A regional vegetation type detection method and system

Abstract

The application discloses a regional vegetation type detection method and system, and the method comprises the following steps: acquiring regional remote sensing image data and performing pretreatment; performing hierarchical feature extraction and fusion according to the pretreated regional remote sensing image data to obtain fused features; and constructing a deep hierarchical semantic segmentation model coupled with vegetation type prior knowledge to input the fused features and output vegetation type classification results; wherein the vegetation type prior knowledge is a hierarchical structure of vegetation types. The technical scheme provided by the application can extract features from the perspectives of space, channels and scales respectively based on an attention mechanism, can realize extraction of differentiated features of the remote sensing image data from multiple dimensions, and can improve the classification effect; by constructing the deep hierarchical semantic segmentation model coupled with the vegetation type prior knowledge, the problem of uneven distribution of vegetation type samples can be overcome, the training speed can be improved, the classification effect can be optimized, and the detection precision can be improved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a method and system for detecting regional vegetation types. Background Technology

[0002] Vegetation type detection is an essential part of forest resource surveys and monitoring. By comprehensively using multi-source, multi-temporal, and multispectral remote sensing data, vegetation type information can be obtained, providing important scientific basis for the rational utilization and protection of forest resources in my country.

[0003] However, current methods for vegetation type detection using this type of data typically employ common data fusion techniques, often neglecting the differentiated characteristics of multi-source, multi-temporal, and multispectral remote sensing images. Furthermore, limitations imposed by the objective distribution of vegetation and the constraints of field survey environments often lead to reduced classification accuracy due to uneven distribution of training samples.

[0004] Furthermore, vegetation type detection typically involves broad classifications such as forests, shrubs, and grasses, lacking more granular analysis and classification. This general approach can only assess the vegetation overview of a large area and is inadequate for in-depth research on ecosystem health, biodiversity conservation, and precision agriculture management. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the current technology, the present invention provides a regional vegetation type detection method. By hierarchically extracting and fusing features from remote sensing image data, a multi-level semantic segmentation model coupled with prior knowledge of vegetation type is constructed, which can optimize the classification effect and improve the detection accuracy.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: A method for detecting regional vegetation types includes the following steps: Acquire regional remote sensing image data and perform preprocessing; Based on the preprocessed regional remote sensing image data, hierarchical feature extraction and fusion are performed to obtain fused features; A deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types is constructed to input fused features and output vegetation type classification results; Among them, the prior knowledge of vegetation types is the hierarchical structure of vegetation types.

[0007] According to one aspect of the present invention, the step of performing hierarchical feature extraction and fusion based on preprocessed regional remote sensing image data to obtain fused features includes: A spatiotemporal feature extraction module is constructed based on joint spatial attention and channel attention to extract spatiotemporal features from preprocessed regional remote sensing image data.

[0008] According to one aspect of the present invention, the step of performing hierarchical feature extraction and fusion based on preprocessed regional remote sensing image data to obtain fused features includes: A scale attention module is constructed to calculate the contribution weights of spatiotemporal features at different scales, and the spatiotemporal features of different image data are fused to obtain fused features.

[0009] According to one aspect of the present invention, constructing a deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types includes: Obtain the multi-level structure of vegetation types; Based on fusion features, and according to the number of vegetation type levels N, an N-level semantic segmentation model coupled with prior knowledge of vegetation type is constructed using the U-Net model as the basic architecture.

[0010] According to one aspect of the present invention, the N-level semantic segmentation model includes: The previous layer of the N-level semantic segmentation model serves as a constraint condition for the next layer. The output of each level of the N-level semantic segmentation model corresponds to the vegetation type result.

[0011] According to one aspect of the present invention, the N-level semantic segmentation model includes: The loss function of the N-level semantic segmentation model is obtained by considering the loss of all levels and weighting and summing the loss of the coarse-grained type level and the loss of the fine-grained type level. During model training, the distribution of loss weights is controlled.

[0012] According to one aspect of the present invention, the construction of a deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types to input fused features and output vegetation type classification results further includes: Based on the hierarchical structure of vegetation types, the classification results of vegetation types are cascaded and fused to construct an optimized model and obtain optimized vegetation type detection results.

[0013] According to one aspect of the present invention, the regional vegetation type detection method further includes: The detection results were verified by combining samples collected in the field, and the deep hierarchical semantic segmentation model was fine-tuned and improved.

[0014] According to one aspect of the present invention, the regional vegetation type detection method further includes: Based on the vegetation type detection results, a regional vegetation type distribution map is constructed.

[0015] A regional vegetation type detection system, characterized in that it is based on the regional vegetation type detection method described above, comprising: The data acquisition module is used to acquire regional remote sensing image data and perform preprocessing. The feature extraction module is used to perform hierarchical feature extraction and fusion to obtain fused features based on the preprocessed regional remote sensing image data; The detection module is used to construct a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types to input fused features and output vegetation type classification results.

[0016] Advantages of implementing this invention: This invention provides a regional vegetation type detection method. Based on the attention mechanism, it can extract features from spatial, channel, and scale perspectives, enabling the extraction of differentiated features from remote sensing image data from multiple dimensions and improving classification performance. By constructing a deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types, it can overcome the problem of uneven distribution of vegetation type samples, improve training speed, optimize classification performance, and enhance detection accuracy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a regional vegetation type detection method according to Embodiment 1 of the present invention; Figure 2 The flowchart of the vegetation type detection method described in this invention is as follows; Figure 3 This is the network structure of the spatiotemporal feature extraction module described in this invention; Figure 4 This is the three-level semantic segmentation model described in this invention; Figure 5 This is a flowchart of a regional vegetation type detection method according to Embodiment 2 of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1 and 2As shown, a method for detecting regional vegetation types includes the following steps: S1: Acquire regional remote sensing image data and perform preprocessing.

[0021] In practical applications, the regional remote sensing image data collected are multi-temporal GF-1 WFV, GF-1 PMS and GF-2 PMS images.

[0022] GF-1 WFV, GF-1 PMS, and GF-2 PMS imagery refer to several different types of remote sensing imagery data acquired by Earth observation satellites launched by China. GF stands for the "Gaofen" series of satellites, which are high-resolution satellites capable of acquiring high-resolution images of the Earth's surface for various scientific research and applications. Specifically, "GF-1" and "GF-2" represent the Gaofen-1 and Gaofen-2 satellites, respectively, while "WFV" and "PMS" represent different types of sensors.

[0023] WFV is an abbreviation for "Wide Field of View." This type of sensor can cover a wide area of ​​the ground and is suitable for large-scale environmental monitoring and resource surveys.

[0024] PMS is an abbreviation for "Panchromatic and Multispectral Camera". This type of sensor can acquire panchromatic and multispectral images. Panchromatic images have high resolution, while multispectral images can provide spectral information about ground features, which helps in land cover classification and environmental monitoring.

[0025] In practical applications, preprocessing methods for regional remote sensing image data include: noise removal, geometric correction, radiometric calibration, atmospheric correction, and image registration.

[0026] S2: Based on the preprocessed regional remote sensing image data, hierarchical feature extraction and fusion are performed to obtain fused features.

[0027] Preferably, the feature extraction part of this method achieves hierarchical feature extraction and fusion by constructing a spatiotemporal feature extraction module and a scale attention module. In this method, the feature extraction part is also referred to as the hierarchical scale-sensitive attention spatiotemporal feature extraction model.

[0028] Attention mechanisms in computer vision can extract differentiated features from multiple dimensions. Based on attention mechanisms, features can be extracted from spatial, channel, and scale perspectives, which can enhance the differentiated features of remote sensing images and improve classification performance.

[0029] 1. Spatiotemporal Feature Extraction Module A spatiotemporal feature extraction module based on joint spatial attention and channel attention is constructed to extract spatiotemporal features from preprocessed regional remote sensing image data.

[0030] Specifically, the network structure of the spatiotemporal feature extraction module is as follows: Figure 3 As shown, the input feature is A, and the final output feature is E. The dimensions of A and E are both c×h×w, where c is the number of channels, h is the number of pixels in the image height, and w is the number of pixels in the image width.

[0031] Preferably, the spatiotemporal feature extraction module includes: (1) Spatial attention module, which considers the correlation of different locations in the image and extracts the spatial features of the regional remote sensing image data.

[0032] like Figure 3 As shown, the spatial attention module takes feature A as input and generates three new feature maps B, F, and D through convolution. The dimensions of B, F, and D are the same as those of A and E, which are all c×h×w. B and F are then processed to generate feature S, which is a 2h×w feature map. Finally, after processing, feature P is output, and the dimension of P is the same as that of E, which is also c×h×w. Specifically, the spatial attention model is calculated as follows:

[0033]

[0034] Where i and j represent the positions of the i-th and j-th pixels in the image; α is the scale parameter, initially set to 0; A j P is the input feature at position j; j S is the output feature at position j; ji This is the core computational step of the spatial attention model, which measures the correlation between the i-th position and the j-th position; M is the number of samples.

[0035] (2) Channel attention module, which considers the correlation between different channels in the image and extracts the channel features of the regional remote sensing image data.

[0036] The input image is usually an RGB image with 3 channels. However, remote sensing images are different from RGB images, having multiple channels, also known as bands. High-resolution images typically have 4 bands, i.e., c=4; other sensors, such as Sentinel-2, have 11 bands, i.e., c=11. Remote sensing equipment can capture electromagnetic waves of different wavelength ranges, with each band corresponding to data within a specific wavelength range.

[0037] like Figure 3As shown, the channel attention module takes feature A as input, performs intermediate operations to generate feature X, and X has a dimension of c×c; finally, after further operations, it outputs feature Q, whose dimension is the same as that of E, also c×h×w. Specifically, the calculation method of the channel attention module is as follows:

[0038]

[0039] Where i and j represent the i-th and j-th bands of the image; β is the weighting parameter, initially set to 0; A j The input feature at position j; Q j X is the output feature at the j-th position; ji The correlation between the i-th band and the j-th band was measured.

[0040] (3) Add the outputs of the spatial attention module and the channel attention module to obtain the final output feature E.

[0041] By integrating the processing results of the spatial attention module and the channel attention module, not only is the spatial consistency of pixels of the same category improved, but the band distinguishability is also enhanced, enabling the extraction of spatiotemporal features from remote sensing input images.

[0042] 2. Scale Attention Module By constructing a scale attention module, the contribution weights of spatiotemporal features at different scales are calculated, and the spatiotemporal features of different image data are fused to obtain fused features.

[0043] Specifically, by constructing a scale-attention model, the contribution weights of spatiotemporal features generated from remote sensing images at different spatial resolutions to the final result are learned. These spatiotemporal features are then weighted and summed to achieve spatiotemporal feature fusion. The weights are calculated as follows:

[0044] Where k∈{1,...,K}, represents the level to which remote sensing images of different spatial resolutions belong. In specific implementation, there are a total of 5 scale levels, K=5, k∈{1,...,5}; w k For weights; h k h t These are the classification probability results at different scale levels.

[0045] S3: Construct a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types to input fused features and output vegetation type classification results.

[0046] Among them, the prior knowledge of vegetation types is the hierarchical structure of vegetation types.

[0047] This method uses a semantic segmentation network to detect vegetation types. Before constructing the semantic segmentation network, the vegetation types need to be classified. Then, the remote sensing image data is labeled to construct a sample dataset.

[0048] In forest ecosystems, traditional vegetation type detection only classifies vegetation into broad categories such as forest, shrub, and grass, which is too general. For example, the single category of "forest" cannot effectively distinguish between different types such as coniferous forests, broad-leaved forests, and mixed forests, which differ significantly in ecological processes, species composition, and responses to environmental changes. Therefore, more granular detection and classification of vegetation types are needed to facilitate in-depth research on ecosystem health, biodiversity conservation, and precision agriculture management.

[0049] In this method, a classification table ranging from coarse to fine divisions is first created based on the detection requirements of vegetation types. In practice, vegetation types are classified into three levels, as shown in the table below:

[0050] Users can also classify and subdivide vegetation types in a more detailed and scientific way based on their actual needs and research objectives.

[0051] Training data were selected from regional remote sensing imagery, and a three-level classification label was created based on the above vegetation type classification table to construct a sample dataset.

[0052] This method provides a detailed introduction to semantic segmentation networks from aspects such as network model architecture design, loss function design, and loss weight design.

[0053] Constructing a semantic segmentation network includes: (1) Obtain the multi-level structure of vegetation types.

[0054] In this method, vegetation types are specifically classified when constructing the sample dataset. Specifically, vegetation types are classified into three levels, with N being 3 levels.

[0055] (2) Based on the fusion features, according to the number of vegetation types N, an N-level semantic segmentation model coupled with prior knowledge of vegetation types is constructed using the U-Net model as the basic architecture.

[0056] In this method, the semantic segmentation network is multi-level and utilizes the hierarchical structure of vegetation types. Therefore, it is also called a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types.

[0057] In practice, the number of vegetation type levels, N, is 3. Based on the classification hierarchy of vegetation types, a three-level semantic segmentation model is constructed using the U-Net model as the basic architecture, such as... Figure 4 As shown. The input to the three-level semantic segmentation model is the fused feature obtained at the end of the feature extraction part in step S2 mentioned above.

[0058] Preferably, the N-level semantic segmentation model includes: The previous layer of the N-level semantic segmentation model serves as a constraint condition for the next layer. The output of each level of the N-level semantic segmentation model corresponds to the vegetation type result.

[0059] Specifically, the first layer of the model serves as the constraint condition for the second layer, and the second layer serves as the constraint condition for the third layer; the first layer outputs the first-level classification result of vegetation type, the second layer outputs the second-level classification result of vegetation type, and the third layer outputs the third-level classification result of vegetation type.

[0060] Preferably, the N-level semantic segmentation model further includes: The loss function of the N-level semantic segmentation model is obtained by considering the loss of all levels and weighting and summing the loss of the coarse-grained type level and the loss of the fine-grained type level. During model training, the distribution of loss weights is controlled to overcome the problem of imbalanced samples.

[0061] In practice, utilizing prior structural information about vegetation types can guide the training of the entire model, making gradient backpropagation easier, thereby reducing the number of training iterations and accelerating convergence. The model's loss function considers losses at all levels and is obtained by weighted summation of the losses at the coarse-grained and fine-grained types. Specifically, the loss function L... U The calculation method is as follows:

[0062] Where N is the number of vegetation type levels, N=3; V n For the loss weights, each V n The value of V is between 0 and 1, and the sum is 1. n The contribution of a specific level to the final loss function is defined; G is the number of objects to be classified; Y is the expected output result with dimensions G×h×w; Output the results for the model.

[0063] To overcome the problem of imbalanced samples, the loss weight V is controlled. n The distribution of the model's output layers allows for flexible control over the types of layers that need to be focused on during training. This not only enables more flexible extraction of the layers that need to be focused on during training, but also effectively prevents gradient vanishing and overcomes imbalanced samples.

[0064] Finally, the vegetation type detection results output from each layer of the N-level semantic segmentation model are integrated to obtain the final output of the detection network. Specifically, the final output has a dimension of G×h×w, where G is the number of objects to be classified, h is the number of channels in the output, and each channel corresponds to the probability of whether each position belongs to that category.

[0065] In practical applications, a comparison threshold can be preset based on the output of the detection model. If none of the classification results for a certain pixel location reach the threshold, then that location is an unidentified type or another type. If any classification result reaches the threshold, then the probabilities of each category are compared, and the vegetation type with the highest probability is selected as the actual category of the current pixel location.

[0066] Preferably, the method further includes: The detection results were verified by combining samples collected in the field, and the deep hierarchical semantic segmentation model was fine-tuned and improved.

[0067] The beneficial effects of this embodiment are as follows: This method, based on the attention mechanism, can extract features from spatial, channel, and scale perspectives, enabling the extraction of differentiated features from remote sensing image data across multiple dimensions and improving classification performance. By constructing a deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types, it can overcome the problem of uneven distribution of vegetation type samples, improve training speed, optimize classification performance, and enhance detection accuracy.

[0068] This method, through fine-grained classification of vegetation types, provides more detailed data support for ecologists, forestry managers, and environmental protection agencies, helping to formulate more scientific and rational ecological protection and restoration strategies. In the agricultural sector, fine-grained vegetation type detection can also help farmers achieve precise management of crop planting and improve resource utilization efficiency.

[0069] Example 2 like Figure 2 and 5 As shown, a method for detecting regional vegetation types includes the following steps: S1: Acquire regional remote sensing image data and perform preprocessing.

[0070] In practical applications, the regional remote sensing image data collected are multi-temporal GF-1 WFV, GF-1 PMS and GF-2 PMS images.

[0071] GF-1 WFV, GF-1 PMS, and GF-2 PMS imagery refer to several different types of remote sensing imagery data acquired by Earth observation satellites launched by China. GF stands for the "Gaofen" series of satellites, which are high-resolution satellites capable of acquiring high-resolution images of the Earth's surface for various scientific research and applications. Specifically, "GF-1" and "GF-2" represent the Gaofen-1 and Gaofen-2 satellites, respectively, while "WFV" and "PMS" represent different types of sensors.

[0072] WFV is an abbreviation for "Wide Field of View." This type of sensor can cover a wide area of ​​the ground and is suitable for large-scale environmental monitoring and resource surveys.

[0073] PMS is an abbreviation for "Panchromatic and Multispectral Camera". This type of sensor can acquire panchromatic and multispectral images. Panchromatic images have high resolution, while multispectral images can provide spectral information about ground features, which helps in land cover classification and environmental monitoring.

[0074] In practical applications, preprocessing methods for regional remote sensing image data include: noise removal, geometric correction, radiometric calibration, atmospheric correction, and image registration.

[0075] S2: Based on the preprocessed regional remote sensing image data, hierarchical feature extraction and fusion are performed to obtain fused features.

[0076] Preferably, the feature extraction part of this method achieves hierarchical feature extraction and fusion by constructing a spatiotemporal feature extraction module and a scale attention module. In this method, the feature extraction part is also referred to as the hierarchical scale-sensitive attention spatiotemporal feature extraction model.

[0077] Attention mechanisms in computer vision can extract differentiated features from multiple dimensions. Based on attention mechanisms, features can be extracted from spatial, channel, and scale perspectives, which can enhance the differentiated features of remote sensing images and improve classification performance.

[0078] 1. Spatiotemporal Feature Extraction Module A spatiotemporal feature extraction module based on joint spatial attention and channel attention is constructed to extract spatiotemporal features from preprocessed regional remote sensing image data.

[0079] Specifically, the network structure of the spatiotemporal feature extraction module is as follows: Figure 3 As shown, the input feature is A, and the final output feature is E. The dimensions of A and E are both c×h×w, where c is the number of channels, h is the number of pixels in the image height, and w is the number of pixels in the image width.

[0080] Preferably, the spatiotemporal feature extraction module includes: (1) Spatial attention module, which considers the correlation of different locations in the image and extracts the spatial features of the regional remote sensing image data.

[0081] like Figure 3 As shown, the spatial attention module takes feature A as input and generates three new feature maps B, F, and D through convolution. The dimensions of B, F, and D are the same as those of A and E, which are all c×h×w. B and F are then processed to generate feature S, which is a 2h×w feature map. Finally, after processing, feature P is output, and the dimension of P is the same as that of E, which is also c×h×w. Specifically, the spatial attention model is calculated as follows:

[0082]

[0083] Where i and j represent the positions of the i-th and j-th pixels in the image; α is the scale parameter, initially set to 0; A j P is the input feature at position j; j S is the output feature at position j; ji This is the core computational step of the spatial attention model, which measures the correlation between the i-th position and the j-th position; M is the number of samples.

[0084] (2) Channel attention module, which considers the correlation between different channels in the image and extracts the channel features of the regional remote sensing image data.

[0085] The input image is usually an RGB image with 3 channels. However, remote sensing images are different from RGB images, having multiple channels, also known as bands. High-resolution images typically have 4 bands, i.e., c=4; other sensors, such as Sentinel-2, have 11 bands, i.e., c=11. Remote sensing equipment can capture electromagnetic waves of different wavelength ranges, with each band corresponding to data within a specific wavelength range.

[0086] like Figure 3 As shown, the channel attention module takes feature A as input, performs intermediate operations to generate feature X, and X has a dimension of c×c; finally, after further operations, it outputs feature Q, whose dimension is the same as that of E, also c×h×w. Specifically, the calculation method of the channel attention module is as follows:

[0087]

[0088] Where i and j represent the i-th and j-th bands of the image; β is the weighting parameter, initially set to 0; A j The input feature at position j; Qj X is the output feature at the j-th position; ji The correlation between the i-th band and the j-th band was measured.

[0089] (3) Add the outputs of the spatial attention module and the channel attention module to obtain the final output feature E.

[0090] By integrating the processing results of the spatial attention module and the channel attention module, not only is the spatial consistency of pixels of the same category improved, but the band distinguishability is also enhanced, enabling the extraction of spatiotemporal features from remote sensing input images.

[0091] 2. Scale Attention Module By constructing a scale attention module, the contribution weights of spatiotemporal features at different scales are calculated, and the spatiotemporal features of different image data are fused to obtain fused features.

[0092] Specifically, by constructing a scale-attention model, the contribution weights of spatiotemporal features generated from remote sensing images at different spatial resolutions to the final result are learned. These spatiotemporal features are then weighted and summed to achieve spatiotemporal feature fusion. The weights are calculated as follows:

[0093] Where k∈{1,...,K}, represents the level to which remote sensing images of different spatial resolutions belong. In specific implementation, there are a total of 5 scale levels, K=5, k∈{1,...,5}; w k For weights; h k h t These are the classification probability results at different scale levels.

[0094] S3: Construct a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types to input fused features and output vegetation type classification results.

[0095] Among them, the prior knowledge of vegetation types is the hierarchical structure of vegetation types.

[0096] This method uses a semantic segmentation network to detect vegetation types. Before constructing the semantic segmentation network, the vegetation types need to be classified. Then, the remote sensing image data is labeled to construct a sample dataset.

[0097] In forest ecosystems, traditional vegetation type detection only classifies vegetation into broad categories such as forest, shrub, and grass, which is too general. For example, the single category of "forest" cannot effectively distinguish between different types such as coniferous forests, broad-leaved forests, and mixed forests, which differ significantly in ecological processes, species composition, and responses to environmental changes. Therefore, more granular detection and classification of vegetation types are needed to facilitate in-depth research on ecosystem health, biodiversity conservation, and precision agriculture management.

[0098] In this method, a classification table ranging from coarse to fine divisions is first created based on the detection requirements of vegetation types. In practice, vegetation types are classified into three levels, as shown in the table below:

[0099] Users can also classify and subdivide vegetation types in a more detailed and scientific way based on their actual needs and research objectives.

[0100] Training data were selected from regional remote sensing imagery, and a three-level classification label was created based on the above vegetation type classification table to construct a sample dataset.

[0101] This method provides a detailed introduction to semantic segmentation networks from aspects such as network model architecture design, loss function design, and loss weight design.

[0102] Constructing a semantic segmentation network includes: (1) Obtain the multi-level structure of vegetation types.

[0103] In this method, vegetation types are specifically classified when constructing the sample dataset. Specifically, vegetation types are classified into three levels, with N being 3 levels.

[0104] (2) Based on the fusion features, according to the number of vegetation types N, an N-level semantic segmentation model coupled with prior knowledge of vegetation types is constructed using the U-Net model as the basic architecture.

[0105] In this method, the semantic segmentation network is multi-level and utilizes the hierarchical structure of vegetation types. Therefore, it is also called a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types.

[0106] In practice, the number of vegetation type levels, N, is 3. Based on the classification hierarchy of vegetation types, a three-level semantic segmentation model is constructed using the U-Net model as the basic architecture, such as... Figure 4 As shown. The input to the three-level semantic segmentation model is the fused feature obtained at the end of the feature extraction part in step S2 mentioned above.

[0107] Preferably, the N-level semantic segmentation model includes: The previous layer of the N-level semantic segmentation model serves as a constraint condition for the next layer. The output of each level of the N-level semantic segmentation model corresponds to the vegetation type result.

[0108] Specifically, the first layer of the model serves as the constraint condition for the second layer, and the second layer serves as the constraint condition for the third layer; the first layer outputs the first-level classification result of vegetation type, the second layer outputs the second-level classification result of vegetation type, and the third layer outputs the third-level classification result of vegetation type.

[0109] Preferably, the N-level semantic segmentation model further includes: The loss function of the N-level semantic segmentation model is obtained by considering the loss of all levels and weighting and summing the loss of the coarse-grained type level and the loss of the fine-grained type level. During model training, the distribution of loss weights is controlled to overcome the problem of imbalanced samples.

[0110] In practice, utilizing prior structural information about vegetation types can guide the training of the entire model, making gradient backpropagation easier, thereby reducing the number of training iterations and accelerating convergence. The model's loss function considers losses at all levels and is obtained by weighted summation of the losses at the coarse-grained and fine-grained types. Specifically, the loss function L... U The calculation method is as follows:

[0111] Where N is the number of vegetation type levels, N=3; V n For the loss weights, each V n The value of V is between 0 and 1, and the sum is 1. n The contribution of a specific level to the final loss function is defined; G is the number of objects to be classified; Y is the expected output result with dimensions G×h×w; Output the results for the model.

[0112] To overcome the problem of imbalanced samples, the loss weight V is controlled. n The distribution of the model's output layers allows for flexible control over the types of layers that need to be focused on during training. This not only enables more flexible extraction of the layers that need to be focused on during training, but also effectively prevents gradient vanishing and overcomes imbalanced samples.

[0113] (3) Based on the hierarchical structure of vegetation types, the classification results of vegetation types are cascaded and fused to construct an optimization model and obtain optimized vegetation type detection results.

[0114] This step involves constructing an optimization model. Based on the interrelationships between types at each level, the classification results from each layer of the N-level semantic segmentation model are cascaded and fused to ensure the consistency of the detection results at each level in hierarchical logic. This process constructs the optimized output of the N-level semantic segmentation model, resulting in optimized vegetation type detection results. In this method, this step is also referred to as the classification result optimization method that combines the uncertainties of results from each layer.

[0115] Specifically, assuming It is the probability of hierarchical classification results generated by the N-level semantic segmentation model, by... The cascaded structures are then used as input to construct an optimization model, calculated as follows:

[0116] in, Z represents the final classification result; Z represents the newly constructed optimized model. In order to enable the newly constructed network model to complete synchronous end-to-end network training with the N-level semantic segmentation model, this method uses ResNet as the basic architecture to construct the optimized model.

[0117] The detection network in this method combines an N-level semantic segmentation model with the aforementioned optimization model. The final loss function L consists of two parts: one part is the weighted loss L generated by the N-level semantic segmentation model. U The other part is the loss L of the optimization model. Z The specific calculation method is as follows:

[0118]

[0119] in, The hyperparameters are used to optimize the model, control the classification results of each layer, and optimize the weights of the results obtained by the model.

[0120] Ultimately, the output dimension of the detection network is G×h×w, where G is the number of objects to be classified, h is the number of channels in the output, and each channel corresponds to the probability of whether each position belongs to that category.

[0121] In practical applications, a comparison threshold can be preset based on the output of the detection model. If none of the classification results for a certain pixel location reach the threshold, then that location is an unidentified type or another type. If any classification result reaches the threshold, then the probabilities of each category are compared, and the vegetation type with the highest probability is selected as the actual category of the current pixel location.

[0122] Preferably, the method further includes: The detection results were verified by combining samples collected in the field, and the deep hierarchical semantic segmentation model was fine-tuned and improved.

[0123] S4: Based on the vegetation type detection results, construct a regional vegetation type distribution map.

[0124] The vegetation type detection results of the detection area are visualized to construct a vegetation type distribution map, which helps environmental monitoring and decision analysis and facilitates subsequent analysis and processing.

[0125] The beneficial effects of this embodiment are as follows: This method also constructs an optimization model to cascade and fuse the classification results of each layer of the N-level semantic segmentation model, ensuring the consistency of the detection results of each layer in hierarchical logic and optimizing the output results; it constructs a visualized regional vegetation type distribution map to help environmental monitoring and decision analysis, and facilitates subsequent analysis and processing.

[0126] Example 3 A regional vegetation type detection system, based on the regional vegetation type detection method as described in Embodiment 1 or 2, includes: The data acquisition module is used to acquire regional remote sensing image data and perform preprocessing. The feature extraction module is used to perform hierarchical feature extraction and fusion to obtain fused features based on the preprocessed regional remote sensing image data; The detection module is used to construct a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types to input fused features and output vegetation type classification results.

[0127] Preferably, the system further includes: The model adjustment module is used to verify the detection results by combining field-collected samples, and to fine-tune and improve the regional vegetation type detection model; The distribution map construction module is used to construct a regional vegetation type distribution map based on the vegetation type detection results.

[0128] Example 4

[0129] A computer program, when executed, implements the steps of the regional vegetation type detection method as described in Embodiment 1 or 2.

[0130] Example 5

[0131] A readable storage medium storing a computer program as described in Embodiment 4, wherein when executed, the computer program implements the steps of the regional vegetation type detection method as described in Embodiment 1 or 2.

[0132] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting regional vegetation types, characterized in that, Includes the following steps: Acquire regional remote sensing image data and perform preprocessing; Based on the preprocessed regional remote sensing image data, hierarchical feature extraction and fusion are performed to obtain fused features; A deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types is constructed to input fused features and output vegetation type classification results; Among them, the prior knowledge of vegetation types is the hierarchical structure of vegetation types; The construction of the deep hierarchical semantic segmentation model coupled with prior knowledge of vegetation types includes: Obtain the multi-level structure of vegetation types; Based on the fusion features, and according to the number of vegetation type levels N, an N-level semantic segmentation model coupled with prior knowledge of vegetation types is constructed using the U-Net model as the basic architecture. The output of the previous layer of the N-level semantic segmentation model serves as the constraint condition for the next layer, and each level outputs the corresponding vegetation type result. The loss function of the N-level semantic segmentation model is obtained by considering the loss of all levels and weighting the loss of the coarse-grained type level and the loss of the fine-grained type level. During model training, the distribution of the loss weights is controlled. Based on the hierarchical structure of vegetation types, the classification results of the vegetation types are cascaded and fused to construct an optimized model and obtain optimized vegetation type detection results.

2. The method for detecting regional vegetation types according to claim 1, characterized in that, The step of extracting and fusing hierarchical features from preprocessed regional remote sensing image data to obtain fused features includes: A spatiotemporal feature extraction module is constructed based on joint spatial attention and channel attention to extract spatiotemporal features from preprocessed regional remote sensing image data.

3. The method for detecting regional vegetation types according to claim 2, characterized in that, The step of extracting and fusing hierarchical features from preprocessed regional remote sensing image data to obtain fused features includes: A scale attention module is constructed to calculate the contribution weights of spatiotemporal features at different scales, and the spatiotemporal features of different image data are fused to obtain fused features.

4. The method for detecting regional vegetation types according to claim 1, characterized in that, The method for detecting regional vegetation types also includes: The detection results were verified by combining samples collected in the field, and the deep hierarchical semantic segmentation model was fine-tuned and improved.

5. The method for detecting regional vegetation types according to claim 1, characterized in that, The method for detecting regional vegetation types also includes: Based on the vegetation type detection results, a regional vegetation type distribution map is constructed.

6. A regional vegetation type detection system, characterized in that, The regional vegetation type detection method according to any one of claims 1 to 5 includes: The data acquisition module is used to acquire regional remote sensing image data and perform preprocessing. The feature extraction module is used to perform hierarchical feature extraction and fusion to obtain fused features based on the preprocessed regional remote sensing image data; The detection module is used to construct a deep hierarchical semantic segmentation model that couples prior knowledge of vegetation types to input fused features and output vegetation type classification results.

Images