CN114120050A

CN114120050A - Method, device and equipment for extracting surface ecological data and storage medium

Info

Publication number: CN114120050A
Application number: CN202111209413.7A
Authority: CN
Inventors: 王福涛; 周艺; 王世新; 王振庆; 王丽涛; 刘文亮; 朱金峰; 侯艳芳; 赵清
Original assignee: Aerospace Information Research Institute of CAS
Current assignee: Aerospace Information Research Institute of CAS
Priority date: 2021-10-18
Filing date: 2021-10-18
Publication date: 2022-03-01

Abstract

The invention provides a method, a device, equipment and a storage medium for extracting surface ecological data, wherein the method for extracting the surface ecological data comprises the steps of firstly taking acquired hyperspectral images shot by aiming at different ground objects on the same surface area as surface data to-be-extracted hyperspectral images, and then inputting the surface data to-be-extracted hyperspectral images into a preset surface ecological data extraction network for data extraction to obtain a target surface ecological data distribution map. The preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted of the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image to obtain a target earth surface ecological data distribution map, so that the earth surface ecological data can be accurately and reliably extracted on the premise of low model complexity, the serious overfitting phenomenon caused by high model complexity is reduced, and the extraction precision of the earth surface ecological data can be greatly improved.

Description

Method, device and equipment for extracting surface ecological data and storage medium

Technical Field

The invention relates to the technical field of image processing, in particular to a method, a device, equipment and a storage medium for extracting surface ecological data.

Background

The ecological environment problem of the earth surface mainly comprises land quality degradation, sharp reduction of wetland and biological diversity, reduction of surface water quality and underground water level, protrusion of a human mole shield and the like, and influences the safety and sustainable development of the ecological environment. Therefore, how to extract the typical ecological data of the earth surface timely and accurately becomes a key problem which needs to be solved urgently at present.

The existing method extracts typical ecological data of the earth surface from the hyperspectral surface feature image through a learning classification model, for example, the data can be extracted in a mode that the features extracted from the hyperspectral surface feature image are sent into a classifier for supervised learning to generate the classification model; and data extraction can also be carried out in a mode of training a multi-feature classifier to carry out image classification by the spectral features of the hyperspectral ground object image and the high-level spatial correlation depth features extracted by the convolutional neural network.

However, the existing method mainly focuses on extracting low-dimensional features such as textures and colors for supervised learning or extracting high-dimensional features when extracting surface feature data, and cannot meet the increasing data extraction precision requirement, and a deep learning model also causes a serious overfitting phenomenon due to high model complexity.

Disclosure of Invention

The invention provides a method, a device, equipment and a storage medium for extracting ecological data of a ground surface, which are used for solving the defects of low data extraction precision and serious overfitting phenomenon in the prior art and achieving the purpose of extracting ecological data of the ground surface with high precision while greatly weakening the overfitting phenomenon.

The invention provides a method for extracting ecological data of a surface, which comprises the following steps:

acquiring hyperspectral images to be extracted of surface data, wherein the hyperspectral images to be extracted of the surface data comprise hyperspectral images aiming at different ground objects in the same surface area;

inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

According to the method for extracting the ecological data of the earth surface, provided by the invention, the preset ecological data of the earth surface comprises a preset multi-mode principal component analysis sub-network and a preset semantic segmentation sub-network, the hyperspectral image to be extracted of the earth surface data is input into the preset ecological data of the earth surface to obtain a target ecological data distribution map of the earth surface, and the method comprises the following steps:

respectively inputting the hyperspectral images to be extracted of the surface data into a preset multi-modal principal component analysis sub-network for multi-modal principal component analysis to obtain each principal component analysis image after each modal principal component analysis;

inputting each principal component analysis image into a semantic segmentation sub-network to respectively perform semantic segmentation and spatial channel joint attention correction, and respectively obtaining each predicted earth surface data distribution map corresponding to each principal component analysis image;

and obtaining a target earth surface ecological data distribution map based on each predicted earth surface data distribution map.

According to the method for extracting the ecological data of the earth surface, the preset multi-mode principal component analysis sub-network comprises a preset principal component analysis sub-network and a preset subsection principal component analysis sub-network, the hyperspectral images of the earth surface data to be extracted are respectively input into the preset multi-mode principal component analysis sub-network for multi-mode principal component analysis, and each principal component analysis image after each mode principal component analysis is obtained comprises the following steps:

respectively carrying out preset data enhancement processing on the hyperspectral images to be extracted of the surface data to obtain target enhanced images;

and inputting the target enhanced image into a preset principal component analysis sub-network and a preset subsection principal component analysis sub-network to perform principal component analysis and subsection principal component analysis respectively to obtain a first principal component analysis image and a first subsection principal component analysis image.

According to the method for extracting the ecological data of the earth surface provided by the invention, each convolution sub-network in the preset semantic segmentation sub-network is respectively connected with a spatial channel joint attention correction sub-network, each principal component analysis image is input into the semantic segmentation sub-network to be respectively subjected to semantic segmentation and spatial channel joint attention correction, and each predicted earth surface data distribution diagram corresponding to each principal component analysis image is respectively obtained, and the method comprises the following steps:

and inputting each principal component analysis image into a preset semantic segmentation sub-network for convolution, spatial channel joint attention correction, up-sampling, down-sampling and jump connection, and then respectively obtaining each predicted earth surface data distribution graph corresponding to each principal component analysis image.

According to the method for extracting the ecological data of the earth surface, provided by the invention, the training process of the preset ecological data extraction network of the earth surface comprises the following steps:

acquiring a training hyperspectral image and a verification hyperspectral image;

training a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and verifying the trained network according to the verification hyperspectral image to obtain the preset earth surface ecological data extraction network.

According to the method for extracting the ecological data of the ground surface, provided by the invention, the training of the preset initial ground surface ecological data extraction network is carried out according to the training hyperspectral image, and the verification of the trained network is carried out according to the verification hyperspectral image to obtain the preset ground surface ecological data extraction network, and the method comprises the following steps:

performing iteration training of a preset number of rounds on a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and acquiring an intermediate earth surface ecological data extraction network obtained after the iteration training of the round;

verifying the middle earth surface ecological data extraction network by using the verification hyperspectral image, and judging whether the parameter precision of the middle earth surface ecological data extraction network meets the preset precision requirement or not;

if the parameter precision meets the preset precision requirement, taking the intermediate earth surface ecological data extraction network obtained after the iterative training as the preset earth surface ecological data extraction network;

and if the parameter precision does not meet the preset precision requirement, training the middle earth surface ecological data extraction network by using the training hyperspectral image to obtain the preset earth surface ecological data extraction network.

The invention also provides a device for extracting the ecological data of the earth surface, which comprises:

the system comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring the hyperspectral images to be extracted of the earth surface data, and the hyperspectral images to be extracted of the earth surface data comprise hyperspectral images shot aiming at different ground objects in the same earth surface area;

the determining module is used for inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

The invention also provides an electronic device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the steps of the surface ecological data extraction method.

The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the surface ecological data extraction method as described in any one of the above.

The present invention also provides a computer program product comprising a computer program which, when executed by a processor, performs the steps of the method for extracting surface ecological data as described in any one of the above.

According to the method, the device and the equipment for extracting the ecological data of the earth surface and the storage medium, firstly, the acquired hyperspectral images shot aiming at different ground objects in the same earth surface area are taken as the hyperspectral images to be extracted from the earth surface data, and then the target earth surface ecological data distribution map is obtained through the process of inputting the hyperspectral images to be extracted from the earth surface data into a preset earth surface ecological data extraction network for data extraction. The preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted of the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image to obtain a target earth surface ecological data distribution map, so that the earth surface ecological data can be accurately and reliably extracted on the premise of low model complexity, the serious overfitting phenomenon caused by high model complexity is reduced, and the extraction precision of the earth surface ecological data can be greatly improved.

Drawings

In order to more clearly illustrate the technical solutions of the present invention or the prior art, the drawings needed for the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

FIG. 1 is a schematic flow chart of a surface ecological data extraction method provided by the present invention;

FIG. 2 is a schematic diagram of a training process of a pre-set earth surface ecological data extraction network provided by the present invention;

FIG. 3 is a schematic diagram of a test spectrum image, a ground truth image, and a network model prediction result according to the present invention;

FIG. 4 is a schematic structural diagram of a surface ecological data extraction device provided by the present invention;

fig. 5 is a schematic structural diagram of an electronic device provided in the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The problem of the earth surface ecological environment is a big important problem generally faced by all mankind, and the change of the spatial distribution and the time sequence of different earth surface ecological data of a region can provide powerful support for analyzing and solving the problem of the earth surface ecological environment of the region. The extraction of typical ecological data (buildings, vegetation, water bodies, bare soil and the like) of the earth surface not only has great auxiliary effects on enhancing ecological environment protection and realizing sustainable development of areas, but also has important application prospects in ecological environment evaluation research, ecological function utilization and development.

At present, a hyperspectral remote sensing image has high spectral resolution and can be used for distinguishing fine spectral features of different ground objects. Due to this advantage, hyperspectral remote sensing has been successfully applied in many fields, such as ground feature classification (Kang et al, 2014; Lu et al, 2016), target detection (Nasrabadi, 2014; Liu et al, 2017), and spectral unmixing (Kizel et al, 2017). In these applications, hyperspectral imagery ground object classification is of great interest due to its importance in ecological environment assessment and monitoring.

In addition, the high-dimensional features of hyperspectral data and the houss phenomenon caused by more limited training samples have been an important challenge in hyperspectral image classification. In order to improve the accuracy of classification of the feature in the spectral image, feature extraction or feature selection is generally performed on the hyperspectral image before classification. After the features are obtained, the features need to be sent to a classifier for supervised learning, and a final classification model is obtained. Benediktsson et al (2005) extract a plurality of principal components from hyperspectral data using Principal Component Analysis (PCA), construct a Morphological Profile (MP) for each principal component and use it together for an extended morphological profile, and finally train and classify it using neural networks. Licciardi et al (2012) performed classification experiments using extended MP constructed from features after dimensionality reduction of nonlinear principal component analysis (NLPCA) implemented by a self-associative neural network, concluded that NLPCA has better classification accuracy than linear PCA. In addition, other scholars have achieved good classification effects by learning using machine learning models such as Support Vector Machines (SVMs) using dimension reduction methods such as Independent Component Analysis (ICA) and Singular Spectral Analysis (SSA) (Villa et al, 2011; Zabalza et al, 2014).

With the rapid development of deep learning, numerous models represented by convolutional neural networks begin to be elegant and colorful on the task of hyperspectral terrain classification. Zhao and Du (2016) propose a classification framework based on spectrum-space features, firstly, a balanced local discrimination embedding algorithm is used for extracting spectrum features, secondly, a convolutional neural network is used for automatically extracting high-level space correlation depth features, and finally, a classifier based on multiple features is trained for image classification. Liu et al (2017) developed an effective classification framework combining Deep Belief Networks (DBNs) with active learning, designed a weighted incremental dictionary learning algorithm, actively selected additional representative samples from unlabeled datasets, and DBNs provided the final predictions. GaoQuliang (2021) directly inputs high-spectrum high-dimensional data into a capsule network, a depth network model is constructed by combining a rolling capsule layer and a residual structure, and finally, a three-dimensional rolling capsule layer is introduced to fully utilize empty spectrum joint information in an image so as to further improve classification precision.

However, machine learning classification models, while behaving somewhat infrequently, have difficulty meeting the ever-increasing practical demands for accuracy due to their limitations. Although the deep learning model improves the precision of data in the paper, relatively serious overfitting phenomenon still exists when less sample data sets are faced.

In view of the above problems, the present invention provides a method for extracting surface ecological data, where an execution main body of the method for extracting surface ecological data may be a device for extracting surface ecological data, and the device for extracting surface ecological data may be implemented as part or all of a terminal device by software, hardware, or a combination of software and hardware. Alternatively, the terminal device may be a Personal Computer (PC), a portable device, a notebook Computer, a smart phone, a tablet Computer, a portable wearable device, and other electronic devices, such as a tablet Computer, a mobile phone, and the like. The present invention does not limit the specific form of the terminal device.

It should be noted that the execution subject of the method embodiments described below may be part or all of the terminal device described above. The following method embodiments take the execution subject as an example of the terminal device.

Fig. 1 is a schematic flow chart of a surface ecological data extraction method provided by the present invention, as shown in fig. 1, the surface ecological data extraction method includes the following steps:

step S110, obtaining surface data to-be-extracted hyperspectral images, wherein the surface data to-be-extracted hyperspectral images comprise hyperspectral images aiming at different ground objects in the same surface area.

Specifically, the terminal device can receive a ground surface data extraction instruction sent by the client, acquire a plurality of ground surface data to-be-extracted hyperspectral images according to the ground surface data extraction instruction, can select a plurality of hyperspectral images aiming at different ground objects in the same region from a prestored hyperspectral image set to serve as the ground surface data to-be-extracted hyperspectral images, and can also start the image acquisition device to shoot a plurality of hyperspectral images aiming at different ground objects in the same ground surface area to serve as the ground surface data to-be-extracted hyperspectral images. And is not particularly limited herein.

And S120, inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map.

The preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

Specifically, when the terminal device acquires the hyperspectral image to be extracted from the surface data, according to a feature extraction algorithm and a feature fusion processing algorithm of a pre-device, the hyperspectral image to be extracted from the surface data is subjected to multi-modal principal component analysis processing to obtain different feature images of different principal component features, and then the different feature images are subjected to semantic segmentation processing and attention mechanism correction processing respectively to perform feature fusion of different scales on the different principal component features, so that a target surface ecological data distribution map corresponding to the hyperspectral image to be extracted from the surface data is obtained

It should be noted that, in order to improve the extraction accuracy and the extraction precision, the feature extraction algorithm used in the present invention includes an algorithm combining PCA and segmented PCA, because the PCA converts hyperspectral band data represented by linear correlation variables into a small amount of "band" data represented by linear independent variables by using orthogonal transformation, and the factors of mutual influence among the components of the original data are eliminated. Although the data dimension reduction is realized, the physical meaning of the spectral band of the image is also changed, so that the meaning of each principal component has certain ambiguity. In addition, non-principal components with small variance may also contain important information beneficial to the classification of the surface features, and since the dimension reduction and discarding may have an influence on the subsequent data processing, the segmented PCA processing is added, wherein the segmented PCA continuously divides the original bands into several groups, and then performs principal component transformation on each group respectively, so as to avoid that local important bands are omitted in the selection. Therefore, the network classification precision can be improved when the features obtained by PCA and the features obtained by segmented PCA are classified and fused. Optionally, the feature fusion processing algorithm may be an Unet + + segmentation network, and each convolution sub-network in the Unet + + segmentation network is connected to one attention mechanism correction sub-network.

Optionally, when the preset surface ecological data extraction network includes a preset multi-modal principal component analysis sub-network and a preset semantic segmentation sub-network, step S120 may be implemented by the following processes:

firstly, respectively inputting the hyperspectral images to be extracted of the surface data into a preset multi-modal principal component analysis sub-network to perform multi-modal principal component analysis, and obtaining each principal component analysis image after each modal principal component analysis; then inputting each principal component analysis image into a semantic segmentation sub-network to respectively perform semantic segmentation and spatial channel joint attention correction, and respectively obtaining each predicted surface data distribution map corresponding to each principal component analysis image; and finally, obtaining a target earth surface ecological data distribution map based on each predicted earth surface data distribution map.

Specifically, when the terminal device obtains the surface data to-be-extracted hyperspectral image, the surface data to-be-extracted hyperspectral image may be input into a preset multi-modal principal component analysis sub-network to perform principal component analysis of multiple modalities, that is, the preset multi-modal principal component analysis sub-network may perform principal component analysis processing of different modalities on the surface data to-be-extracted hyperspectral image, for example, perform PCA and segmented PCA on the surface data to-be-extracted hyperspectral image, so as to obtain each principal component analysis image after principal component analysis of each modality, and each principal component analysis image contains different component feature information with surface data attributes.

Furthermore, when the terminal device determines that the preset multi-modal principal component analysis sub-network outputs each principal component analysis image after each modal principal component analysis, each principal component analysis image can be further input into the preset semantic segmentation sub-network to perform feature fusion processing of different levels and different scales, so that the typical earth surface ecological data with obvious features of each category and the typical earth surface ecological data with unobvious features can be accurately and precisely captured, and each predicted earth surface data distribution map is obtained; each predicted surface data distribution map comprises specific category information corresponding to each pixel and the probability thereof.

And finally, performing fusion addition processing on each predicted surface data distribution map, for example, each predicted surface data distribution map comprises a first predicted surface data distribution map and a second predicted surface data distribution map, the size of each predicted surface data distribution map is the same, the first pixel in the first predicted surface data distribution map is marked as a water body, the probability of each first pixel is 60%, the second pixel in the first predicted surface data distribution map is marked as a non-water body, the probability of each second pixel is 40%, the first pixel in the second predicted surface data distribution map is marked as a water body, the probability of each second pixel in the second predicted surface data distribution map is 20%, the second pixel in the second predicted surface data distribution map is marked as a non-water body, the probability of each second pixel in the target surface ecological data distribution map obtained by performing fusion addition processing on the first predicted surface data distribution map and the second predicted surface data distribution map is marked as a water body, the probability of each first pixel is 80%, and the probability of each second pixel is 120%. Therefore, a target earth surface ecological data distribution diagram with richer and more comprehensive earth surface typical ecological data is obtained.

According to the method for extracting the ecological data on the earth surface, the preset multi-mode principal component analysis sub-network with different modal principal component analysis processing is used for extracting different component characteristic information with earth surface data attributes, and the preset semantic segmentation sub-network with different levels and different scale characteristic fusion functions is used for fusing the different component characteristic information into the target earth surface ecological data distribution diagram with rich earth surface data types and complete earth surface data, so that the extraction accuracy and the extraction reliability of the ecological data on the earth surface are improved.

Optionally, the preset multi-modal principal component analysis sub-network includes a preset principal component analysis sub-network and a preset segmented principal component analysis sub-network, and the method includes the following steps of respectively inputting the hyperspectral images of the surface data to be extracted into the preset multi-modal principal component analysis sub-network for multi-modal principal component analysis to obtain each principal component analysis image after each modal principal component analysis:

respectively carrying out preset data enhancement processing on the hyperspectral images to be extracted of the surface data to obtain target enhanced images; and inputting the target enhanced image into a preset principal component analysis sub-network and a preset subsection principal component analysis sub-network to perform principal component analysis and subsection principal component analysis respectively to obtain a first principal component analysis image and a first subsection principal component analysis image.

The preset data enhancement processing comprises channel enhancement and data enhancement.

Specifically, the terminal device performs preset data enhancement processing on the surface data to-be-extracted hyperspectral image, namely, the channel enhancement purpose is realized by selecting any wave band in the surface data to-be-extracted hyperspectral image to perform rejection processing, and then the data enhancement purpose is realized by performing random horizontal mirror image, vertical mirror image and diagonal data enhancement on the rejected surface data to-be-extracted hyperspectral image, so that a target enhanced image is obtained.

Then, performing PCA and segmented PCA on the target enhanced image, wherein the process of performing PCA on the target enhanced image comprises the following steps:

for a target enhanced image X comprising n pixels and m wave bands, firstly, a covariance matrix C of the target enhanced image X is calculated, then n eigenvalues of the covariance matrix C and eigenvectors corresponding to each eigenvalue are solved, the n eigenvectors are arranged from large to small according to the magnitude relation between the corresponding eigenvalues and then are arranged into a first matrix according to rows, the first k rows of the first matrix are selected to form a second matrix P, the second matrix P comprises k eigenvectors corresponding to large eigenvalues, and then the result of multiplying the second matrix P and the target enhanced image X is used as a first principal component analysis image obtained after PCA processing. The calculation formula of the covariance matrix C is as follows:

wherein x is_iI-th pixel, i-1, 2, … n, X representing the target enhanced image X_jRepresents the jth band of the target enhanced picture X, the superscript T represents the transpose operation, Σ () represents the sum operation, j is 1, 2, … m.

The process of segmented PCA on the target enhanced image X comprises the following steps:

for a target enhanced image X comprising n pixels and m bands, it is first set that each of the m bands comprises n ' pixels, i ' is formed by [1, n ']Then, calculating the correlation coefficient r of Pearson's product moment between every two adjacent wave bands in the target enhanced image X_j，j+1The calculation formula is as follows:

wherein r is_j，j+1Representing Pearson product moment correlation coefficients between a jth wave band and a j +1 wave band, wherein j and j +1 represent two adjacent wave bands in m wave bands, and the initial value of j is 1; each of the m bands includes n ' pixels, i ' e [1, n ']；A_ji′Representing the pixel value of the i' th pixel in the j-th band,

denotes the mean value of the j-th band, B_(j+1)i′Represents the pixel value of the i' th pixel element in the j +1 th band,

represents the mean of the j +1 th band, j ∈ [1, m [ ]]。

Further, judging the correlation coefficient r of the Pearson product moment between every two adjacent wave bands_j，j+1Whether it is greater than a preset coefficient threshold value T, when determining r_j，j+1When the number of the wavebands is more than T, storing the jth waveband in a segmented index linked list, adding 1 to the value of j, continuously calculating the Pearson product moment correlation coefficient between the next two adjacent wavebands and judging the Pearson product moment correlation coefficient and the preset coefficient threshold value T until m wavebands are traversed, and dividing the target enhanced image X into a hyperspectral image F containing D groups of wavebands by using each waveband stored in the segmented index linked list; for example, when the segment index linked list stores the 1 st band, the 3 rd band, the 7 th band, and the 11 th band, the hyperspectral image may be divided into 4 groups, and at this time, the hyperspectral image F may be considered to include 4 groups of bands, i.e., the 1 st band, the 3 rd band, the 6 th band, the 7 th band, the 10 th band, and the 11 th band.

Then, further performing PCA processing on each group of wave bands in the hyperspectral image F, wherein the PCA processing process is the same as the above process, that is, firstly, feature vectors of covariance matrices of each group of wave bands are sorted from large to small according to magnitude relations between corresponding feature values, then, the first matrix formed by arranging rows is arranged, then, the first k rows in the formed first matrix are respectively formed into a second matrix, the second matrix comprises k feature vectors corresponding to large feature values, then, the result of multiplying each second matrix with the corresponding group of wave bands is obtained to obtain D PCA processing results, finally, the D PCA processing results are subjected to superposition processing (such as channel fusion processing), and the result after the superposition processing is used as a first segmented principal component analysis image obtained after the segmented PCA processing.

It should be noted that, because the hyperspectral images to be extracted from the surface data are usually obtained for different ground features in a certain surface area with the surface ecological environment problem, most of the surfaces with the surface ecological environment problem have the appearance of land quality degradation, sharp reduction of wetland and biological diversity, surface water quality and groundwater level reduction and human spear shield protrusion, the hyperspectral images to be extracted from the obtained surface data need to be enhanced, so that the defect of network overfitting caused by less data is avoided; and when the fuzzy wave band of the hyperspectral image to be extracted of the earth surface data is determined, fuzzy wave band elimination processing is also carried out. Therefore, the detection precision of the hyperspectral image to be extracted from the earth surface data with the earth surface ecological environment problem is improved.

In addition, if the hyperspectral image of the surface data to be extracted contains b fuzzy wave bands, the mth wave band is the last wave band in the remaining wave bands after the b fuzzy wave bands in the hyperspectral image of the surface data to be extracted and the channel enhancement treatment are removed; on the contrary, if the hyperspectral image to be extracted does not contain the fuzzy wave band, the mth wave band is the last wave band in the rest wave bands of the hyperspectral image to be extracted of the earth surface data after the channel enhancement processing is carried out.

According to the method for extracting the ecological data of the earth surface, provided by the invention, the aim of improving the classification balance of the earth surface data while avoiding the phenomenon of network overfitting is fulfilled by performing the preset enhancement treatment on the hyperspectral image to be extracted of the earth surface data, and then performing the principal component analysis and the segmented principal component analysis treatment, so that a reliable basis is provided for obtaining a high-precision target earth surface ecological data distribution map subsequently.

Optionally, each convolution sub-network in the preset semantic segmentation sub-network is respectively connected to a spatial channel joint attention correction sub-network, and the input of each principal component analysis image into the semantic segmentation sub-network is respectively subjected to semantic segmentation and spatial channel joint attention correction to respectively obtain each predicted surface data distribution map corresponding to each principal component analysis image, including:

Specifically, the preset semantic segmentation sub-networks include a preset Unet + + sub-network and a preset scSE sub-network, the preset Unet + + sub-network is a network obtained after an initial Unet + + sub-network is trained, the preset scSE sub-network is a network obtained after the initial scSE sub-network is trained, the Unet + + sub-network retains each convolution sub-network, each down-sampling sub-network, each up-sampling sub-network and each hop connection sub-network, and also includes a plurality of scSE sub-networks, each convolution sub-network is connected with one scSE sub-network, and the number of the convolution sub-networks is the same as that of the scSE sub-networks.

Therefore, when each principal component analysis image is input into the preset semantic segmentation sub-network, the preset convolution sub-network performs convolution calculation on the principal component analysis image to obtain a feature map after convolution, the scSE sub-network performs spatial correction and channel correction on the feature map output by the corresponding preset convolution sub-network to obtain a corrected feature map, the preset down-sampling sub-network performs pooling operation on the input feature map to reduce the size of the feature map to obtain a feature map after size reduction, the preset up-sampling sub-network performs size doubling operation on the input feature map to obtain a feature map after size doubling, the preset skip connection sub-network fuses the feature maps with different sizes, and finally, after spatial and channel correction is performed on the last preset scSE sub-network in the preset semantic segmentation sub-network, each predicted surface data distribution map corresponding to each principal component analysis image is output. For example, when each principal component analysis image includes a first principal component analysis image and a first segmented principal component analysis image, each of the corresponding predicted surface data distribution maps that can be obtained includes a first principal component predicted surface data distribution map and a first segmented predicted surface data distribution map, respectively.

It should be noted that the preset scSE sub-network performs spatial and channel correction on the feature map output by the preset convolution sub-network; presetting a jump connection sub-network to fuse the feature maps with different scales to obtain fused feature maps, wherein the scales of the feature maps obtained after each fusion are respectively different; the feature maps with different scales can be obtained after multiple times of upsampling, multiple times of downsampling, multiple times of fusion or multiple times of space and channel correction, the depth of the scales is increased along with the increase of the times of processing operation as the number of times of upsampling is increased, and the depth of the current feature map is larger than that of the previous feature map. For example, the scale of the feature map obtained after the 2 nd convolution operation is deeper than the scale of the feature map obtained after the 1 st convolution operation, and the scale of the feature map obtained after the 4 th fusion is deeper than the scale of the feature map obtained after the 3 rd fusion.

According to the method for extracting the ecological data on the earth surface, provided by the invention, the preset semantic segmentation sub-network is used for respectively carrying out convolution, spatial channel joint attention correction, up-sampling, down-sampling and jump connection on each principal component analysis image, so that the purpose of carrying out fusion processing on different scales and different depths on different characteristic information extracted by multi-mode principal component analysis is realized, and the extraction precision and the category richness of the ecological data on the earth surface are effectively improved.

Optionally, the training process of the preset surface ecological data extraction network includes:

firstly, acquiring a training hyperspectral image and a verification hyperspectral image; and then training a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and verifying the trained network according to the verification hyperspectral image to obtain the preset earth surface ecological data extraction network.

Specifically, M hyperspectral images of different ground objects on the same ground surface area are obtained in advance, each hyperspectral image comprises K channels, each hyperspectral image comprises an original remote sensing image and a label remote sensing image y, each pixel point in the label remote sensing image y is correspondingly marked with a category value, for example, when a certain pixel in the label remote sensing image y is a tree and the preset category value of a vegetation is 1, the pixel can be marked with 1; dividing the M hyperspectral images into a training image set (for example, M1 takes a value of 72) composed of M1 training hyperspectral images and a verification image set (for example, M2 can take a value of 18) composed of M2 verification hyperspectral images according to a preset proportion (preferably 4: 1).

Further, the M1 training hyperspectral images are used for training a preset initial earth surface ecological data extraction network, wherein the preset initial earth surface ecological data extraction network comprises an initial multi-modal principal component analysis sub-network and an initial semantic segmentation sub-network, the initial semantic segmentation sub-network comprises an initial Unet + + sub-network and an initial scSE sub-network, the initial multi-modal principal component analysis sub-network comprises an initial principal component analysis sub-network and an initial segmentation principal component analysis sub-network, and when the preset initial earth surface ecological data extraction network is trained to be convergent, the preset earth surface ecological data extraction network comprising the preset multi-modal principal component analysis sub-network and the preset semantic segmentation sub-network can be obtained.

According to the method for extracting the ecological data of the earth surface, the purpose of quickly obtaining the preset ecological data extraction network of the earth surface is achieved by using the mode of training the initial ecological data extraction network of the earth surface by using the hyperspectral image and verifying the trained network by using the verification hyperspectral image, the purpose of distinguishing fine spectrum characteristics of different ground objects can be achieved by using the hyperspectral resolution of the hyperspectral image, the purposes of reducing the complexity of network training and selecting a proper training sample can be achieved based on the mode of combining training and verification, the overfitting phenomenon is greatly weakened, the training speed and precision of the network are improved, and a foundation is laid for quickly and accurately extracting the ecological data of the earth surface subsequently.

Optionally, the training is performed on the preset initial earth surface ecological data extraction network according to the training hyperspectral image, and the training is verified on the network according to the verification hyperspectral image, so as to obtain the preset earth surface ecological data extraction network, including:

firstly, carrying out iteration training of a preset number of rounds on a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and obtaining an intermediate earth surface ecological data extraction network obtained after the iteration training of the round; then, the verification hyperspectral image is used for verifying the middle earth surface ecological data extraction network, and whether the parameter precision of the middle earth surface ecological data extraction network meets the preset precision requirement or not is judged; if the parameter precision meets the preset precision requirement, taking the intermediate earth surface ecological data extraction network obtained after the iterative training as the preset earth surface ecological data extraction network; and if the parameter precision does not meet the preset precision requirement, training the middle earth surface ecological data extraction network by using the training hyperspectral image to obtain the preset earth surface ecological data extraction network.

Specifically, the terminal device trains a preset initial earth surface ecological data extraction network by using M1 training hyperspectral images, and after the M1 training hyperspectral images train the initial earth surface ecological data extraction network once, the training is called as one round of training. And the verification process can be executed after multiple rounds of initial earth surface ecological data extraction network training, and can also be executed after one round of training. Preferably, verification is performed after a round of training,

the terminal equipment obtains an intermediate network after each round of training by using M1 training hyperspectral images, verifies the intermediate network after the round of training by further using M2 verification hyperspectral images to judge whether the network parameter precision of the intermediate network after the round of training meets the preset precision requirement, stops training when the network parameter precision of the intermediate network after the round of training meets the preset precision requirement, and determines the corresponding intermediate network as a preset earth surface ecological data extraction network when the training stops; on the contrary, if the intermediate network after the training of the current round does not meet the preset precision requirement, the number of rounds of the intermediate network after the training of the current round is executed by using the M1 training hyperspectral images again, and the preset earth surface ecological data extraction network is obtained until the network parameter precision of the network after the M2 verification hyperspectral images verification training meets the preset precision requirement.

In addition, to every training hyperspectral image, every training hyperspectral image all participates in a training in every round of training, and can all calculate a loss function after every training hyperspectral image participates in a training, and every training hyperspectral image participates in the process of a training and corresponds the computational process of loss function, include:

as shown in fig. 2, the training hyperspectral image is subjected to PCA processing and segmented PCA processing respectively to obtain a first principal component analysis training image after PCA processing and a first segmented principal component analysis training image after segmented PCA, the first principal component analysis training image and the first segmented principal component analysis training image are input into an initial semantic segmentation subnetwork respectively, and then a first principal component prediction surface data training distribution map corresponding to the first principal component analysis training image and a first segmented prediction surface data training distribution map corresponding to the first segmented principal component analysis training image are output respectively.

Before PCA and segmented PCA are carried out on training hyperspectral images, fuzzy wave band elimination processing is carried out on the training hyperspectral images, then any channel in the training hyperspectral images after noise wave bands are eliminated is eliminated, offline enhanced training images are obtained, and random horizontal mirror images, vertical mirror images and diagonal angle data enhancement are carried out on the offline enhanced training images, so that target enhanced training images are obtained. The method comprises the steps of training hyperspectral images, removing the hyperspectral images, performing enhancement processing, and performing enhancement processing.

Then, the process of conducting PCA on the target enhanced training image comprises the following steps: for a target enhanced training image Y comprising n pixels and m wave bands, calculating a training covariance matrix of the target enhanced training image Y, then solving n eigenvalues of the training covariance matrix and eigenvectors corresponding to each eigenvalue, arranging the n eigenvectors from large to small according to the corresponding eigenvalue size and arranging the n eigenvectors into a first training matrix according to rows, taking the first k rows of the first training matrix to form a second training matrix, wherein the second training matrix comprises k eigenvectors corresponding to large eigenvalues, and then taking the result of multiplying the second training matrix by the target enhanced training image Y as a first principal component analysis training image after PCA processing.

The process of performing segmented PCA on the target-enhanced training image Y includes: for a target enhanced training image X comprising n pixels, m bands, each of the m bands comprising n ' pixels, i ' being [1, n ']Calculating the correlation coefficient r of Pearson's product moment between every two adjacent bands in the target enhanced training image Y_j',j'+1And further determining when r_j',j'+1If the wave band is larger than T, the jth wave band is stored in the segmented index linked list, then the value of j 'is added with 1, the calculation of the correlation coefficient of the Pearson product moment between the next two adjacent wave bands and the judgment of the size of the correlation coefficient and the threshold value T are calculated until m wave bands are traversed, and j' belongs to [1, m]And at the moment, dividing the target enhanced training image Y into hyperspectral training images comprising D groups of wave bands by using each wave band stored in the segmented index linked list, and then performing PCA (principal component analysis) processing on each group of wave bands in the hyperspectral training images respectively, wherein the PCA processing process is the same as the process, so that D PCA processing training results are obtained, and the D PCA processing training results are subjected to superposition processing (such as channel fusion processing), and the result after the superposition processing can be used as a first segmented principal component analysis training image after the segmented PCA. The PCA processing training results of each group of wave bands in the hyperspectral training image for PCA processing all comprise one wave band, and the superimposed first-segment principal component analysis training image comprises D wave bands.

The process of respectively inputting the first principal component analysis training image and the first subsection principal component analysis training image into an initial semantic segmentation sub-network for semantic segmentation processing comprises the following steps:

as shown in fig. 2, the triangular mesh structure of the middle portion is an initial semantic segmentation sub-network, each upward arrow represents an initial up-sampling sub-network, each downward arrow represents an initial down-sampling sub-network, each dotted line represents an initial jump connection sub-network, each circle represents an initial convolution sub-network, each circle is followed by a small connection box representing an initial scSE sub-network, so that the first principal component analysis training image and the first segmented principal component analysis training image enter the initial semantic segmentation sub-network and are subjected to convolution, down-sampling, spatial channel joint attention correction, up-sampling and jump connection processing, and finally, after spatial channel joint attention correction is performed through the last initial scSE sub-network, the first principal component prediction surface data training profile corresponding to the first principal component analysis training image and the first segmented principal component analysis training profile corresponding to the first principal component analysis training image are output respectively And (5) training a distribution diagram by sectionally predicting the surface data.

Further, the first principal component predicted surface data training distribution map and the first subsection predicted surface data training distribution map respectively include specific category information and probability corresponding to each pixel, for example, the first pixel in the first principal component predicted surface data training distribution map is marked with bare soil with a probability of 40%, the second pixel is marked with non-bare soil with a probability of 70%, the first pixel in the first subsection predicted surface data training distribution map is marked with bare soil with a probability of 50%, the second pixel is marked with non-bare soil with a probability of 80%, the first principal component predicted surface data training distribution map and the first subsection predicted surface data training distribution map are subjected to fusion and addition processing to obtain a trained surface ecological data distribution map P obtained after the training of the trained hyperspectral image is performed, the first pixel in the trained surface ecological data distribution map P is marked with bare soil with a probability of 90% >, and the first subsection surface ecological data training distribution map is subjected to fusion and addition processing, The second pixel is marked as non-bare soil and the probability is 150%.

Then, calculating a Loss function Loss of the training ground surface ecological data distribution diagram P by using the label remote sensing image y_ceThe formula is as follows:

wherein N represents the number of pixels respectively contained in the label remote sensing image y and the training earth surface ecological data distribution map P, K represents the number of preset categories (for example, when the categories include buildings, vegetation, water and bare soil, K takes the value of 4),

a c-th value in the one-hot code representing the actual category of the ith pixel in the training ground surface ecological data distribution map P (for example, when the actual category of the ith pixel is vegetation and is 1, the 1 st bit in the one-hot code is 1, and the rest bits are 0, at this time, c is 1 and indicates vegetation);

and the predicted probability value of the category c at the ith' pixel in the label remote sensing image y (such as the predicted probability value of the vegetation class when c is 1) is represented. Optionally, a joint function Lossseg of soft cross entropy and Dice may also be used as the loss function:

wherein, TP is a positive class classified accurately, FP is a negative class misclassified as a positive class, and FN is a positive class misclassified as a negative class.

Based on the process that the training hyperspectral images participate in one training and the calculation process of the corresponding loss function, the training process of each training hyperspectral image and the calculation operation of the loss function can be executed, and the corresponding parameters of the network after the training are adaptively adjusted according to the value of the loss function calculated each time, so that the number of training rounds is as small as possible, and the accuracy of network training is higher.

According to the method for extracting the ecological data of the earth surface, the aim of quickly and efficiently obtaining the preset ecological data of the earth surface is fulfilled by training the initial ecological data extraction network preset round number training preset by the hyperspectral image training and using the verification hyperspectral image to carry out precision verification on the intermediate ecological data extraction network of the earth surface obtained after the iterative training, the network precision is ensured, the training complexity is greatly reduced, and the network training speed is accelerated

In the actual processing process, in order to verify the effectiveness of the method, the training image set obtained by dividing the M hyperspectral images according to the proportion of 4:1 comprises 72 training hyperspectral images. And through the comparison and observation of the wave spectrum, the hyperspectral image comprises 32 wave bands, the 1 st wave band and the 32 th wave band have relatively obvious noise relative to other wave bands, and the existence of the noise is not beneficial to network benign learning, so the invention adopts the measure of eliminating the two wave bands. And then, one channel is selected from the remaining 30 channels of the hyperspectral image subjected to band rejection operation for rejection, and then the remaining 29 channels are subjected to multi-mode principal component analysis and transformation.

In order to alleviate the class imbalance, a multi-modal principal component analysis transform with a probability of 1 may be performed on an image with a bare soil proportion of > 1% and a background proportion of < 50%, and a multi-modal principal component analysis transform with a probability of 1/2 may be performed on the other images, thereby obtaining about 72+5 × 30+67 × 15 — 1092 images. In addition, in the iterative training of the network, an AdamW optimization algorithm with an initial learning rate of 0.001 and weight attenuation of 0.001 is used for backward propagation iteration for 100 times, parameters in the network are optimized layer by layer, and network parameters under the optimal round number of intersection comparison IoU of the verification image sets are stored.

In order to quantitatively describe the ecological data extraction precision of the model, the method adopts three indexes of a Kappa coefficient, an intersection-to-parallel ratio IoU and an average intersection-to-parallel ratio mIoU to evaluate the result performance, wherein the Kappa coefficient represents the error reduction ratio of the extraction result obtained by the network compared with the result obtained by completely randomly classifying, IoU represents the intersection-to-parallel ratio of the actual normal sample and the predicted extracted normal sample, and the mIoU is the average value of IoU of each class, and the calculation formula of each index is as follows:

wherein, P_oRepresenting correctly classified pels divided by the total number of pels, P_eThe quotient of the sum of the product of the actual and predicted sample numbers of each class and the square of the total pixel number is represented, TP represents a positive class which is classified accurately, FN represents a positive class which is misclassified as a negative class, FP represents a negative class which is misclassified as a positive class, and K represents the preset number of classes.

The accuracy of the evaluation of the results of the surface ecological data by using the invention is shown in table 1, wherein IoU in the four ecological data is the highest water body and the water body is as high as 0.833, vegetation IoU of 0.827 is followed by 0.781 buildings IoU, bare soil IoU of 0.507 is ranked finally, the network total Kappa value is 0.822, and the mIoU is 0.737. Fig. 3 is a schematic diagram of a network model prediction result of the test spectrum image, the corresponding ground real value image and the network ecological element extraction result, in fig. 3, (a) is a schematic diagram of a list of test spectrum images, (b) is a schematic diagram of a list of ground real value images, and (c) is a schematic diagram of a list of network model prediction results.

TABLE 1

Through precision comparison analysis, the extraction precision of the surface ecological data by using the preset surface ecological data extraction network in the method is high, and the effect is better.

The surface ecological data extracting device provided by the invention is described below, and the surface ecological data extracting device described below and the surface ecological data extracting method described above can be correspondingly referred to.

Fig. 4 illustrates a surface ecological data extracting apparatus, as shown in fig. 4, the surface ecological data extracting apparatus 400 includes: the acquisition module 410 is configured to acquire hyperspectral images to be extracted of earth surface data, where the hyperspectral images to be extracted of earth surface data include hyperspectral images captured for different ground objects in the same earth surface area; the determining module 420 is configured to input the hyperspectral image to be extracted from the surface data into a preset surface ecological data extraction network, so as to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

Optionally, the determining module 420 may be specifically configured to input the to-be-extracted hyperspectral images of the surface data into a preset multi-modal principal component analysis sub-network respectively to perform multi-modal principal component analysis, so as to obtain each principal component analysis image after each modal principal component analysis; inputting each principal component analysis image into a semantic segmentation sub-network to respectively perform semantic segmentation and spatial channel joint attention correction, and respectively obtaining each predicted earth surface data distribution map corresponding to each principal component analysis image; and obtaining a target earth surface ecological data distribution map based on each predicted earth surface data distribution map.

Optionally, the determining module 420 may be further configured to perform preset data enhancement processing on the to-be-extracted hyperspectral images of the surface data, respectively, to obtain target enhanced images; and inputting the target enhanced image into a preset principal component analysis sub-network and a preset subsection principal component analysis sub-network to perform principal component analysis and subsection principal component analysis respectively to obtain a first principal component analysis image and a first subsection principal component analysis image.

Optionally, the determining module 420 may be further configured to input each principal component analysis image into a preset semantic segmentation sub-network to perform convolution, spatial channel joint attention correction, upsampling, downsampling, and skip connection, and then obtain each predicted surface data distribution map corresponding to each principal component analysis image.

Optionally, the obtaining module 410 may be further configured to obtain a training hyperspectral image and a verification hyperspectral image.

Optionally, the determining module 420 may be further configured to train a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and verify the trained network according to the verification hyperspectral image to obtain the preset earth surface ecological data extraction network.

Optionally, the determining module 420 may be further specifically configured to perform iteration training of a preset number of rounds on a preset initial earth surface ecological data extraction network according to the training hyperspectral image, and obtain an intermediate earth surface ecological data extraction network obtained after the iteration training of the round; verifying the middle earth surface ecological data extraction network by using the verification hyperspectral image, and judging whether the parameter precision of the middle earth surface ecological data extraction network meets the preset precision requirement or not; if the parameter precision meets the preset precision requirement, taking the intermediate earth surface ecological data extraction network obtained after the iterative training as the preset earth surface ecological data extraction network; and if the parameter precision does not meet the preset precision requirement, training the middle earth surface ecological data extraction network by using the training hyperspectral image to obtain the preset earth surface ecological data extraction network.

Fig. 5 illustrates a physical structure diagram of an electronic device, and as shown in fig. 5, the electronic device 500 may include: a processor (processor)510, a communication Interface (Communications Interface)520, a memory (memory)530 and a communication bus 540, wherein the processor 510, the communication Interface 520 and the memory 530 communicate with each other via the communication bus 540. Processor 510 may invoke logic instructions in memory 530 to perform a surface ecology data extraction method comprising: acquiring hyperspectral images to be extracted of surface data, wherein the hyperspectral images to be extracted of the surface data comprise hyperspectral images aiming at different ground objects in the same surface area; inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

Furthermore, the logic instructions in the memory 530 may be implemented in the form of software functional units and stored in a computer readable storage medium when the software functional units are sold or used as independent products. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

In another aspect, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, a computer is capable of executing the surface ecological data extraction method provided by the above methods, the method including: acquiring hyperspectral images to be extracted of surface data, wherein the hyperspectral images to be extracted of the surface data comprise hyperspectral images aiming at different ground objects in the same surface area; inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which when executed by a processor, implements a method for surface ecological data extraction provided by the above methods, the method comprising: acquiring hyperspectral images to be extracted of surface data, wherein the hyperspectral images to be extracted of the surface data comprise hyperspectral images aiming at different ground objects in the same surface area; inputting the hyperspectral image to be extracted of the surface data into a preset surface ecological data extraction network to obtain a target surface ecological data distribution map; the preset earth surface ecological data extraction network is used for performing multi-mode principal component analysis processing on the hyperspectral image to be extracted from the earth surface data, performing semantic segmentation and attention mechanism correction on the processed hyperspectral image, and then obtaining the target earth surface ecological data distribution map.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting surface ecological data is characterized by comprising the following steps:

2. The method for extracting ecological data on earth's surface according to claim 1, wherein the preset ecological data on earth's surface extraction network comprises a preset multi-modal principal component analysis sub-network and a preset semantic segmentation sub-network, and the step of inputting the hyperspectral image to be extracted from the ecological data on earth's surface into the preset ecological data on earth's surface extraction network to obtain a target ecological data on earth's surface distribution map comprises the following steps:

3. The method for extracting surface ecological data according to claim 2, wherein the preset multi-modal principal component analysis sub-network comprises a preset principal component analysis sub-network and a preset subsection principal component analysis sub-network, and the step of inputting the hyperspectral images of the surface data to be extracted into the preset multi-modal principal component analysis sub-network respectively to perform multi-modal principal component analysis to obtain each principal component analysis image after each modal principal component analysis comprises the steps of:

4. The method according to claim 2, wherein each convolution sub-network in the preset semantic segmentation sub-network is respectively connected to a spatial channel joint attention correction sub-network, and the step of inputting each principal component analysis image into the semantic segmentation sub-network for semantic segmentation and spatial channel joint attention correction respectively to obtain each predicted surface data distribution map corresponding to each principal component analysis image comprises:

5. The method for extracting earth surface ecological data according to claims 1 to 4, wherein the training process of the preset earth surface ecological data extraction network comprises:

6. The method for extracting ecological data on earth's surface according to claim 5, wherein the training a preset initial network for extracting ecological data on earth's surface according to the training hyperspectral image and verifying the trained network according to the verification hyperspectral image to obtain the preset network for extracting ecological data on earth's surface comprises:

7. A surface biological data extraction device, comprising:

8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the surface ecological data extracting method according to any one of claims 1 to 6 when executing the program.

9. A non-transitory computer-readable storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the surface ecological data extraction method as claimed in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, carries out the steps of the surface ecological data extraction method according to any one of claims 1 to 6.