A method for fine classification of typical urban land features applicable to high-resolution hyperspectral data

Abstract

This invention relates to a method for fine classification of typical urban features using high-resolution hyperspectral data. The method includes: acquiring raw hyperspectral data from multiple flight strips of the original region; preprocessing all raw hyperspectral data to obtain multiple training sample data containing at least one typical urban feature; training the original hyperspectral classification model based on the multiple training sample data to obtain a target hyperspectral classification model; acquiring target hyperspectral data of the target region; and obtaining the classification result of typical urban features based on the target hyperspectral classification model. This invention increases the amount of sample data and designs a convolutional autoencoder network to construct and train the hyperspectral classification model. Utilizing the trained semantic segmentation model, it achieves fine classification of typical features while also providing technical support for the expanded application of novel hyperspectral payloads in urban remote sensing.

Description

Technical Field

[0001] This invention relates to the fields of hyperspectral technology, artificial intelligence, and urban remote sensing applications, and in particular to a method for fine classification of typical urban land features applicable to high-resolution hyperspectral data. Background Technology

[0002] Hyperspectral technology has been a hot topic in recent years, attracting widespread attention due to its spatial-spectral integration, which gives it a natural advantage in remote sensing Earth observation. With the continuous advancement of hyperspectral payload development technology in my country, domestic satellite hyperspectral payloads such as Gaofen-5 and Ziyuan-1-02D ​​have been launched, and in recent years, airborne hyperspectral payloads with sub-meter spatial resolution have been developed. Traditional hyperspectral data, due to insufficient spatial resolution, is often used primarily for large-scale natural elements such as forests, water bodies, and crops. Improved spatial resolution allows hyperspectral data to play a crucial role in urban remote sensing. However, because hyperspectral data comprises hundreds of bands, the increased spatial resolution leads to a massive increase in data volume, posing new challenges for its subsequent applications.

[0003] Traditional methods for classifying typical land cover based on hyperspectral data can be summarized into three categories: image enhancement methods, spectral feature recognition methods, and traditional machine learning methods. Image enhancement methods, such as Minimum Noise Separation (MNF), Principal Component Analysis (PCA), and band ratio methods, are simple and efficient at extracting target features, but cannot directly complete classification. Spectral feature recognition methods are based on spectral feature extraction and spectral matching methods. These methods require expert knowledge guidance and have limited generalization ability in complex scenarios. Traditional machine learning methods, such as Support Vector Machines and Random Forests, have better generalization ability than the first two methods, but because they do not fully consider the spatial characteristics of the data, they often suffer from significant "salt and pepper effect."

[0004] The rapid development of deep learning technology in remote sensing has driven the application of convolutional neural networks in land cover classification. However, current classification models, primarily based on purely supervised learning, have high sample requirements, and due to the scarcity of hyperspectral data, most applications involve small sample classification. Therefore, designing deep learning models for high-resolution hyperspectral data from the perspectives of sample augmentation and training strategy optimization, and applying these models to the classification of typical urban land covers, remains a pressing challenge. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for fine classification of typical urban land features applicable to high-resolution hyperspectral data.

[0006] The technical solution of the present invention for a fine classification method of typical urban land features applicable to high-resolution hyperspectral data is as follows:

[0007] S1. Obtain raw hyperspectral data of multiple flight strips in the original area, and preprocess all raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature.

[0008] S2. Train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model;

[0009] S3. Obtain the target hyperspectral data of the target area, and based on the target hyperspectral classification model, obtain the urban typical land cover classification results of the target hyperspectral data.

[0010] The beneficial effects of the present invention's method for fine classification of typical urban land features applicable to high-resolution hyperspectral data are as follows:

[0011] The method of this invention increases the amount of sample data and designs a convolutional autoencoder network to construct and train a hyperspectral classification model. By using the trained semantic segmentation model, it can achieve fine classification of typical land features and provide technical support for the expansion of new hyperspectral payloads in the field of urban remote sensing.

[0012] Based on the above scheme, the method for fine classification of typical urban land features applicable to high-resolution hyperspectral data of the present invention can be further improved as follows.

[0013] Furthermore, S1 includes:

[0014] S11. Perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data in sequence to obtain mosaicked hyperspectral data.

[0015] S12. Perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data.

[0016] S13. Construct multiple band hyperspectral data based on the reflectance data and the transformed hyperspectral data;

[0017] S14. Slice the hyperspectral data for each band to obtain and enhance all sliced ​​hyperspectral data to obtain multiple training sample data.

[0018] Furthermore, S13 includes:

[0019] S131. Obtain a preset number of band hyperspectral data from the transformed hyperspectral data;

[0020] S132. Obtain the true color band data and classification true value label data from the reflectance data;

[0021] S133. Construct the multiple band hyperspectral data based on the preset number of band hyperspectral data, the true color band data, and the classification truth label data.

[0022] Furthermore, the original hyperspectral classification model includes: an original convolutional autoencoder network and an original semantic segmentation network, then S2 includes:

[0023] S21. The original convolutional autoencoder network is iteratively pre-trained using the multiple training sample data to obtain the target convolutional autoencoder network.

[0024] S22. The original semantic segmentation network is iteratively trained using the multiple training sample data to obtain the target semantic segmentation network;

[0025] S23. Based on the target convolutional autoencoder network and the target semantic segmentation network, the target hyperspectral classification model is obtained.

[0026] Furthermore, the original convolutional autoencoder network includes an original encoder and an original decoder. Then, S21 includes: using each training sample data as the input feature of the original encoder and the output feature of the original decoder for iterative pre-training, and using the root mean square error as the loss function until the loss function decreases and tends to converge, to obtain the target convolutional autoencoder network.

[0027] Furthermore, the original semantic segmentation network includes: an original semantic segmentation decoder and a target encoder of the target convolutional autoencoder network. Then, S22 includes: locking the parameters of the target encoder and iteratively training the original semantic segmentation decoder using the multiple training sample data to obtain the target semantic segmentation network.

[0028] Furthermore, it also includes: using the validation sample data of the original region, correcting the target hyperspectral classification model based on the validation sample data, and determining the corrected target hyperspectral classification model as the target hyperspectral classification model.

[0029] The technical solution of a typical land cover classification system of the present invention is as follows:

[0030] It includes: a processing module, a training module, and a runtime module;

[0031] The processing module is used to: acquire raw hyperspectral data of multiple flight strips in the original area, and preprocess all the raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature;

[0032] The training module is used to: train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model;

[0033] The operating module is used to: acquire target hyperspectral data of the target area, and obtain the urban typical land cover classification results of the target hyperspectral data based on the target hyperspectral classification model.

[0034] The beneficial effects of a typical land cover classification system of the present invention are as follows:

[0035] The system of this invention increases the amount of sample data and designs a convolutional autoencoder network to construct and train a hyperspectral classification model. By using the trained semantic segmentation model, it can achieve fine classification of typical land features and provide technical support for the expansion of new hyperspectral payloads in the field of urban remote sensing.

[0036] Based on the above scheme, a typical land cover classification system of the present invention can be further improved as follows.

[0037] Furthermore, the processing module includes: a first processing module, a second processing module, a third processing module, and a fourth processing module;

[0038] The first processing module is used to: sequentially perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data to obtain mosaicked hyperspectral data;

[0039] The second processing module is used to: perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data;

[0040] The third processing module is used to: construct multiple band hyperspectral data based on the reflectance data and the transformed hyperspectral data;

[0041] The fourth processing module is used to: slice the hyperspectral data of each band, obtain and enhance all the sliced ​​hyperspectral data to obtain multiple training sample data.

[0042] Furthermore, the third processing module is specifically used for:

[0043] From the transformed hyperspectral data, a preset number of band hyperspectral data are acquired;

[0044] Obtain the true color band data and classification ground truth label data from the reflectance data;

[0045] The multiple band hyperspectral data are constructed based on the preset number of band hyperspectral data, the true color band data, and the classification true value label data. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating a method for fine classification of typical urban land features using high-resolution hyperspectral data, according to an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the data slicing process in a method for fine classification of typical urban land features using high-resolution hyperspectral data, according to an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the data augmentation process in a method for fine classification of typical urban land features using high-resolution hyperspectral data, according to an embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram of the RCB convolutional unit of the original encoder in a fine classification method for typical urban land features using high-resolution hyperspectral data according to an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of the structure of a target hyperspectral classification model in a fine classification method for typical urban land features applicable to high-resolution hyperspectral data according to an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the original semantic segmentation decoder in a fine classification method for typical urban land features using high-resolution hyperspectral data according to an embodiment of the present invention.

[0052] Figure 7 This is a schematic diagram of the target hyperspectral data of the target area in a fine classification method for typical urban land features using high-resolution hyperspectral data according to an embodiment of the present invention.

[0053] Figure 8 This is a schematic diagram of the classification results of typical urban features in a target area in a fine classification method for typical urban features using high-resolution hyperspectral data according to an embodiment of the present invention.

[0054] Figure 9 This is a schematic diagram of the structure of a fine classification system for typical urban land features applicable to high-resolution hyperspectral data, according to an embodiment of the present invention. Detailed Implementation

[0055] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for fine classification of typical urban land features applicable to high-resolution hyperspectral data, comprising the following steps:

[0056] S1. Obtain raw hyperspectral data of multiple flight strips in the original area, and preprocess all raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature.

[0057] The original region is a specific urban area. The original hyperspectral data consists of hyperspectral data (hyperspectral images) acquired via an airborne hyperspectral sensor during flight, using multiple flight bands. Since hyperspectral data comprises hundreds of bands, and the sensor is airborne with flight bands, the acquired data is also segmented by flight band. There is some overlap between flight band data, and the data acquired for each flight band constitutes the original hyperspectral data. The training sample data consists of hyperspectral data obtained after image (video) preprocessing. Each training sample includes at least one type of typical urban feature. Typical urban features include, but are not limited to, roads, crops, buildings, and water bodies.

[0058] S2. Train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model.

[0059] The original hyperspectral classification model is an untrained network model consisting of a convolutional autoencoder network and a semantic segmentation network. The target hyperspectral classification model is a network model obtained by training the original hyperspectral classification model with multiple training sample data, which can be used to classify typical land cover types in hyperspectral data of a certain region.

[0060] S3. Obtain the target hyperspectral data of the target area, and based on the target hyperspectral classification model, obtain the urban typical land cover classification results of the target hyperspectral data.

[0061] The acquisition method for the target hyperspectral data is the same as that for the original hyperspectral data of multiple flight strips in the original region. The target region is the area where typical urban land cover classification and identification is required. This region can be the same as or different from the original region; no restrictions are set here.

[0062] Preferably, S1 includes:

[0063] S11. Perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data in sequence to obtain mosaicked hyperspectral data.

[0064] The process of correcting the radiometric uniformity of the original hyperspectral data refers to the technical solution for correcting hyperspectral data in "A Photometric Correction Method for Airborne Hyperspectral Remote Sensing Data". The purpose is to use the functional relationship between the sensor observation angle and the radiance value to eliminate the radiometric non-uniformity between multiple flight strip images of airborne hyperspectral data caused by the dihedrality of the ground surface.

[0065] The process of image mosaicking of the original hyperspectral data after radiation homogeneity correction is an existing technology.

[0066] S12. Perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data.

[0067] Atmospheric correction is an existing technique and will not be elaborated upon here. The reflectance data refers to the image data after atmospheric correction. The process of filtering the mosaic hyperspectral data includes removing bad bands such as those affected by water vapor from the mosaic reflectance data. The PCA transformation process for each band is an existing technique; its purpose is to perform feature filtering on hyperspectral data with hundreds of bands, minimizing redundant information and retaining effective classification information. The transformed hyperspectral data refers to the hyperspectral data after PCA transformation.

[0068] S13. Based on the reflectance data and the transformed hyperspectral data, construct multiple band hyperspectral data.

[0069] Wherein, S13 includes:

[0070] S131. Obtain a preset number of band hyperspectral data from the transformed hyperspectral data.

[0071] S132. Obtain the true color band data and classification true value label data from the reflectance data.

[0072] S133. Construct the multiple band hyperspectral data based on the preset number of band hyperspectral data, the true color band data, and the classification truth label data.

[0073] Specifically, for example, the first N bands after PCA transformation are extracted and combined with the red, green, and blue true-color bands from the reflectance data and the classification ground truth labels to form hyperspectral data with N+4 bands. Here, N is a positive integer, and in this embodiment, N is taken as 1 / 10 of the total number of bands.

[0074] S14. Slice the hyperspectral data for each band to obtain and enhance all sliced ​​hyperspectral data to obtain multiple training sample data.

[0075] The process of slicing the hyperspectral data for each band is as follows: Figure 2As shown, the hyperspectral data is randomly cropped to form several 256×256×(N+4) hyperspectral data cubes. The specific process is as follows: a random point coordinate (a, b) is set in the image. A 256×256 cropping box is generated using this point as the vertex, and the data is cropped using this box. This process is repeated n times to complete the slicing operation, obtaining all the sliced ​​hyperspectral data. The random point coordinates a and b range from 0 to w-5 and 0 to h-5, respectively, where w and h represent the length and width of the sliced ​​hyperspectral data.

[0076] The process of enhancing all slice hyperspectral data is as follows: the resulting n slices are used as dataset DATA-A, such as... Figure 3 As shown, random geometric augmentation and color enhancement are performed on the dataset DATA-A. Geometric augmentation includes operations such as flipping, rotating, scaling, and offsetting, while color enhancement includes changing the saturation, brightness, contrast, and sharpness of the data. The above operations are performed on all slices of the dataset DATA-A to expand the dataset DATA-A to 6 times its original size, forming a new dataset DATA-B (multiple training sample data).

[0077] Preferably, the original hyperspectral classification model includes: an original convolutional autoencoder network and an original semantic segmentation network, then S2 includes:

[0078] The CAE network was built in the TensorFlow environment. The original hyperspectral classification model consisted of an original convolutional autoencoder network and an original semantic segmentation network. The original convolutional autoencoder network (CAE) was composed of an original encoder and an original decoder with a mirror structure. The original encoder consisted of five layers of RCB convolutional units (e.g., ...). Figure 4 As shown in Table 1, each unit consists of a 3D depth-separable convolutional layer, a normalization layer, and a ReLU activation layer. The structural parameters of the original convolutional autoencoder network are shown in Table 1.

[0079] Table 1:

[0080]

[0081] S21. The original convolutional autoencoder network is iteratively pre-trained using the multiple training sample data to obtain the target convolutional autoencoder network.

[0082] S21 includes: performing iterative pre-training with each training sample data as the input feature of the original encoder and the output feature of the original decoder, and using the root mean square error as the loss function until the loss function decreases and tends to converge, thereby obtaining the target convolutional autoencoder network.

[0083] Specifically, the image data from the constructed training sample datasets are used as the input and output of the network for iterative pre-training. During training, the root mean square error (RMSE) is used as the loss function to evaluate the network's reconstruction accuracy. The model training is considered complete when the reconstruction error decreases and stabilizes. During pre-training, the high-dimensional hyperspectral data is dimensionality-reduced while preserving semantic features for subsequent classification tasks.

[0084] S22. The original semantic segmentation network is iteratively trained using the multiple training sample data to obtain the target semantic segmentation network.

[0085] Wherein, the original semantic segmentation network includes: an original semantic segmentation decoder and a target encoder of the target convolutional autoencoder network, then S22 includes: locking the parameters of the target encoder and iteratively training the original semantic segmentation decoder using the multiple training sample data to obtain the target semantic segmentation network.

[0086] S23. Based on the target convolutional autoencoder network and the target semantic segmentation network, the target hyperspectral classification model is obtained.

[0087] Specifically, a raw semantic segmentation network is constructed in the TensorFlow environment for the classification of typical land features in hyperspectral data. This network mainly consists of two parts: a CAE network target encoder and a raw semantic segmentation decoder (e.g., ...). Figure 5 The original semantic segmentation decoder consists of two upsampling layers, two CBR convolutional units, and an attention module, as shown in the structure below. Figure 6 As shown, the training process for the original semantic segmentation decoder is as follows: the parameters of the target encoder are locked, only the original semantic segmentation decoder is trained, and the target semantic segmentation network is obtained through multiple rounds of iterative training using all training sample data. The loss function used during training is cross-entropy, and the optimizer is adam.

[0088] Preferably, the method further includes: using the validation sample data of the original region, modifying the target hyperspectral classification model based on the validation sample data, and determining the modified target hyperspectral classification model as the target hyperspectral classification model.

[0089] Specifically, the prediction accuracy of the target hyperspectral classification model is evaluated using validation sample data. Evaluation metrics include precision, recall, F1 score, and overall classification accuracy. For example, the target hyperspectral classification model is applied to... Figure 7 From the target flight strip data of the target area in the middle, we obtain Figure 8 The typical land cover classification results for the target area are shown in Table 2. The classification accuracy of the typical land cover in the target area is shown in Table 2.

[0090] Table 2:

[0091]

[0092] Among them, 1-city road; 2-village road; 3-greenhouse; 4-building; 5-non-major crops; 6-pond; 7-woodland; 8-rice; 9-water body; 10-corn; 11-grassland; 12-bare land; 13-other.

[0093] The technical solution of this embodiment adopts the method of increasing the amount of sample data and designs a convolutional autoencoder network to construct and train the hyperspectral classification model. By using the trained semantic segmentation model, it can achieve fine classification of typical land features, and also provides technical support for the expansion of new hyperspectral payloads in the field of urban remote sensing.

[0094] like Figure 9 As shown, a typical land cover classification system 200 according to an embodiment of the present invention includes: a processing module 210, a training module 220 and a running module 230;

[0095] The processing module 210 is used to: acquire raw hyperspectral data of multiple flight strips in the original area, and preprocess all the raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature;

[0096] The training module 220 is used to: train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model;

[0097] The operation module 230 is used to: acquire target hyperspectral data of the target area, and obtain the urban typical land cover classification results of the target hyperspectral data based on the target hyperspectral classification model.

[0098] Preferably, the processing module 210 includes: a first processing module, a second processing module, a third processing module, and a fourth processing module;

[0099] The first processing module is used to: sequentially perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data to obtain mosaicked hyperspectral data;

[0100] The second processing module is used to: perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data;

[0101] The third processing module is used to: construct multiple band hyperspectral data based on the reflectance data and the transformed hyperspectral data;

[0102] The fourth processing module is used to: slice the hyperspectral data of each band, obtain and enhance all the sliced ​​hyperspectral data to obtain multiple training sample data.

[0103] Preferably, the third processing module is specifically used for:

[0104] From the transformed hyperspectral data, a preset number of band hyperspectral data are acquired;

[0105] Obtain the true color band data and classification ground truth label data from the reflectance data;

[0106] The multiple band hyperspectral data are constructed based on the preset number of band hyperspectral data, the true color band data, and the classification true value label data.

[0107] The technical solution of this embodiment adopts the method of increasing the amount of sample data and designs a convolutional autoencoder network to construct and train the hyperspectral classification model. By using the trained semantic segmentation model, it can achieve fine classification of typical land features, and also provides technical support for the expansion of new hyperspectral payloads in the field of urban remote sensing.

[0108] The parameters and steps for implementing the corresponding functions of each module in the typical land cover classification system 200 of the present invention described above can be referred to the parameters and steps in the embodiment of a fine classification method for typical urban land covers applicable to high-resolution hyperspectral data, which will not be repeated here.

[0109] An embodiment of the present invention provides a storage medium, comprising: the storage medium storing instructions, which, when a computer reads the instructions, cause the computer to execute the steps of a fine classification method for typical urban features applicable to high-resolution hyperspectral data as described above. For details, please refer to the parameters and steps in the embodiment of the fine classification method for typical urban features applicable to high-resolution hyperspectral data described above, which will not be repeated here.

[0110] Computer storage media, such as USB flash drives and external hard drives.

[0111] An electronic device provided by an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor, when executing the computer program, causes the computer to perform the steps of a fine classification method for typical urban features applicable to high-resolution hyperspectral data as described above. Specific details regarding the parameters and steps in the embodiment of the fine classification method for typical urban features applicable to high-resolution hyperspectral data described above are provided but will not be repeated here.

[0112] Those skilled in the art will know that the present invention can be implemented as a method, system, storage medium, and electronic device.

[0113] Therefore, the present invention can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the present invention can also be implemented as a computer program product contained in one or more computer-readable media, which contains computer-readable program code. Any combination of one or more computer-readable media can be used. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. Although embodiments of the invention have been shown and described above, it is to be understood that these embodiments are exemplary and should not be construed as limiting the invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the invention.

Claims

1. A method for fine classification of typical urban land features applicable to high-resolution hyperspectral data, characterized in that, include: S1. Obtain raw hyperspectral data of multiple flight strips in the original area, and preprocess all raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature. S2. Train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model; S3. Obtain the target hyperspectral data of the target area, and based on the target hyperspectral classification model, obtain the urban typical land cover classification results of the target hyperspectral data; S1 includes: S11. Perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data in sequence to obtain mosaicked hyperspectral data. S12. Perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data. S13. Construct multiple band hyperspectral data based on the reflectance data and the transformed hyperspectral data; S14. Slice the hyperspectral data for each band to obtain and enhance all sliced ​​hyperspectral data to obtain multiple training sample data. S13 includes: S131. Obtain a preset number of band hyperspectral data from the transformed hyperspectral data; S132. Obtain the true color band data and classification true value label data from the reflectance data; S133. Based on the preset number of band hyperspectral data, the true color band data, and the classification truth label data, construct the multiple band hyperspectral data; wherein, extract the first N bands after PCA transformation, and combine them with the red, green, and blue true color bands from the reflectance data and the classification truth label to form N+4 band hyperspectral data; N is a positive integer, which is 1 / 10 of the total number of bands.

2. The method for fine classification of typical urban land features applicable to high-resolution hyperspectral data as described in claim 1, characterized in that, The original hyperspectral classification model includes: an original convolutional autoencoder network and an original semantic segmentation network, then S2 includes: S21. The original convolutional autoencoder network is iteratively pre-trained using the multiple training sample data to obtain the target convolutional autoencoder network. S22. The original semantic segmentation network is iteratively trained using the multiple training sample data to obtain the target semantic segmentation network; S23. Based on the target convolutional autoencoder network and the target semantic segmentation network, the target hyperspectral classification model is obtained.

3. The method for fine classification of typical urban land features applicable to high-resolution hyperspectral data as described in claim 2, characterized in that, The original convolutional autoencoder network includes an original encoder and an original decoder. S21 includes: using each training sample data as the input feature of the original encoder and the output feature of the original decoder for iterative pre-training, and using the root mean square error as the loss function until the loss function decreases and tends to converge, to obtain the target convolutional autoencoder network.

4. The method for fine classification of typical urban land features applicable to high-resolution hyperspectral data as described in claim 3, characterized in that, The original semantic segmentation network includes an original semantic segmentation decoder and a target encoder of the target convolutional autoencoder network. S22 includes locking the parameters of the target encoder and iteratively training the original semantic segmentation decoder using the multiple training sample data to obtain the target semantic segmentation network.

5. A method for fine classification of typical urban land features applicable to high-resolution hyperspectral data according to any one of claims 1-4, characterized in that, Also includes: Using the validation sample data of the original region, the target hyperspectral classification model is corrected based on the validation sample data, and the corrected target hyperspectral classification model is determined as the target hyperspectral classification model.

6. A fine classification system for typical urban land features suitable for high-resolution hyperspectral data, characterized in that, include: Processing module, training module, and execution module; The processing module is used to: acquire raw hyperspectral data of multiple flight strips in the original area, and preprocess all the raw hyperspectral data to obtain multiple training sample data containing at least one typical urban land feature; The training module is used to: train the original hyperspectral classification model based on the multiple training sample data to obtain the target hyperspectral classification model; The operating module is used to: acquire target hyperspectral data of the target area, and based on the target hyperspectral classification model, obtain the urban typical land cover classification results of the target hyperspectral data; The processing module includes: a first processing module, a second processing module, a third processing module, and a fourth processing module; The first processing module is used to: sequentially perform radiometric homogeneity correction and image mosaicking on all the original hyperspectral data to obtain mosaicked hyperspectral data; The second processing module is used to: perform atmospheric correction processing on the mosaic hyperspectral data to obtain reflectance data, and perform PCA transformation on each band in the filtered reflectance data to obtain transformed hyperspectral data; The third processing module is used to: construct multiple band hyperspectral data based on the reflectance data and the transformed hyperspectral data; The fourth processing module is used to: slice the hyperspectral data of each band to obtain and enhance all the sliced ​​hyperspectral data to obtain multiple training sample data; The third processing module is specifically used for: From the transformed hyperspectral data, a preset number of band hyperspectral data are acquired; Obtain the true color band data and classification ground truth label data from the reflectance data; Based on the preset number of band hyperspectral data, the true-color band data, and the classification truth label data, the multiple band hyperspectral data are constructed; wherein, the first N bands after PCA transformation are extracted and combined with the red, green, and blue true-color bands from the reflectance data and the classification truth label to form N+4 band hyperspectral data; N is a positive integer, which is 1 / 10 of the total number of bands.

Images