV3+ and multi-scale spatial-spectral feature fusion network models for comparative experiments, and several predicted images are randomly selected for visualization.

Images