Data processing method and device, and urban ground object classification method and device

Abstract

Embodiments of the present specification provide a data processing method and device, a city ground object classification method and device, wherein the data processing method comprises: collecting hyperspectral regional data for a target region, and extracting target region features in the hyperspectral regional data; inputting the target region features into an object classification model to classify regional objects in the target region, and obtaining object classification information of the regional objects in the target region, wherein the object classification model optimizes the identification accuracy of the object classification model for mixed pixel regions through a hierarchical sensitive spectral loss function; comparing the hyperspectral regional data with historical hyperspectral regional data of the target region, and obtaining regional change information of the target region; and constructing a visual regional report corresponding to the target region based on the object classification information and the regional change information.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of data processing, in particular to a data processing method and device, and a city feature classification method and device. BACKGROUND

[0002] City feature classification and dynamic monitoring is a core task in city planning and management, aiming to accurately obtain the spatial distribution and change information of city features, and to provide support for land use planning, green space supervision, and building distribution optimization. However, traditional monitoring methods, such as satellite remote sensing and manual verification, are inefficient in handling large-scale and complex data. In high-density urban environments with frequent changes in green coverage, the monitoring accuracy is insufficient, and is easily affected by environmental factors such as light and weather, which cannot meet the dynamic monitoring needs of rapid urban development. The use of unmanned aerial vehicles (UAVs) for feature monitoring has become an alternative solution. Its flexible deployment, efficient collection, and low cost make it particularly suitable for urban environment monitoring, especially when equipped with hyperspectral sensors, which can provide more accurate feature information. However, in the prior art, although the UAV can achieve feature classification by collecting spectral data with a hyperspectral sensor, there are still certain errors in feature classification in complex environments due to the inability of the algorithm to meet the accuracy requirements. Therefore, an effective solution is needed to solve the above problems. SUMMARY

[0003] In view of the above, the embodiments of the present specification provide a data processing method. One or more embodiments of the present specification also relate to a city feature classification method, a data processing device, a city feature classification device, a computing device, a computer-readable storage medium, and a computer program product, to solve the technical defects in the prior art.

[0004] According to a first aspect of the embodiments of the present specification, a data processing method is provided, comprising:

[0005] Collecting hyperspectral regional data for a target region, and extracting target region features from the hyperspectral regional data;

[0006] Inputting the target region features into an object classification model to classify regional objects in the target region, and obtaining object classification information of the regional objects in the target region, wherein the object classification model optimizes the recognition accuracy of the object classification model for mixed pixel regions through a hierarchical sensitive spectral loss function;

[0007] Comparing the hyperspectral regional data with historical hyperspectral regional data of the target region, and obtaining regional change information of the target region;

[0008] construct a visual region report corresponding to the target region based on the object classification information and the region change information.

[0009] According to a second aspect of the embodiments of the present specification, a city ground object classification method is provided, including:

[0010] Collecting hyperspectral region data for a city region, and extracting city region features from the hyperspectral region data;

[0011] Inputting the city region features into a classification model to classify ground object objects in the city region, and obtaining ground object classification information of the ground object objects in the city region, wherein the classification model optimizes the recognition accuracy of the classification model for mixed pixel regions through a hierarchical sensitive spectrum loss function;

[0012] Comparing the hyperspectral region data with historical hyperspectral region data of the city region, and obtaining region change information of the city region;

[0013] Constructing a city region change report corresponding to the city region based on the ground object classification information and the region change information.

[0014] According to a third aspect of the embodiments of the present specification, a data processing apparatus is provided, including:

[0015] A collection module configured to collect hyperspectral region data for a target region, and extract target region features from the hyperspectral region data;

[0016] An input module configured to input the target region features into an object classification model to classify region objects in the target region, and obtain object classification information of the region objects in the target region, wherein the object classification model optimizes the recognition accuracy of the object classification model for mixed pixel regions through a hierarchical sensitive spectrum loss function;

[0017] A comparison module configured to compare the hyperspectral region data with historical hyperspectral region data of the target region, and obtain region change information of the target region;

[0018] A construction module configured to construct a visual region report corresponding to the target region based on the object classification information and the region change information.

[0019] According to a fourth aspect of the embodiments of the present specification, a city ground object classification apparatus is provided, including:

[0020] A collection data module configured to collect hyperspectral region data for a city region, and extract city region features from the hyperspectral region data;

[0021] The input model module is configured to input the urban area feature into a classification model to classify objects in the urban area, and obtain ground object classification information of the objects in the urban area, wherein the classification model is optimized by a hierarchical sensitive spectral loss function to improve the identification accuracy of the classification model for mixed pixel regions.

[0022] The comparison data module is configured to compare the hyperspectral region data with historical hyperspectral region data of the urban area, and obtain region change information of the urban area.

[0023] The report construction module is configured to construct an urban region change report corresponding to the urban area based on the ground object classification information and the region change information.

[0024] According to a fifth aspect of the embodiments of the present specification, a computing device is provided, comprising:

[0025] a memory and a processor;

[0026] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, and the computer executable instructions, when executed by the processor, implement the steps of the above data processing method or urban ground object classification method.

[0027] According to a sixth aspect of the embodiments of the present specification, a computer readable storage medium is provided, which stores computer executable instructions, and the instructions, when executed by a processor, implement the steps of the above data processing method or urban ground object classification method.

[0028] According to a seventh aspect of the embodiments of the present specification, a computer program product is provided, comprising a computer program or instructions, and the computer program or instructions, when executed by a processor, implement the steps of the above data processing method or urban ground object classification method.

[0029] The data processing method provided in this embodiment, in order to improve the accuracy of ground feature identification and support efficient and accurate regional planning, first collects hyperspectral regional data for the target area and extracts target area features from the hyperspectral regional data. Based on this, to improve object classification efficiency, the target area features can be input into an object classification model to classify regional objects in the target area, thereby obtaining object classification information for regional objects in the target area based on the classification results. To enable the classification model to have strong object boundary recognition capabilities, during the training phase, the recognition accuracy of the object classification model for mixed pixel areas can be optimized through a hierarchical sensitive spectral loss function. This allows the optimized object classification model to maintain strong boundary recognition accuracy even with data collected in complex environments, thus ensuring that the object classification information of regional objects in the target area output by the model is more accurate. Subsequently, to facilitate use in regional planning and management, the hyperspectral regional data can be compared with historical hyperspectral regional data of the target area to obtain regional change information of the target area. Then, a visualized regional report corresponding to the target area can be constructed based on the object classification information and regional change information. This effectively improves accuracy and efficiency in ground feature analysis and processing, thereby facilitating downstream business use for regional planning and management. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a data processing method provided in one embodiment of this specification;

[0031] Figure 2 This is a flowchart illustrating an urban feature classification method provided in one embodiment of this specification;

[0032] Figure 3 This is a flowchart illustrating the processing procedure of a data processing method provided in one embodiment of this specification.

[0033] Figure 4 This is a schematic diagram of the structure of a data processing apparatus provided in one embodiment of this specification;

[0034] Figure 5 This is a schematic diagram of the structure of an urban land cover classification device provided in one embodiment of this specification;

[0035] Figure 6 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0036] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. However, the present description can be practiced without the specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the present description.

[0037] The terminology used in this description is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present description. As used in this description and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0038] It will be understood that, although the terms first, second, etc. can be used herein to describe various information, these terms are not intended to denote a temporal or chronological order. Rather, these terms are used solely to distinguish one from another only. For example, a first can be termed a second, and, similarly, a second can be also termed a first, without departing from the scope of one or more embodiments of the present description. As used herein, the term "if' can be construed to mean "when" or "in response to determining" or "in response to a determination" depending on the context.

[0039] In addition, it should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present description are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

[0040] First, the terms involved in one or more embodiments of the present description are explained.

[0041] Convolutional Neural Network (CNN) is a deep learning algorithm or model, which is particularly suitable for processing data with grid topology structure, such as image data (which can be regarded as a two-dimensional pixel grid).

[0042] Hyperspectral regional data (also known as imaging spectral data) is a multi-dimensional information set, which adds a spectral dimension to the traditional two-dimensional image data (spatial dimension and spectral dimension), forming a three-dimensional cubic data structure.

[0043] Urban feature classification refers to categorizing various natural and man-made objects on the urban surface according to their formation process, properties, uses, size, and other characteristics. This classification helps to better understand the structure and characteristics of urban geospatial space and further infer the urban environment and human activities.

[0044] A Gaussian filter is a low-pass filter that is widely used in image processing and other fields. It can smooth images, reduce noise, and preserve the edge and detail information of the image.

[0045] Z-score standardization is a common data preprocessing technique that transforms raw data into deviations from the mean expressed in standard deviations. The specific formula is: Z = (X - μ) / σ, where X is the original data points, μ is the mean of the dataset, and σ is the standard deviation of the dataset. Standardized data has a mean of 0 and a variance of 1, allowing data of different magnitudes to be compared and analyzed on the same scale. This method plays a crucial role in data analysis and statistical modeling, eliminating the influence of different data units and magnitudes, and improving the accuracy and reliability of data analysis and statistical modeling. Furthermore, Z-score standardization helps identify outliers in the dataset, providing important data cleaning and preprocessing data.

[0046] Euclidean distance, also known as Euclidean distance, is the "ordinary" (i.e., straight-line) distance between two points in Euclidean space. In two-dimensional space, the formula for Euclidean distance is d = sqrt((x1-x2)^2+(y1-y2)^2); in three-dimensional space, the formula is d=sqrt((x1-x2)^2+(y1-y2)^2+(z1-z2)^2). Euclidean distance is a commonly used definition of distance that reflects the true distance between two points or the natural length of a vector. In data analysis, machine learning, and other fields, Euclidean distance is often used to calculate the distance between sample points for tasks such as clustering and classification.

[0047] Spatial interpolation is a geospatial data analysis technique that explores and analyzes the inherent patterns in collected sample point data, then extrapolates these patterns to the entire study area, transforming it into area data. Essentially, spatial interpolation uses known point data, along with specific mathematical models and algorithms, to scientifically predict data for unknown points. This method has wide applications in Geographic Information Systems (GIS), environmental science, meteorology, geology, and other fields. Various spatial interpolation methods exist, including nearest neighbor interpolation, linear interpolation, bilinear interpolation, spline interpolation, and Kriging interpolation. Different interpolation methods are suitable for different application scenarios and data characteristics; choosing an appropriate interpolation method can improve the accuracy and reliability of data prediction.

[0048] In the present specification, a data processing method is provided. One or more embodiments of the present specification also relate to a city feature classification method, a data processing device, a city feature classification device, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0049] In practical applications, existing hyperspectral regional data classification algorithms still face many challenges when processing city feature monitoring tasks, such as the decline in classification accuracy caused by mixed pixel effects, and weak recognition ability for small-scale features such as temporary buildings and narrow green spaces. In addition, the existing technology lacks efficient classification and change detection mechanisms that adapt to urban environments. Therefore, how to combine unmanned aerial vehicle spectral data acquisition with deep learning algorithms to improve the accuracy of feature classification and dynamic monitoring has become the key to solving the technical problems of city feature monitoring.

[0050] The data processing method provided by the present embodiment can improve feature recognition accuracy and support efficient and accurate regional planning. First, hyperspectral regional data can be collected for a target region, and target region features can be extracted from the hyperspectral regional data. Then, to improve object classification efficiency, the target region features can be input into an object classification model to classify regional objects in the target region, and object classification information of the regional objects in the target region can be obtained based on the classification results. To enable the classification model to have strong object boundary recognition ability, the object classification model can be optimized for recognition accuracy of mixed pixel regions by using a hierarchical sensitive spectral loss function during the training phase. The optimized object classification model can still have strong boundary recognition accuracy for data collected in complex environments, so as to ensure that the object classification information of the regional objects in the target region output by the model is more accurate. Subsequently, to facilitate the use of regional planning and management, the hyperspectral regional data can be compared with historical hyperspectral regional data of the target region to obtain regional change information of the target region. Then, a visual regional report corresponding to the target region can be constructed based on the object classification information and the regional change information. In the feature analysis and processing, the accuracy and efficiency can be effectively improved, so as to facilitate the use of downstream business for regional planning and management.

[0051] Referring to Figure 1 , Figure 1 A flowchart of a data processing method according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0052] In step S102, hyperspectral regional data is collected for a target region, and target region features are extracted from the hyperspectral regional data.

[0053] The data processing method provided in this embodiment can be applied to any object classification scene, such as a city ground feature classification scene, a housing article classification scene, a park building classification scene, and the like. This embodiment takes the city ground feature classification scene as an example to describe the data processing method, and the description of other scenes can refer to the same or corresponding description in this embodiment, which will not be described in detail herein.

[0054] Specifically, the target region specifically refers to a region that needs to be classified and processed, such as a city region; correspondingly, the hyperspectral region data specifically refers to data obtained by setting a device to collect spectral data for the target region, which can be collected by a hyperspectral sensor configured on a UAV, or by other devices configured with a hyperspectral sensor, such as an airplane. Correspondingly, the target region feature specifically refers to a vector representation of the target region ground feature information extracted from the hyperspectral region data.

[0055] Based on this, in order to improve the ground feature recognition accuracy and support efficient and accurate regional planning, the hyperspectral region data of the target region can be collected first, and the target region feature can be extracted from the hyperspectral region data. On this basis, in order to improve the object classification efficiency, the target region feature can be input into the object classification model to classify the regional objects in the target region, so as to obtain the object classification information of the regional objects in the target region according to the classification result. In order to enable the classification model to have strong object boundary recognition ability, the identification accuracy of the object classification model for mixed pixel regions can be optimized by using a hierarchical sensitive spectrum loss function in the training stage. The optimized object classification model can still have strong boundary recognition accuracy for data collected in complex environments, so as to ensure that the object classification information of the regional objects in the target region output by the model is more accurate. Thereafter, in order to facilitate the use of regional planning and management, the hyperspectral region data can be compared with the historical hyperspectral region data of the target region to obtain the regional change information of the target region. Then, the visual regional report corresponding to the target region can be constructed based on the object classification information and the regional change information. In the ground feature analysis and processing, the accuracy and efficiency can be effectively improved, so as to facilitate the use of downstream business for regional planning and management.

[0056] Further, in order to enable the UAV to collect high-quality spectral data, the flight strategy of the UAV can be planned in the following manner. In this embodiment, the specific implementation is as follows:

[0057] The flight strategy of the UAV is built according to the spectral collection range information, the spectral resolution information, the spatial resolution information, and the flight attribute information; and the UAV is driven to collect the hyperspectral region data of the target region based on the flight strategy.

[0058] Specifically, the collection range information specifically refers to the range of multi-band spectral data collected by the hyperspectral sensor of the unmanned aerial vehicle for the target area, such as 400nm to 2500nm; correspondingly, the spectral resolution information specifically refers to the resolution set by the hyperspectral sensor, such as 5nm; the spatial resolution specifically refers to the resolution that can be achieved for collecting hyperspectral regional data for the target area, such as 10cm; the flight attribute information specifically refers to the flight control information corresponding to the unmanned aerial vehicle, including but not limited to flight speed, heading angle, height, etc.

[0059] Based on this, when collecting hyperspectral regional data for the target area by the unmanned aerial vehicle, in order to ensure that the collected hyperspectral regional data carries more ground object information, a flight strategy can be built for the unmanned aerial vehicle according to the spectral collection range information, the spectral resolution information, the spatial resolution information, and the flight attribute information; thereafter, the unmanned aerial vehicle can be driven to collect hyperspectral regional data for the target area based on the flight strategy, so as to facilitate subsequent ground object classification processing using the hyperspectral regional data collected at this time.

[0060] In summary, by combining multiple dimensions of information to build a flight strategy for the unmanned aerial vehicle, the data collected by the unmanned aerial vehicle when collecting hyperspectral regional data can carry more ground object information, thereby facilitating subsequent high-precision ground object classification.

[0061] Furthermore, in order to ensure that the data collected at different time periods is consistent, the collected data needs to be preprocessed, so as to ensure that the preprocessed data is highly similar. In this embodiment, the specific implementation is as follows:

[0062] Collect initial hyperspectral regional data for the target area by the hyperspectral sensor of the unmanned aerial vehicle; preprocess the initial hyperspectral regional data to obtain hyperspectral regional data, wherein the hyperspectral regional data has a consistent relationship with the historical hyperspectral regional data.

[0063] Specifically, preprocessing specifically refers to data denoising and / or correction and other processing operations on the initial hyperspectral regional data, for making the hyperspectral regional data consistent with the historical hyperspectral regional data collected in the historical period, so as to be used for subsequent regional change information construction.

[0064] Based on this, in order to enable the hyperspectral regional data to be used for the regional change information of the target area after the ground object classification processing, the initial hyperspectral regional data can be collected for the target area by the hyperspectral sensor of the unmanned aerial vehicle, and then the initial hyperspectral regional data can be preprocessed, so that the hyperspectral regional data obtained after preprocessing has a consistent relationship with the historical hyperspectral regional data, thereby facilitating subsequent use.

[0065] On this basis, illumination correction, noise removal and difference correction processing can be used to realize preprocessing. In this embodiment, the specific implementation is as follows:

[0066] The initial hyperspectral regional data is subjected to illumination correction processing to obtain intermediate hyperspectral regional data; the intermediate hyperspectral regional data is subjected to noise removal processing using a Gaussian filter, and the intermediate hyperspectral regional data after noise removal processing is subjected to view angle difference correction processing to obtain target hyperspectral regional data; the signal-to-noise ratio of the target hyperspectral regional data is calculated, and the target hyperspectral regional data is taken as the hyperspectral regional data when it is determined that the target hyperspectral regional data meets the data quality condition according to the signal-to-noise ratio.

[0067] Specifically, the illumination correction processing specifically refers to the processing of correcting the influence of external illumination on the initial hyperspectral regional data, and the intermediate hyperspectral regional data is the hyperspectral regional data obtained after eliminating the influence of illumination. The noise removal processing specifically refers to the processing of eliminating the noise interference carried in the intermediate hyperspectral regional data, which is used to avoid the influence of noise and make the processing result deviate from the true result. Correspondingly, the view angle difference correction processing specifically refers to the operation of image alignment processing on the collected hyperspectral regional data, which is mainly to correct the error caused by the inconsistency of the data collected at different view angles. Correspondingly, the data quality condition specifically refers to the condition of detecting whether the preprocessed hyperspectral regional data can be used for ground object classification.

[0068] Based on this, after obtaining the initial hyperspectral regional data collected by the unmanned aerial vehicle, in order to eliminate the influence of redundant information and make the processed hyperspectral regional data better reflect the ground object information in the region, first, the initial hyperspectral regional data can be subjected to illumination correction processing to eliminate the influence of illumination, and then intermediate hyperspectral regional data is obtained; second, the intermediate hyperspectral regional data can be subjected to noise removal processing using a Gaussian filter to remove redundant information; third, the intermediate hyperspectral regional data after noise removal processing is subjected to view angle difference correction processing to correct the influence of view angle deviation, and target hyperspectral regional data is obtained; finally, the signal-to-noise ratio of the target hyperspectral regional data can be calculated, and the data quality can be detected according to the signal-to-noise ratio, so that the target hyperspectral regional data is taken as the hyperspectral regional data when it is determined that the target hyperspectral regional data meets the data quality condition, so as to be used for subsequent ground object classification.

[0069] In actual application, the collection and preprocessing of hyperspectral regional data can use the hyperspectral sensor carried by the unmanned aerial vehicle to collect data of the target region. The collected data can be subjected to illumination correction, noise removal and view angle difference correction to ensure that the data of different collection periods and regions have consistency and high quality, providing a reliable basis for subsequent analysis and processing.

[0070] In practical implementation, for urban land cover classification scenarios, hyperspectral data acquisition can be achieved by using a drone equipped with a hyperspectral sensor to perform a comprehensive scan of the target urban area. The acquisition range can be set to multi-band spectral data from 400nm to 2500nm, with a spectral resolution of approximately 5nm and a spatial resolution of up to 10cm. The sensor used covers the visible to near-infrared bands, thus enabling detailed recording of land cover information in the target urban area. To ensure comprehensive data coverage, multiple flight paths with different headings and altitudes can be planned during the acquisition process. The drone's flight altitude (e.g., 300 meters) and speed (e.g., 10 m / s) are set according to the area of ​​the target urban area and the sensor resolution to ensure sufficiently fine spatial resolution. The acquired data is stored as a three-dimensional spectral data cube. That is, hyperspectral region data, in which For spatial coordinates, λ is the wavelength.

[0071] Furthermore, after obtaining the hyperspectral region data, it is necessary to preprocess the acquired hyperspectral region data. In this process, illumination correction can be performed first to eliminate the influence of external illumination on the image. In practical applications, atmospheric correction can be performed using the FLAASH method, which is achieved through the following formula (1):

[0072] (1)

[0073] To achieve adjustment of spectral data, among which This is the original data. This is the blackbody radiation value. is the radiation conversion coefficient.

[0074] Furthermore, noise removal can be achieved by using a Gaussian filter to smooth the data, thereby reducing noise interference. The filtering operation formula is as follows: Formula (2):

[0075] (2)

[0076] in, For Gaussian kernel function, For the original image, The filtered image. This is the size of the filter window.

[0077] Based on this, the viewpoint difference correction can be performed; images acquired from different viewpoints can be aligned using feature point matching and registration algorithms. During registration, the error is minimized using the following formula (3):

[0078] (3)

[0079] wherein, and are two influences, is the matching point coordinate, is the total number of matching points.

[0080] Based on the above processing, the geometric correction of the oblique image collected by the unmanned aerial vehicle can be performed, and the hyperspectral regional data is converted into the standard map coordinate system, and then the orthographic image can be generated. In specific implementation, the oblique image can be accurately mapped into the standard map coordinate system through projection transformation or orthographic image generation. The geometric distortion of the image is corrected based on the attitude parameters and GPS positioning information of the unmanned aerial vehicle, and the requirements of the standard map coordinate system are ensured. The geometric correction process can be realized through the following formula (4):

[0081] (4)

[0082] wherein, is the original pixel coordinate, is the corrected coordinate, is the transformation parameter, which is calculated by the calibration model.

[0083] After the pre-processing of the hyperspectral regional data is completed, in order to ensure that the data can be used for subsequent calculation, the quality of the data can be evaluated again to ensure the consistency and reliability of the data. The evaluation index is the signal-to-noise ratio (SNR), which is used to measure the data quality by calculating the ratio of signal to noise, which can be realized through the following formula (5):

[0084] (5)

[0085] wherein, is the signal mean, is the noise standard deviation. After the above processing determines that the hyperspectral regional data collected for the target region can meet the subsequent land feature classification requirements, the subsequent processing operation can be performed.

[0086] In summary, by pre-processing the collected hyperspectral regional data, the data can have high quality and can contain more rich ground feature information, so as to improve the accuracy of subsequent land feature classification.

[0087] After obtaining the hyperspectral regional data, in order to ensure that the extracted features can fully reflect the relevant attributes of the target region, a convolutional neural network can be used to extract regional features. In this embodiment, the specific implementation is as follows:

[0088] The hyperspectral region data is subjected to standardization processing in feature dimension to obtain model output information; the model input information is input into a compressed excitation convolutional neural network for feature extraction to obtain target region features.

[0089] Specifically, the compressed excitation convolutional neural network specifically refers to a convolutional neural network for feature processing of hyperspectral region data. Based on this, in order to ensure that the extracted target region features carry more accurate ground object information, the hyperspectral region data can be subjected to standardization processing in feature dimension, so that the model output information obtained after processing meets the model input requirements. Then the model input information can be input into the compressed excitation convolutional neural network for feature extraction, so as to obtain the target region features for subsequent use.

[0090] In actual application, by selecting the compressed excitation convolutional neural network to process the hyperspectral region data, the model can efficiently capture the subtle differences between different ground object categories, and the sensitivity of the model to complex ground object boundaries (such as the transition zone between buildings and green land) can be improved by enhancing key features, thereby further optimizing the classification accuracy.

[0091] For example, a city planning and development zone needs to provide high-precision city area change reports for each construction stage to facilitate subsequent work guidance. Therefore, in the report construction stage, it is necessary to first accurately classify the ground objects in the development zone, and then mark the change areas. In this process, the unmanned aerial vehicle can first control the hyperspectral sensor carried by the unmanned aerial vehicle to collect hyperspectral region data according to the above scanning settings. After collecting the hyperspectral region data, it can be preprocessed according to formulas (1) to (4) to eliminate the effects of light, noise interference and view angle correction, so that higher quality hyperspectral region data can be obtained; further, after obtaining the hyperspectral region data, the corresponding signal-to-noise ratio S can be calculated by formula (5), and compared with the set threshold value, and if it is greater than the threshold value, it means that the quality of the preprocessed hyperspectral region data is high, and therefore it can be used for subsequent ground object classification and regional change report construction of the development zone.

[0092] In summary, by using the compressed excitation convolutional neural network to extract target region features, the model can better capture the differences between ground objects in the region, thereby improving the subsequent ground object classification accuracy.

[0093] In step S104, the target region features are input into an object classification model to classify the region objects in the target region to obtain object classification information of the region objects in the target region, wherein the object classification model optimizes the recognition accuracy of the object classification model for mixed pixel regions through a hierarchical sensitive spectral loss function.

[0094] Specifically, after obtaining the hyperspectral region data corresponding to the target region, in order to accurately identify the object classification information of the region object in the target region, an object classification model can be used to determine the object classification information. Specifically, the target region features can be input into the object classification model to classify the region objects in the target region, and the object classification information of the region objects in the target region can be output through the classification model. In order to improve the classification accuracy of the region objects, a hierarchical sensitive spectral loss function can be used to optimize the identification accuracy of the object classification model for mixed pixel regions, so that the model can have strong identification ability for the boundary positions of the region objects in the prediction stage, and each region object can be accurately allocated, such as lakes, traffic lights, parks, bridges, etc., so as to facilitate subsequent use.

[0095] The object classification model specifically refers to a model that inputs target region features corresponding to a target region and outputs object classification information corresponding to a region object. The model can be optimized by a hierarchical sensitive spectral loss function, so that the model can learn the ability to divide the boundaries of different region objects. The hierarchical sensitive spectral loss function specifically refers to a loss function used for tuning the model in the training stage. This function can enhance the model's ability to identify the boundaries of different region objects, thereby distinguishing different region objects and outputting real object classification information corresponding to each region object. Accordingly, the mixed pixel region is the region where different region objects meet.

[0096] Further, when classifying the region objects in the target region using the object classification model, in order to improve the classification accuracy, multi-scale and multi-level processing can be used. In this embodiment, the specific implementation is as follows:

[0097] The target region features are input into the object classification model, the hyperspectral cube features corresponding to the region objects in the target region features are convoluted by the multi-scale convolution kernel in the object classification model, and the feature map corresponding to the hyperspectral cube features is obtained. The feature map is dynamically reweighted by the compression excitation unit in the object classification model, and the multi-scale features and multi-level features corresponding to the hyperspectral cube features are determined according to the processing result. The multi-scale features and the multi-level features are spliced into classification features, and the classification unit in the object classification model is used to process the classification features, and the object classification information of the region objects in the target region is obtained.

[0098] Specifically, the multi-scale convolution kernel specifically refers to a plurality of convolution kernels corresponding to different scales; correspondingly, the hyperspectral cube feature specifically refers to the feature expression corresponding to different pixel points in the target region feature, and the feature carries the band information corresponding to the pixel point. Correspondingly, the feature map specifically refers to the feature map obtained after convolution processing. Correspondingly, dynamic reweighting specifically refers to the operation of recalculating the weight and weighting processing, which is realized through the compression excitation unit. Correspondingly, the multi-scale feature and the multi-level feature specifically refer to the feature expression corresponding to the height and depth of the target region.

[0099] Based on this, when the classification processing of the regional object in the target region is performed through the object classification model, the target region feature can be input into the object classification model first, the hyperspectral cube feature corresponding to the regional object in the target region feature is convoluted through the multi-scale convolution kernel in the object classification model, and the feature map corresponding to the hyperspectral cube feature can be obtained according to the processing result; thereafter, the feature map can be dynamically reweighted and processed through the compression excitation unit in the object classification model, so that the multi-scale feature and the multi-level feature corresponding to the hyperspectral cube feature can be determined according to the processing result; on this basis, the multi-scale feature and the multi-level feature can be spliced into the classification feature, so that the classification unit in the object classification model is used to process the classification feature, and the object classification information of the regional object in the target region can be obtained for subsequent use.

[0100] In actual application, when multi-scale classification and fine-grained feature recognition are performed, various features in the urban environment can be classified through the multi-scale convolution network (object classification model). This process can optimize the features with large scale differences such as high-rise buildings, roads and green lands in the city, so as to ensure that small-scale features (such as temporary buildings and narrow green lands) can also be accurately recognized, and the overall classification effect is improved.

[0101] In specific implementation, for the urban feature classification scene, the feature dimension of the above-mentioned preprocessed hyperspectral data can be standardized first, so as to ensure that the input features are within the same scale range, which helps to accelerate model convergence and improve training stability. In specific implementation, the Z-score standardization method can be used, and the standardization formula is as follows:

[0102] (6)

[0103] wherein, and are the mean and standard deviation of the first band respectively. The standardized data can effectively reduce the numerical difference between different bands, and ensure that the convolution operation can learn the features of all bands evenly.

[0104] On this basis, the hyperspectral regional data can be preliminarily feature-extracted by using the convolutional neural network. In this way, the subtle differences between different ground object categories can be effectively captured, and the sensitivity to the boundaries of complex ground objects can be improved by enhancing key features, which is particularly helpful for distinguishing buildings, green land and other ground objects in urban environments, and provides valuable features for subsequent classification tasks.

[0105] In this process, in order to further improve the classification accuracy and optimize the recognition of different scale ground objects in urban environments, a multi-scale convolutional network can be used. By using convolution kernels of different sizes, the model can capture ground object features from multiple scales, especially for small-scale ground objects and fine-grained regions with significant scale differences, which improves the recognition ability. Each layer of convolution kernel extracts corresponding features according to its scale characteristics, ensuring that macro and micro details are effectively captured. Specifically, the input data is a pre-processed hyperspectral cube , the convolution operation uses multi-scale convolution kernels (such as and ), which aims to capture ground object features of different scales. The convolution calculation formula is as follows formula (7):

[0106] (7)

[0107] wherein, is the output feature map, is the th convolution kernel, is the bias, is the activation function (such as ReLU).

[0108] After that, the compression excitation module is introduced again to dynamically re-weight the features between channels and highlight the information of key bands. Specifically, the weight of each channel can be calculated by global average pooling, which can be realized by the following formula (8):

[0109] (8)

[0110] wherein, is the global feature description of the channel , and and are the height and width of the feature map, respectively. Further, the weight vector of the channel is generated by two fully connected layers and an activation function, and the original features are weighted, which is realized by the following formula (9):

[0111] (9)

[0112] Here, i and j are the spatial coordinate parameters of the feature map, where i represents the vertical coordinate of the feature map and j represents the horizontal coordinate of the feature map. i and j together define the position of each pixel in the feature map. Therefore, multiple convolutional layers can be used to extract deep features from the weighted features, thereby further improving the network's ability to distinguish complex terrain features. In addition, to avoid overfitting, regularization measures can be introduced after the convolution operation.

[0113] Finally, the extracted multi-scale and multi-level features are integrated. This involves concatenating the feature maps output from all convolutional layers along the channel dimension to form a global feature vector, as shown in formula (10):

[0114] (10)

[0115] in, This indicates a feature concatenation operation. For the first The integrated feature vectors will serve as input to the subsequent classification module, providing the model with richer spectral and spatial information. Based on this, each feature vector will be processed by the classification layer in the model to obtain the classification information for each land cover in the urban area, facilitating the subsequent construction of visualization reports.

[0116] In summary, by using an object classification model to process the features of the target region, the recognition accuracy of objects at different scales can be enhanced from multiple perspectives during the processing, thereby ensuring that the final output object classification information is more consistent with the actual situation.

[0117] Furthermore, to achieve high recognition accuracy in the object classification model and to effectively remove object boundaries within a region, two loss functions can be combined for model optimization. In this embodiment, the training of the object classification model includes:

[0118] The object classification sample is obtained and input into an initial object classification model for processing to obtain predicted classification information. A boundary loss value is calculated based on the predicted classification information and sample classification information corresponding to the object classification sample according to a boundary recognition loss function, and an object loss value is calculated based on the predicted classification information and the sample classification information according to an object recognition loss function. The initial object classification model is optimized based on the boundary loss value and the object loss value until the object classification model that meets the training stop condition is obtained. The boundary recognition loss function and the object recognition loss function constitute the hierarchical sensitive spectral loss function, and the initial object classification model is optimized by the boundary recognition loss function to improve the recognition accuracy of the object boundary, and the initial object classification model is optimized by the object recognition loss function to improve the recognition accuracy of the regional object.

[0119] Specifically, the object classification sample specifically refers to a sample corresponding to the target region and having the same target region feature structure as the model training phase. Correspondingly, the predicted classification information specifically refers to the classification information obtained after identifying the sample region object in the sample. The boundary recognition loss function specifically refers to a loss function that enhances the model learning of the object boundary recognition accuracy. The object recognition loss function specifically refers to a loss function that enhances the model learning of the object recognition accuracy. Correspondingly, the training stop condition specifically refers to the condition for stopping training the object classification model, which includes but is not limited to the loss value comparison condition, the iteration number condition, or the validation set verification condition, etc. In specific implementation, it can be selected according to actual needs, and the present embodiment does not make any limitation here.

[0120] Based on this, in order to enable the object classification model to have strong object recognition ability in the application stage, not only can each regional object be accurately recognized, but also different regional objects can be effectively divided, so as to construct a visual report for subsequent use. A double-loss function can be used to train it. Specifically, the object classification sample can be obtained first, and input into the initial object classification model for processing to obtain the predicted classification information. After obtaining the predicted classification information, according to the label corresponding to the sample, the boundary recognition loss value can be calculated based on the predicted classification information and the sample classification information corresponding to the object classification sample according to the boundary recognition loss function, and the object loss value can be calculated based on the predicted classification information and the sample classification information according to the object recognition loss function. The boundary loss value can reflect the accuracy of the object classification model in the current stage for boundary recognition, and the object loss value can reflect the recognition accuracy of the object classification model in the current stage for the object. Therefore, the initial object classification model can be optimized based on the boundary loss value and the object loss value. If the optimized model does not meet the training stop condition, new samples can be selected for training until an object classification model that meets the training stop condition is obtained and deployed for use. The boundary recognition loss function and the object recognition loss function constitute a hierarchical sensitive spectral loss function, and the initial object classification model is optimized for the recognition accuracy of the object boundary by the boundary recognition loss function, and the recognition accuracy of the regional object by the object recognition loss function.

[0121] In practical applications, when optimizing the object classification model, in order to improve the recognition accuracy of the model at the boundary, a hierarchical sensitive spectral loss function can be used to optimize the classification accuracy of hyperspectral regional data, especially in complex mixed pixel regions. By improving the recognition ability of different ground object type boundaries such as buildings and green areas, the problem of fine classification confusion in hyperspectral regional data can be solved, and the accuracy of the classification result can be ensured.

[0122] In specific implementation, for the city ground object classification scene, the training of the model can first optimize the classification accuracy in the mixed pixel region of the hyperspectral regional data. By weighting each layer of features, the recognition ability of the model for the boundaries of ground object types such as buildings and green areas is improved. The specific loss function formula is as follows formula (11):

[0123] (11)

[0124] wherein, is the loss of the i-th layer, is a weighting coefficient.

[0125] Further, for the mixed pixel region, a local sensitive spectral loss function ​The recognition of fine-grained objects is strengthened. The loss function formula is as follows formula (12):

[0126] (12)

[0127] wherein, is a predicted spectrum, is a real spectrum, is a category weighting coefficient.

[0128] On this basis, the classification loss and the optimized hierarchical sensitive spectrum loss are integrated to obtain the final optimized loss function, as follows formula (13)

[0129] (13)

[0130] The optimized model can effectively improve the classification accuracy of complex mixed pixel regions and improve the overall classification effect of hyperspectral data.

[0131] In the above example, after obtaining the high-quality hyperspectral region data corresponding to the development zone, the feature expression V corresponding to the hyperspectral region data can be extracted first, and then the hyperspectral region data can be input into the object classification model optimized by formula (11) to (13) for processing. After the hyperspectral region data is input into the object classification model, the model can complete the classification processing of each object in the development zone according to formula (6) to (10). After the classification processing, the type information of each position corresponding object can be marked on the development zone map matched with the hyperspectral region data according to the mapping relationship between them, so as to be used later.

[0132] In summary, by optimizing the object classification model using the boundary recognition loss function and the object recognition loss function, the model can improve the recognition accuracy of the object boundary while ensuring the object recognition accuracy, thereby effectively improving the object classification accuracy in any scene in the application stage.

[0133] Step S106, comparing the hyperspectral region data with the historical hyperspectral region data of the target region to obtain the region change information of the target region.

[0134] Specifically, after obtaining the object classification information corresponding to each region object in the target region, it is explained that the object classification processing of the region objects in the target region is completed at this time. In order to provide accurate reference for the planning and management of the target region, the hyperspectral region data can be compared with the historical hyperspectral region data of the target region, and the region change information of the target region can be obtained according to the comparison result. Subsequently, the visualization region report corresponding to the target region can be constructed by combining the region change information and the object classification information, thereby providing a reference for the planning and management of the target region.

[0135] The historical hyperspectral region data is specifically hyperspectral region data collected in the previous collection cycle according to the same setting as the hyperspectral region data. By comparing the hyperspectral region data collected in the current collection cycle with the historical hyperspectral region data, the regional change of the target region within the time between the two collection cycles can be determined, so as to facilitate subsequent construction of the report. The regional change information is specifically information of the regional change in the visualization dimension obtained by comparing the hyperspectral region data of two adjacent cycles. It can be understood as image difference information existing after comparison of spectral images.

[0136] Further, after obtaining the regional change information, in order to avoid the influence caused by inaccurate data collection, the regional change condition detection can be performed before the construction of the visual regional report, thereby improving the accuracy. In this embodiment, the specific implementation is as follows:

[0137] At least two change regions corresponding to the regional change information are determined, and spatial aggregation is performed on the at least two change regions to obtain a global change region. It is detected whether the global change region satisfies a regional change condition corresponding to the target region. If yes, the step of constructing the visual regional report corresponding to the target region based on the object classification information and the regional change information is performed.

[0138] Specifically, the global change region is specifically a change region obtained by spatial aggregation processing on at least two change regions corresponding to the regional change information. Correspondingly, the regional change condition is specifically a condition for detecting whether the change region corresponding to the regional change information determined in the current cycle satisfies the true situation.

[0139] Based on this, when constructing the visual regional report, at least two change regions corresponding to the regional change information can be determined first. At this time, spatial aggregation can be performed on the at least two change regions to obtain a global change region. Then, it is detected whether the global change region satisfies a regional change condition corresponding to the target region. If not, it is further indicated that the regional change information determined by processing the data collected in the current cycle does not conform to the true situation, which further indicates that the data may be incorrect. Therefore, data collection and processing can be performed again. If yes, step S108 can be performed.

[0140] In actual application, time series analysis technology can be used to compare hyperspectral region data at different time nodes to identify changes in urban regions. Through the change detection module, new buildings, green land occupation and other change regions can be accurately positioned, and detailed change reports can be generated to provide decision basis for dynamic management of cities.

[0141] That is, for the urban feature classification scene, time series analysis techniques can be used to compare and analyze hyperspectral regional data at different time nodes. By calculating the change of each pixel in the time dimension, areas with significant changes are identified. The specific change detection method can be expressed as formula (14) as follows:

[0142] (14)

[0143] wherein, and are hyperspectral data at time nodes and , respectively, is a waveband, represents the spectral difference of pixel position at different time points.

[0144] Further, the change detection module can be used to accurately locate the change area. By setting a change threshold , the difference image is binarized to identify new buildings, green space occupation and other change areas. The process can be achieved through formula (15) as follows:

[0145] (15)

[0146] If the difference value exceeds a certain preset threshold , it is considered that the position has changed. The threshold is obtained by analyzing historical data to distinguish between changed and unchanged areas.

[0147] Further, spatial aggregation of all change areas can further analyze whether the changed area conforms to the actual change, such as new buildings, green space expansion, etc. At this time, the change value of each pixel is spatially aggregated to obtain the overall change value of each region, which is calculated through formula (16) as follows:

[0148] (16)

[0149] If is greater than a certain global threshold , it is considered that the area has changed. After obtaining the regional change information, subsequent construction of a visual regional report for the target area can be achieved for planning and management of the target area.

[0150] With the above example, after obtaining the classification information corresponding to each ground object in the development zone, in order to determine the construction changes in the development zone in the current period and the last period, the corresponding historical hyperspectral regional data of the development zone can be obtained. Then the regional change information corresponding to the adjacent two periods of the development zone can be calculated through the above formulas (14) to (16), so as to generate a change report of the development zone in combination with the classification information of the ground object in the subsequent period, for the planning and management construction of the development zone.

[0151] In summary, by using the comparison method of adjacent period hyperspectral data to determine the regional change information, the change of the target region can be accurately determined, so as to be used in the construction report.

[0152] Step S108, constructing a visual regional report corresponding to the target region based on the object classification information and the regional change information.

[0153] Specifically, after obtaining the regional change information and the object classification information corresponding to the regional object, in order to provide a reference for the planning and management of the target region, the visual regional report corresponding to the target region can be constructed by combining the object classification information and the regional change information. The visual regional report specifically refers to a report that directly reflects the regional change of the target region and the type corresponding to each regional object, for the planning and management of the target region.

[0154] Further, in constructing the visual regional report, the fusion of classification information and change information can be realized by constructing a classification map and updating. In this embodiment, the specific implementation is as follows:

[0155] According to the object classification information, a classification map corresponding to the target region is constructed; the classification map is updated according to the regional change information to obtain an initial change region marking map; a spatial difference algorithm is used to smooth the initial change region marking map, and a visual regional report corresponding to the target region is constructed according to the smoothing result.

[0156] Specifically, the classification map specifically refers to a map constructed for the target region according to the object classification information, which carries the type description of each regional object. Correspondingly, the initial change region marking map specifically refers to a map recording the regional change. Correspondingly, the smoothing processing specifically refers to repairing the initial change region marking map, so that the visually displayed map is more in line with the viewing needs.

[0157] Based on this, when constructing a visual region report corresponding to the target region, a classification map corresponding to the target region can be constructed first based on the object classification information; then, the classification map is updated according to the region change information to obtain an initial change region marker map, thereby enabling the region change information and object classification information to be carried in the map. Finally, the initial change region marker map can be smoothed using a preset spatial difference algorithm to construct a visual region report corresponding to the target region based on the smoothing result.

[0158] In practical applications, when generating visual regional reports, the changing areas and the classification information corresponding to each regional object can be marked on the spatial distribution map of the target area. This makes it easier to intuitively display the combination of classifications and the monitored areas, which is useful for regional management and planning.

[0159] In practical implementation, for urban land cover classification scenarios, a classification map can be constructed based on the above classification results and change detection results. ,in Indicates position The classification of land features can be based on the Euclidean distance metric, which measures the distance between each category in the feature space. Specifically, for each pixel... The Euclidean distance between its spectral feature vector and the center vector of each land cover category is calculated using the following formula (17):

[0160] (17)

[0161] in, Location in hyperspectral region data First Spectral values ​​of the band It is a category exist Mean over the band This represents the number of spectral bands. By calculating the Euclidean distance between each pixel and the center of each category, each pixel will be classified into the category with the smallest distance.

[0162] Furthermore, the change detection results can be combined with the classification results to generate a change area marker map. By marking the change areas in the classification results, new buildings, green space expansions, and other change areas can be effectively identified and marked. The calculation formula (18) is as follows:

[0163] (18)

[0164] in, The map is marked with a 1 to indicate that the area has changed. This indicates the original category label.

[0165] Based on this, the classification results can be smoothed using a spatial interpolation method to avoid noise or unclear boundaries that may exist in the classification process. For example, a bilinear interpolation method can be used to smooth the classification results, making the spatial distribution map more coherent. The interpolation formula (19) is as follows:

[0166] (19)

[0167] wherein, is the smoothed classification result, is the position of the adjacent pixel.

[0168] On this basis, a visual report can be generated. In specific implementation, the classification results and the change area can be intuitively displayed through color coding and symbol labeling, etc. For example, different land types are marked with different colors, and the change area can be highlighted for easy observation. The generation formula of the visualization map (20) is as follows:

[0169] (20)

[0170] wherein, is the final generated visualization image, is a function for combining the classification results and the change area for visualization.

[0171] The final generated visualization area report will show the spatial distribution of various land features in the city and mark the areas that have changed, helping the city management department to intuitively understand the city development and change, and further supporting city planning and decision-making.

[0172] Following the above example, after determining the area change information and the land feature classification information, the construction of the visualization area report of the development zone can be completed by combining the above formulas (17) to (20). Through the report, the areas that have changed in the development zone and the description of the areas that have been completed can be recorded, thereby facilitating the management and planning of the development zone by the management department.

[0173] The data processing method provided by the embodiment can improve the ground object recognition accuracy and support efficient and accurate regional planning. The hyperspectral regional data of the target region can be collected, and the target region features can be extracted from the hyperspectral regional data. On this basis, the target region features can be input into an object classification model to classify the regional objects in the target region, so as to obtain the object classification information of the regional objects in the target region according to the classification result. In the training stage, the object classification model can be optimized by using a hierarchical sensitive spectrum loss function to improve the recognition accuracy of the mixed pixel region, so that the optimized object classification model can still have strong boundary recognition accuracy for the data collected in a complex environment, thereby ensuring that the object classification information of the regional objects in the target region output by the model is more accurate. Then, the hyperspectral regional data can be compared with the historical hyperspectral regional data of the target region to obtain the regional change information of the target region, and then a visual regional report corresponding to the target region can be constructed based on the object classification information and the regional change information. In the ground object analysis and processing, the accuracy and efficiency can be effectively improved, thereby facilitating the downstream business to use the regional planning and management.

[0174] Referring to Figure 2 , Figure 2 A flowchart of a city ground object classification method according to one embodiment of the present specification is shown, which specifically includes the following steps.

[0175] In step S202, hyperspectral regional data of a city region is collected, and city region features are extracted from the hyperspectral regional data.

[0176] In step S204, the city region features are input into a classification model to classify the ground object in the city region, and ground object classification information of the ground object in the city region is obtained, wherein the classification model is optimized by using a hierarchical sensitive spectrum loss function to improve the recognition accuracy of the classification model for the mixed pixel region.

[0177] In step S206, the hyperspectral regional data is compared with the historical hyperspectral regional data of the city region to obtain the regional change information of the city region.

[0178] In step S208, a city regional change report corresponding to the city region is constructed based on the ground object classification information and the regional change information.

[0179] The city ground object classification method provided by the embodiment is applied to the scene of classifying ground objects in a city area, and can be used for planning and management of the city according to the classification result. The description of the city ground object classification method can be referred to the same or corresponding description in the above embodiments, which will not be repeated here.

[0180] The following describes the data processing method in combination with the accompanying Figure 3 The data processing method is further described by taking the application of the data processing method in the city planning scene as an example. Wherein, Figure 3 A processing process flow diagram of a data processing method provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0181] In step S302, the initial hyperspectral region data is collected by the hyperspectral sensor of the unmanned aerial vehicle for the target region, the initial hyperspectral region data is subjected to illumination correction processing, and intermediate hyperspectral region data is obtained.

[0182] In step S304, the intermediate hyperspectral region data is subjected to denoising processing by using a Gaussian filter, and the intermediate hyperspectral region data after denoising processing is subjected to view angle difference correction processing, and target hyperspectral region data is obtained.

[0183] In step S306, the signal-to-noise ratio of the target hyperspectral region data is calculated, and the target hyperspectral region data is determined as the hyperspectral region data when the signal-to-noise ratio satisfies the data quality condition, wherein the hyperspectral region data has a consistent relationship with the historical hyperspectral region data.

[0184] In step S308, the hyperspectral region data is subjected to standardization processing of feature dimensions, and model output information is obtained. The model input information is input into the compressed excitation convolutional neural network for feature extraction, and target region features are obtained.

[0185] In step S310, the target region features are input into the object classification model, and the hyperspectral cube features of the corresponding region objects in the target region features are subjected to convolution processing by the multi-scale convolution kernel in the object classification model, and the feature map corresponding to the hyperspectral cube features is obtained.

[0186] In step S312, the feature map is subjected to dynamic reweighting processing by the compressed excitation unit in the object classification model, and the multi-scale features and multi-level features corresponding to the hyperspectral cube features are determined according to the processing result.

[0187] In step S314, the multi-scale features and multi-level features are spliced into classification features, and the classification features are processed by the classification unit in the object classification model, and the object classification information of the region objects in the target region is obtained.

[0188] Step S316, comparing the hyperspectral region data with historical hyperspectral region data of the target region to obtain region change information of the target region.

[0189] Step S318, constructing a classification map corresponding to the target region according to the object classification information, updating the classification map according to the region change information, and obtaining an initial change region marking map.

[0190] Step S320, smoothing the initial change region marking map by using a preset spatial difference algorithm, and constructing a visual region report corresponding to the target region according to a smoothing result.

[0191] To sum up, in order to improve the feature recognition accuracy and support efficient and accurate region planning, the hyperspectral region data of the target region can be collected first, and the target region features can be extracted from the hyperspectral region data. On this basis, in order to improve the object classification efficiency, the target region features can be input into the object classification model to classify the region objects in the target region, so as to obtain the object classification information of the region objects in the target region according to the classification result. In order to enable the classification model to have strong object boundary recognition ability, the recognition accuracy of the object classification model for mixed pixel regions can be optimized by using a hierarchical sensitive spectrum loss function in the training stage. The optimized object classification model can still have strong boundary recognition accuracy for data collected in complex environments, so as to ensure that the object classification information of the region objects in the target region output by the model is more accurate. Thereafter, in order to facilitate the use of region planning and management, the hyperspectral region data can be compared with the historical hyperspectral region data of the target region to obtain the region change information of the target region. Then, the visual region report corresponding to the target region can be constructed based on the object classification information and the region change information. In the feature analysis and processing, the accuracy and efficiency can be effectively improved, so as to facilitate the use of region planning and management by downstream businesses.

[0192] Corresponding to the method embodiments described above, the present specification also provides data processing device embodiments, Figure 4 A structural schematic diagram of a data processing device provided by one embodiment of the present specification is shown. As shown in the figure, Figure 4 The device comprises:

[0193] The collection module 402 is configured to collect hyperspectral region data for a target region, and extract target region features from the hyperspectral region data.

[0194] The input module 404 is configured to input the target region feature into an object classification model to classify region objects in the target region and obtain object classification information of the region objects in the target region, wherein the object classification model is optimized by a hierarchical sensitive spectral loss function to improve the recognition accuracy of the object classification model for mixed pixel regions.

[0195] The comparison module 406 is configured to compare the hyperspectral region data with historical hyperspectral region data of the target region to obtain region change information of the target region.

[0196] The construction module 408 is configured to construct a visual region report corresponding to the target region based on the object classification information and the region change information.

[0197] In an optional embodiment, the acquisition module 402 is further configured to:

[0198] acquire initial hyperspectral region data for the target region by using a hyperspectral sensor of a UAV; and pre-process the initial hyperspectral region data to obtain the hyperspectral region data, wherein the hyperspectral region data has a consistent relationship with the historical hyperspectral region data.

[0199] In an optional embodiment, the acquisition module 402 is further configured to:

[0200] perform illumination correction processing on the initial hyperspectral region data to obtain intermediate hyperspectral region data; perform denoising processing on the intermediate hyperspectral region data by using a Gaussian filter, and perform perspective difference correction processing on the intermediate hyperspectral region data after the denoising processing to obtain target hyperspectral region data; calculate a signal-to-noise ratio of the target hyperspectral region data, and determine, according to the signal-to-noise ratio, whether the target hyperspectral region data meets a data quality condition, and if so, use the target hyperspectral region data as the hyperspectral region data.

[0201] In an optional embodiment, the acquisition module 402 is further configured to:

[0202] perform standardization processing on a feature dimension of the hyperspectral region data to obtain model input information; and input the model input information into a compressed excitation convolutional neural network to extract features to obtain target region features.

[0203] In an optional embodiment, the input module 404 is further configured to:

[0204] input the target region feature into an object classification model, perform convolution processing on the hyperspectral cube feature corresponding to the region object in the target region feature by using a multi-scale convolution kernel in the object classification model, and obtain a feature map corresponding to the hyperspectral cube feature; perform dynamic reweighting processing on the feature map by using a compression excitation unit in the object classification model, and determine a multi-scale feature and a multi-level feature corresponding to the hyperspectral cube feature according to a processing result; splice the multi-scale feature and the multi-level feature into a feature to be classified, and perform processing on the feature to be classified by using a classification unit in the object classification model, and obtain object classification information of the region object in the target region.

[0205] In an optional embodiment, the training of the object classification model comprises:

[0206] An object classification sample is obtained, and the object classification sample is input into an initial object classification model for processing to obtain predicted classification information. A boundary loss value is calculated based on the predicted classification information and sample classification information corresponding to the object classification sample according to a boundary recognition loss function, and an object loss value is calculated based on the predicted classification information and the sample classification information according to an object recognition loss function. The initial object classification model is optimized based on the boundary loss value and the object loss value until the object classification model meeting a training stop condition is obtained. The boundary recognition loss function and the object recognition loss function constitute the hierarchical sensitive spectral loss function, and the initial object classification model is optimized by using the boundary recognition loss function to improve the recognition accuracy of the object boundary and by using the object recognition loss function to improve the recognition accuracy of the region object.

[0207] In an optional embodiment, the apparatus further comprises:

[0208] The detection module is configured to determine at least two change regions corresponding to the region change information, perform spatial aggregation on the at least two change regions, and obtain a global change region. It is determined whether the global change region meets a region change condition corresponding to the target region. If yes, the step of constructing a visual region report corresponding to the target region based on the object classification information and the region change information is performed.

[0209] In an optional embodiment, the construction module 408 is further configured to:

[0210] construct a classification map corresponding to the target region according to the object classification information; update the classification map according to the region change information to obtain an initial change region marking map; perform smoothing processing on the initial change region marking map by using a preset spatial difference algorithm, and construct a visual region report corresponding to the target region according to a smoothing processing result.

[0211] In an optional embodiment, the collection module 402 is further configured to:

[0212] construct a flight strategy for the unmanned aerial vehicle according to the spectral collection range information, the spectral resolution information, the spatial resolution information, and the flight attribute information; and drive the unmanned aerial vehicle to collect the hyperspectral region data for the target region based on the flight strategy.

[0213] The data processing apparatus provided in this embodiment can improve the feature recognition accuracy of ground objects and support efficient and accurate region planning. The hyperspectral region data for the target region can be collected, and the target region features can be extracted from the hyperspectral region data. On this basis, the target region features can be input into an object classification model to classify the region objects in the target region, so as to obtain the object classification information of the region objects in the target region according to the classification result. In the training stage, the object classification model can be optimized by using a hierarchical sensitive spectral loss function to improve the recognition accuracy of the mixed pixel region, so that the optimized object classification model can still have strong boundary recognition accuracy for the data collected in a complex environment, thereby ensuring that the object classification information of the region objects in the target region output by the model is more accurate. Thereafter, the hyperspectral region data can be compared with the historical hyperspectral region data of the target region to obtain the region change information of the target region, and then a visual region report corresponding to the target region can be constructed based on the object classification information and the region change information. In the analysis and processing of ground objects, the accuracy and efficiency can be effectively improved, thereby facilitating the use of region planning and management by downstream businesses.

[0214] The above is a schematic scheme of the data processing apparatus of this embodiment. It should be noted that the technical scheme of the data processing apparatus belongs to the same concept as the technical scheme of the data processing method described above, and the details of the technical scheme of the data processing apparatus that are not described in detail can be referred to the description of the technical scheme of the data processing method.

[0215] Corresponding to the method embodiments described above, the present specification also provides a city ground object classification apparatus embodiment, Figure 5 A structural schematic diagram of a city ground object classification apparatus provided by one embodiment of the present specification is shown. As shown in the figure, Figure 5 the apparatus comprises:

[0216] The data collection module 502 is configured to collect hyperspectral regional data for a city area and extract city area features from the hyperspectral regional data;

[0217] The input model module 504 is configured to input the city area features into a classification model to classify objects in the city area, and obtain object classification information of the objects in the city area, wherein the classification model is optimized by a hierarchical sensitive spectral loss function to improve the identification accuracy of the classification model for mixed pixel regions;

[0218] The data comparison module 506 is configured to compare the hyperspectral regional data with historical hyperspectral regional data of the city area, and obtain regional change information of the city area;

[0219] The report construction module 508 is configured to construct a city area change report corresponding to the city area based on the object classification information and the regional change information.

[0220] The above is a schematic scheme of the city object classification device of the embodiment. It should be noted that the technical scheme of the city object classification device and the technical scheme of the city object classification method described above belong to the same concept. The technical scheme of the city object classification device is not described in detail, and the details can be referred to the description of the technical scheme of the city object classification method.

[0221] Figure 6 A structural block diagram of a computing device 600 according to an embodiment of the present specification is shown. The components of the computing device 600 include but are not limited to a memory 610 and a processor 620. The processor 620 is connected to the memory 610 through a bus 630, and a database 650 is used to save data.

[0222] The computing device 600 also includes an access device 640 that enables the computing device 600 to communicate via one or more networks 660. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or combinations of such networks, such as the Internet. The access device 640 can include one or more of any type of network interface (for example, a network interface card (NIC)) such as a wired or wireless network interface, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC).

[0223] In one embodiment of the present specification, the above-mentioned components of the computing device 600 and other components not shown in the Figure 6 may be connected to each other, for example, through a bus. It should be understood that Figure 6 The computing device structure diagram shown is only for the purpose of example, and is not a limitation on the scope of the present specification. Those skilled in the art can add or replace other components as needed.

[0224] The computing device 600 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (for example, a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, and the like), a mobile phone (for example, a smartphone), a wearable computing device (for example, a smart watch, smart glasses, and the like), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 600 can also be a mobile or stationary server.

[0225] The processor 620 is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned data processing method or urban ground object classification method.

[0226] The above is a schematic scheme of the computing device of the embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the data processing method or the urban ground object classification method belong to the same concept, and the details of the technical scheme of the computing device that are not described in detail can be referred to the description of the technical scheme of the data processing method or the urban ground object classification method.

[0227] An embodiment of the present specification further provides a computer readable storage medium storing computer executable instructions, and the computer executable instructions realize the steps of the data processing method or the urban ground object classification method when executed by a processor.

[0228] The above is a schematic scheme of the computer readable storage medium of the embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the data processing method or the urban ground object classification method belong to the same concept, and the details of the technical scheme of the storage medium that are not described in detail can be referred to the description of the technical scheme of the data processing method or the urban ground object classification method.

[0229] An embodiment of the present specification further provides a computer program product, comprising a computer program or instructions, and the computer program or instructions realize the steps of the data processing method or the urban ground object classification method when executed by a processor.

[0230] The above is a schematic scheme of the computer program product of the embodiment. It should be noted that the technical scheme of the computer program product and the technical scheme of the data processing method or the urban ground object classification method belong to the same concept, and the details of the technical scheme of the computer program product that are not described in detail can be referred to the description of the technical scheme of the data processing method or the urban ground object classification method.

[0231] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order than the order in which they are recited and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0232] The computer readable medium can include any entity or apparatus capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, software distribution medium, etc. It should be noted that the computer readable medium can include appropriate additions or subtractions according to the requirements of patent practice. For example, according to the patent practice in some regions, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0233] It should be noted that, for the foregoing method embodiments, in order to facilitate description, each is described as a combination of a series of acts, but those skilled in the art should appreciate that the embodiments of the present specification are not limited by the order of the described acts, because according to the embodiments of the present specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should appreciate that the embodiments described in the specification are all preferred embodiments, and the acts and modules involved are not necessarily essential to the embodiments of the present specification.

[0234] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0235] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details and do not limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and changes can be made. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and use the present specification.

Claims

1. A data processing method, characterized in that, include: Hyperspectral data is collected for the target region, and the target region features are extracted from the hyperspectral data. The target region features are input into an object classification model to classify regional objects in the target region, thereby obtaining object classification information for the regional objects in the target region. The object classification model optimizes the recognition accuracy of the object classification model for mixed pixel regions through a hierarchical sensitive spectral loss function. The hierarchical sensitive spectral loss function consists of a boundary recognition loss function and an object recognition loss function. The boundary recognition loss function is used to optimize the object classification model's recognition accuracy for object boundaries, and the object recognition loss function is used to optimize the object classification model's recognition accuracy for regional objects. The hyperspectral region data is compared with the historical hyperspectral region data of the target region to obtain the regional change information of the target region; Based on the object classification information and the regional change information, a visual regional report corresponding to the target region is constructed.

2. The data processing method according to claim 1, characterized in that, The acquisition of hyperspectral data for the target region includes: Initial hyperspectral region data was collected for the target area using the hyperspectral sensor of the drone. The initial hyperspectral region data is preprocessed to obtain hyperspectral region data, wherein the hyperspectral region data has a consistent relationship with the historical hyperspectral region data.

3. The data processing method according to claim 2, characterized in that, The preprocessing of the initial hyperspectral region data to obtain hyperspectral region data includes: The initial hyperspectral region data is subjected to illumination correction processing to obtain intermediate hyperspectral region data; The intermediate hyperspectral region data is denoised using a Gaussian filter, and the denoised intermediate hyperspectral region data is then subjected to viewpoint difference correction to obtain the target hyperspectral region data. Calculate the signal-to-noise ratio (SNR) of the target hyperspectral region data, and if the target hyperspectral region data meets the data quality conditions based on the SNR, then use the target hyperspectral region data as the hyperspectral region data.

4. The data processing method according to claim 1, characterized in that, Extracting target region features from the hyperspectral data includes: The hyperspectral region data is standardized in terms of feature dimensions to obtain the model output information; The model input information is fed into a compressed excitation convolutional neural network for feature extraction to obtain the features of the target region.

5. The data processing method according to claim 1, characterized in that, The step of inputting the target region features into an object classification model to classify regional objects in the target region and obtaining object classification information for the regional objects in the target region includes: The target region features are input into an object classification model. The hyperspectral cubic features corresponding to the objects in the target region features are convolved by the multi-scale convolution kernel in the object classification model to obtain the feature map corresponding to the hyperspectral cubic features. The feature map is dynamically reweighted by the compression excitation unit in the object classification model, and the multi-scale features and multi-level features corresponding to the hyperspectral cubic features are determined based on the processing results. The multi-scale features and the multi-level features are concatenated into features to be classified, and the classification units in the object classification model are used to process the features to be classified to obtain the object classification information of regional objects in the target region.

6. The data processing method according to any one of claims 1 to 5, characterized in that, The training of the object classification model includes: Obtain object classification samples and input them into an initial object classification model for processing to obtain predicted classification information; The boundary loss value is calculated based on the predicted classification information and the sample classification information corresponding to the object classification sample according to the boundary recognition loss function, and the object loss value is calculated based on the predicted classification information and the sample classification sample according to the object recognition loss function; The initial object classification model is optimized based on the boundary loss value and the object loss value until the object classification model that meets the training stopping condition is obtained. The boundary recognition loss function and the object recognition loss function together constitute the hierarchical sensitive spectral loss function. The boundary recognition loss function is used to optimize the initial object classification model's accuracy in recognizing object boundaries, and the object recognition loss function is used to optimize the initial object classification model's accuracy in recognizing regional objects.

7. The data processing method according to claim 1, characterized in that, Before the step of constructing a visual region report corresponding to the target region based on the object classification information and the region change information is executed, the following steps are also included: Identify at least two changing regions corresponding to the regional change information, and spatially aggregate the at least two changing regions to obtain the global changing region; Detect whether the global change region satisfies the region change condition corresponding to the target region; If so, proceed with the step of constructing a visual region report corresponding to the target region based on the object classification information and the region change information.

8. The data processing method according to claim 1, characterized in that, The process of constructing a visual region report corresponding to the target region based on the object classification information and the region change information includes: Construct a classification map corresponding to the target area based on the object classification information; The classification map is updated according to the regional change information to obtain an initial map of changed regional markers; The initial changed area marker map is smoothed using a preset spatial difference algorithm, and a visualization report of the target area is constructed based on the smoothing result.

9. The data processing method according to any one of claims 1 to 5, characterized in that, The acquisition of hyperspectral data for the target region includes: A flight strategy is constructed for the UAV based on spectral acquisition range information, spectral resolution information, spatial resolution information, and flight direction attribute information; The flight strategy described above drives the UAV to collect hyperspectral data in the target area.

10. A method for classifying urban land features, characterized in that, include: Hyperspectral data is collected for urban areas, and urban area features are extracted from the hyperspectral data. The urban area features are input into an object classification model to classify ground objects in the urban area, thereby obtaining ground object classification information. The object classification model optimizes its recognition accuracy for mixed pixel regions using a hierarchical sensitive spectral loss function, which consists of a boundary recognition loss function and an object recognition loss function. The boundary recognition loss function optimizes the object classification model's accuracy for object boundaries, while the object recognition loss function optimizes its accuracy for regional objects. The hyperspectral region data is compared with the historical hyperspectral region data of the urban area to obtain the regional change information of the urban area; Based on the land feature classification information and the regional change information, an urban area change report corresponding to the urban area is constructed.

11. A data processing apparatus, characterized in that, include: The acquisition module is configured to acquire hyperspectral data for a target region and extract target region features from the hyperspectral data. The input module is configured to input the target region features into an object classification model to classify regional objects in the target region and obtain object classification information of the regional objects in the target region. The object classification model optimizes its recognition accuracy for mixed pixel regions using a hierarchical sensitive spectral loss function, which consists of a boundary recognition loss function and an object recognition loss function. The boundary recognition loss function optimizes the object classification model's accuracy for recognizing object boundaries, and the object recognition loss function optimizes its accuracy for recognizing regional objects. The comparison module is configured to compare the hyperspectral region data with the historical hyperspectral region data of the target region to obtain the regional change information of the target region; The construction module is configured to construct a visual region report corresponding to the target region based on the object classification information and the region change information.

12. An urban land cover classification device, characterized in that, include: The data acquisition module is configured to acquire hyperspectral data for urban areas and extract urban area features from the hyperspectral data. The input model module is configured to input the urban area features into an object classification model to classify the ground objects in the urban area and obtain the ground object classification information of the ground objects in the urban area. The object classification model optimizes the recognition accuracy of the object classification model for mixed pixel areas through a hierarchical sensitive spectral loss function. The hierarchical sensitive spectral loss function consists of a boundary recognition loss function and an object recognition loss function. The boundary recognition loss function is used to optimize the object classification model for the recognition accuracy of object boundaries, and the object recognition loss function is used to optimize the object classification model for the recognition accuracy of regional objects. The comparison data module is configured to compare the hyperspectral region data with the historical hyperspectral region data of the urban area to obtain regional change information of the urban area. The report building module is configured to build an urban area change report corresponding to the urban area based on the land feature classification information and the regional change information.

13. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 10.

15. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 10.

Images