Stress vegetation indicative information extraction method, device and equipment and storage medium

Abstract

The present application provides a kind of stress vegetation indicative information extraction method, comprising: building stress vegetation indicative information extraction model based on ViT framework;Make source domain and target domain sample;The source domain sample is divided into training set and verification set, as the input of model;Pretreatment operation is carried out to input data;The data after pretreatment is handled by FOPE module, obtains fragmented data, and the feature map is zero-filled;After the fragmented data is encoded by Transform encoder, source domain coding data is obtained;Target domain sample is used as the input data of model, and the above-mentioned pretreatment, fragmentation and zero value filling, coding steps are repeated, and target domain coding data is obtained;The difference between source domain and target domain is calculated by adaptive layer, and the model is adjusted, when the difference is less than the set threshold, the training is ended, and the target domain recognition result is output by Transform decoder.The present application can realize the fast large-scale search of stress vegetation in unknown high vegetation coverage area, and can realize higher recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rock and mineral alteration information extraction, and particularly relates to a stress vegetation indicative information extraction method, device, equipment and storage medium. BACKGROUND

[0002] Rock and mineral alteration information extraction is an indispensable part of mineral resource exploration work. It is extremely difficult to carry out prospecting work using traditional geological means, and combining remote sensing technology with prospecting work can effectively solve the problems of slow and difficult extraction of alteration information.

[0003] For medium and low vegetation coverage areas, passive hyperspectral remote sensing has achieved relatively significant results by inverting mineralization alteration information anomaly spectra. However, for high vegetation coverage areas, the extraction of alteration information is more difficult, and the actual extraction effect is not good. The research has been limited by data acquisition means and intelligent processing algorithms, and has not formed a set of rock and mineral alteration information inversion method suitable for high vegetation coverage areas. Therefore, the abnormal reflection spectrum of stressed vegetation caused by mineralization alteration can be obtained to indirectly judge the composition and distribution of minerals. For stressed vegetation, the spectral information reflected by the bottom of the canopy is significantly stronger than the middle and top of the canopy, showing obvious differential layering characteristics. However, in high vegetation coverage areas, it is difficult to obtain the reflection spectrum of the bottom of the canopy, and it exists in small regional blocks in hyperspectral remote sensing images, so the algorithm needs to have high accuracy when automatically processed.

[0004] The existing related technologies are mainly as follows:

[0005] The Chinese patent "Uranium mineralization alteration information remote sensing identification method for vegetation coverage area, application number CN202111660772.4" provides a volcanic rock type uranium mineralization information remote sensing identification method for vegetation coverage area. This patent is based on the pixel two-part model before extracting alteration information, estimates the vegetation coverage of the remote sensing image, and divides the image into high vegetation coverage area and medium and low vegetation coverage area; after removing the interference of vegetation, the vegetation in the medium and low vegetation coverage area is suppressed, the vegetation in the high vegetation coverage area is masked, and the image map after removing the interference factors in the medium and low vegetation coverage area is obtained; the principal component analysis method is used to extract uranium mineralization alteration anomaly information, and the uranium mineralization alteration information remote sensing identification in the vegetation coverage area is completed. The technical problem of reducing the influence of vegetation interference on remote sensing alteration information extraction is solved.

[0006] The Chinese patent "A method for searching for orogenic gold mines, application number CN202211375733" provides a method for searching for orogenic gold mines, which mainly selects an orogenic belt, then processes remote sensing images and selects regions according to the selected orogenic belt, uses a FieldSpecProFR spectrometer to measure the spectrum of known orogenic gold mine samples, and compares and analyzes them with typical altered mineral models, summarizes their mineral assemblages and alteration characteristics, establishes a spectral dataset of known orogenic gold mine lithology and country rock, then uses the FieldSpecProFR spectrometer to carry out field spectral measurement in favorable gold mineralization areas, identifies altered minerals, establishes a spectral dataset of altered mineral zoning, AI machine automatically selects orogenic gold ore prospecting target area, and finally uses trenching to verify the spectral of the ore prospecting target area to determine the orogenic gold ore body. Through hyperspectral remote sensing combined with surface verification, orogenic gold ore bodies can be quickly delineated within 1 year, greatly shortening the ore prospecting and exploration work cycle, greatly reducing production costs, and saving a lot of manpower and material resources.

[0007] The existing patent mainly solves the problem of extracting alteration information in low-vegetation-covered areas. In the face of high-vegetation-covered areas, the underlying alteration information is more abundant, and the context association is close, which will affect the extraction effect of the existing method. At the same time, in actual application, the model trained by the existing data set will have poor extraction effect when facing the target area due to the differences in region, resolution, scale and style.

[0008] Therefore, in the face of high-vegetation-covered areas, the alteration information extraction effect is poor, which is a technical problem to be solved in the field. SUMMARY

[0009] The main technical problem to be solved by the present application is to realize the rapid and large-scale search for stressed vegetation in unknown high-vegetation-covered areas and improve the identification accuracy.

[0010] In order to solve the above technical problems, the technical scheme adopted by the present application is: a deep learning method of a ViT (Vision Transformer) model constructed by a FOPE (Feature Overlapped Patch Embedding) module is adopted, the deep learning method is designed to better obtain local information and context information, and a Transformer Encoder more suitable for semantic segmentation and a simple and lightweight MLP Decoder have good information extraction capability; since in actual application, known label data sets and target region data sets may differ in region, resolution, size and style, which leads to different feature distributions, therefore, the present application establishes a transferable semantic segmentation model, uses the spatial distribution of alteration information of a known region to quickly migrate to the spatial distribution of the same type of alteration information of other regions, so as to realize the rapid and large-scale search for stressed vegetation in unknown high-vegetation-covered areas. Under the premise of reducing human workload and improving work efficiency, the present application also has high recognition accuracy.

[0011] According to a first aspect of the present application, the present application provides a stressed vegetation indicative information extraction method, comprising the following steps:

[0012] S1: a stressed vegetation indicative information extraction model based on a ViT framework is built, comprising a FOPE module, a Transformer encoder and a Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder;

[0013] S2: source domain samples and target domain samples are made;

[0014] S3: the source domain samples are divided into a training set and a validation set as input data of the model;

[0015] S4: a pretreatment operation is performed on the input data to obtain pretreated data;

[0016] S5: the pretreated data is processed by the FOPE module to obtain fragmented data, and the feature map is zero-filled;

[0017] S6: after the fragmented data passes through the Transformer encoder, source domain encoding data is obtained;

[0018] S7: the target domain samples are used as input data of the model, steps S4-S6 are repeated to obtain target domain encoding data;

[0019] S8: calculating the difference between the source domain coded data and the target domain coded data through the adaptive layer, adjusting the model to reduce the difference, and when the difference is less than a set threshold, the training is completed, and the target domain recognition result is output through the Transformer decoder.

[0020] Further, the step of making the source domain sample comprises:

[0021] In the high vegetation coverage research area, the hyperspectral image of the stressed vegetation canopy bottom is selected according to the hyperspectral data and the laser radar point cloud data.

[0022] The hyperspectral image of the stressed vegetation canopy bottom is cropped into an image with a size of 512*512 pixels and made into a source domain sample.

[0023] Further, the step of making the target domain sample comprises:

[0024] In the high vegetation coverage research area, the hyperspectral data is cropped into an image with a size of 512*512 pixels, and all of them are used as target domain samples.

[0025] Further, the step of dividing the source domain sample into a training set and a validation set comprises:

[0026] The source domain sample is divided into a training set and a validation set according to a ratio of 8:2.

[0027] Further, the step of pre-processing the input data to obtain pre-processed data comprises:

[0028] The image size, color space and channel number of the input data are uniformly standardized. The hyperspectral data has multiple spectral bands in addition to red, green and blue. Through principal component analysis, the dimensionality is reduced to 10 dimensions, and the channel number of the model is also set to 10 to obtain more spectral information data.

[0029] Further, the FOPE module comprises a convolution layer, a maximum pooling layer and an Overlapped patch embedding connected in sequence.

[0030] Further, the step of processing the pre-processed data through the FOPE module to obtain patch data and zero-padding the feature map comprises:

[0031] A plurality of convolution kernels are used to perform convolution operation on the pre-processed data to obtain local features; wherein the number of convolution kernels is consistent with the number of channels of the model, so that the dimension of the data does not change;

[0032] The obtained local features are subjected to maximum pooling processing.

[0033] The feature map is zero-filled, and then patch data is obtained through overlapped patch embedding.

[0034] According to a second aspect of the present application, the present application provides a stress vegetation indicative information extraction device, comprising the following modules:

[0035] The building module is configured to build a stress vegetation indicative information extraction model based on a ViT framework, comprising a FOPE module, a Transformer encoder and a Transformer decoder; and an adaptive layer is added to the last layer of the Transformer encoder;

[0036] The making module is configured to make source domain samples and target domain samples;

[0037] The dividing module is configured to divide the source domain samples into a training set and a validation set as input data of the model;

[0038] The preprocessing module is configured to perform a preprocessing operation on the input data to obtain preprocessed data;

[0039] The patching and zero-padding module is configured to process the preprocessed data through the FOPE module to obtain patch data, and perform zero padding on the feature map;

[0040] The encoding module is configured to obtain source domain encoding data by processing the patch data through the Transformer encoder;

[0041] The encoding module is further configured to repeat the preprocessing, patching and zero-padding operations by taking the target domain samples as input data of the model, to obtain target domain encoding data;

[0042] The training module is configured to calculate the difference between the source domain encoding data and the target domain encoding data through the adaptive layer, to adjust the model to reduce the difference, and to output a target domain recognition result through the Transformer decoder when the difference is less than a set threshold.

[0043] According to a third aspect of the present application, the present application provides 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 stress vegetation indicative information extraction method when executing the program.

[0044] According to a fourth aspect of the present application, the present application provides a storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the stress vegetation indicative information extraction method.

[0045] The technical solution provided by the present application has the following beneficial effects:

[0046] Compared with the prior art, the convolutional network can obtain better local information but lacks global information, the ViT can better obtain global information but is not sensitive to local information, and adjacent information of different patches is prone to be lost. In the technical scheme provided in the present application, the FOPE module is used to better obtain local information and also obtain global information and context information, so that the model obtains rich spatial position information and context information, improves the utilization rate of information, and thus the network learns more features conducive to the classification of stressed plants. Moreover, by adding an adaptive layer, the problem that the performance of the trained model greatly decreases when facing a new research area is solved, the problem that the traditional classification method cannot cope with the extraction of mineral alteration vegetation stress information under high vegetation coverage due to the lack of accurate samples and autonomous learning ability is solved, and a method for extracting rock and mineral alteration information in a high vegetation coverage area is established. The method also improves the efficiency of target ore district positioning. BRIEF DESCRIPTION OF DRAWINGS

[0047] The present application will be further described below in conjunction with the drawings and embodiments, wherein:

[0048] Figure 1 It is a general flowchart of the stress vegetation indicative information extraction method of the present application.

[0049] Figure 2 It is a schematic diagram of the Overlapped patch embedding of the present application.

[0050] Figure 3 It is a structural schematic diagram of the stress vegetation indicative information extraction device of the present application.

[0051] Figure 4 It is a structural schematic diagram of the electronic device of the present application. DETAILED DESCRIPTION

[0052] In order to have a clearer understanding of the technical features, objectives and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the drawings.

[0053] REFERENCE Figure 1 , Figure 1 It is a general flowchart of the stress vegetation indicative information extraction method of the present application, which specifically includes the following steps:

[0054] S1: build a stress vegetation indicative information extraction model based on a ViT framework, including a FOPE module, a Transformer encoder and a Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder;

[0055] S2: source domain samples and target domain samples are made;

[0056] S3: the source domain samples are divided into a training set and a validation set as input data of the model;

[0057] S4: a pretreatment operation is performed on the input data to obtain pretreated data;

[0058] S5: the pretreated data is processed by a FOPE module to obtain fragmented data, and the feature map is zero-filled;

[0059] S6: the fragmented data is obtained after passing through a Transformer encoder;

[0060] S7: the target domain samples are used as input data of the model, and steps S4-S6 are repeated to obtain target domain encoded data;

[0061] S8: the difference between the source domain encoded data and the target domain encoded data is calculated through an adaptive layer, the model is adjusted to reduce the difference, and when the difference is less than a set threshold, the training is completed, and the target domain recognition result is output through a Transformer decoder.

[0062] In this embodiment, the threshold is preferably 0.3.

[0063] Based on but not limited to the above method, the making of the source domain samples in S2 specifically includes the following steps:

[0064] In a high vegetation coverage research area, hyperspectral images of the bottom of the stressed vegetation canopy are selected according to hyperspectral data and laser radar point cloud data;

[0065] The hyperspectral images of the bottom of the stressed vegetation canopy are cropped into images with a size of 512x512 pixels, and source domain samples are made.

[0066] Based on but not limited to the above method, the making of the target domain samples in S2 specifically includes the following steps:

[0067] In a high vegetation coverage research area, the hyperspectral data are cropped into images with a size of 512x512 pixels, and all of them are used as target domain samples.

[0068] Based on but not limited to the above method, in step S3, the source domain samples are divided into a training set and a validation set according to a ratio of 8:2.

[0069] Based on but not limited to the above method, in step S4, the step of performing a pretreatment operation on the input data to obtain pretreated data includes:

[0070] The image size, color space and channel number of the input data are uniformly standardized, the hyperspectral data has multiple spectral bands in addition to red, green and blue, the dimensionality is reduced to 10 dimensions through principal component analysis, and the channel number of the model is also set to 10 to obtain more spectral information data.

[0071] Based on but not limited to the above method, in step S5, the FOPE module includes a convolution layer, a maximum pooling layer and an overlapped patch embedding connected in sequence.

[0072] Based on the FOPE module, the specific implementation of step S5 is as follows:

[0073] The preprocessed data is convolved using multiple convolution kernels to obtain local features, wherein the number of convolution kernels is consistent with the number of channels of the model, so that the dimension of the data does not change;

[0074] The obtained local features are subjected to maximum pooling processing.

[0075] The feature map is zero-padded, and then the patch data is obtained through overlapped patch embedding. In this embodiment, the obtained patch data is as shown in Figure 2

[0076] The key points of the embodiment of the application are also as follows:

[0077] Key point 1: The Transformer module constructed using the new FOPE module can better extract regional local detail features and global context information. First, the convolution layer is used to obtain the features of the local region, and the maximum pooling is used to retain the main features, extract the texture information of the features, and expand the encoding features from the previous level to the next level through the overlapped patch merging operation. Specifically, the patch window is expanded, the regions of adjacent windows are partially overlapped, and the feature map is zero-padded to retain the resolution.

[0078] Key point 2: In actual application, due to the difference in area, resolution, scale and style between the labeled training data and the unlabeled target data, different feature distributions are caused, which is called data shift. The labeled data set is usually defined as the source domain (SD), and the unlabeled data set is defined as the target domain (TD).

[0079] The application adds an adaptive layer at the last layer of the model encoder, and calculates the feature distribution difference from the source domain and the target domain through the adaptive layer.

[0080] ​It should be noted that the adaptive layer is used to measure the approximation degree between the two distributions by constructing a measurement criterion. The most commonly used measurement criterion is the maximum mean difference (MMD), which is mainly used to calculate the distance between two distributions in the reproding kernel hilbert space (RKHS, Reproding Kernel Hilbert Space, RKHS). Given a supervised source domain D s ={x i ,y i} and an unsupervised target domain D t ={x i ,y i}, and assume that the data distribution P(X s ) and P(X t ) of the two fields are different, that is, P(X s )≠P(X t ). The MMD distance between the two fields can be expressed as:

[0081]

[0082] Where f(·) is a mapping function used to map data to a reproding kernel hilbert space. Let the number of source domain and target domain samples be n x and n t . The reason for adding an adaptive layer at the last layer of the encoder is that the last layer of the encoder extracts high-level semantic features in high-resolution images, hyperspectral images and laser radar images, which makes the adaptive layer more accurately calculate the distribution difference between the source domain and the target domain, so that the model can obtain better results on the target domain.

[0083] A stress vegetation indicative information extraction device provided by the present application is described below. The stress vegetation indicative information extraction device described below can be mutually corresponding and referred to with the stress vegetation indicative information extraction method described above.

[0084] As shown in Figure 3 , a stress vegetation indicative information extraction device includes the following modules:

[0085] The building module 001 is used to build a stress vegetation indicative information extraction model based on the ViT framework, including: FOPE module, Transformer encoder and Transformer decoder; an adaptive layer is added at the last layer of the Transformer encoder;

[0086] The manufacturing module 002 is used to manufacture source domain samples and target domain samples;

[0087] The division module 003 is used to divide the source domain samples into a training set and a validation set as input data of the model;

[0088] a preprocessing module 004, configured to perform a preprocessing operation on input data to obtain preprocessed data;

[0089] a slicing and zero value padding module 005, configured to process the preprocessed data through a FOPE module to obtain sliced data, and perform zero value padding on a feature map;

[0090] an encoding module 006, configured to obtain source domain encoded data by performing a Transformer encoder on the sliced data;

[0091] The encoding module 006 is further configured to repeat the operations of preprocessing, slicing and zero value padding by taking a target domain sample as input data of the model to obtain target domain encoded data.

[0092] a training module 007, configured to calculate a difference between the source domain encoded data and the target domain encoded data through an adaptive layer, adjust the model to reduce the difference, and output a target domain recognition result through a Transformer decoder when the difference is less than a set threshold.

[0093] In this embodiment, the threshold is preferably 0.3.

[0094] Based on but not limited to the above method, the production module 002 is specifically configured to perform the following steps:

[0095] The source domain sample is produced as follows: in a high vegetation coverage research area, a hyperspectral image at the bottom of a stressed vegetation canopy is selected according to hyperspectral data and laser radar point cloud data; the hyperspectral image at the bottom of the stressed vegetation canopy is cropped into an image with a size of 512*512 pixels, and is produced into a source domain sample.

[0096] The target domain sample is produced as follows: in a high vegetation coverage research area, the hyperspectral data is cropped into an image with a size of 512*512 pixels, and is all taken as a target domain sample.

[0097] Based on but not limited to the above device, the division module 003 is specifically configured to divide the source domain sample into a training set and a verification set according to a ratio of 8:2.

[0098] Based on but not limited to the above device, the preprocessing module 004 is specifically configured to perform the following steps:

[0099] The image size, color space and channel number of the input data are uniformly standardized. The hyperspectral data has multiple spectral bands in addition to red, green and blue. The dimensionality is reduced to 10 dimensions through principal component analysis, and the channel number of the model is also set to 10 to obtain more spectral information data.

[0100] Based on but not limited to the above device, the FOPE module comprises a convolution layer, a maximum pooling layer and an Overlapped patch embedding connected in sequence.

[0101] Based on the FOPE module, the fragmentation and zero value padding module 005 specifically performs the following steps:

[0102] The preprocessed data is convolved using a plurality of convolution kernels to obtain local features; wherein the number of convolution kernels is consistent with the number of channels of the model, so that the dimension of the data does not change; the obtained local features are maximum pooled; the feature map is zero padded, and then the fragmented data is obtained through the overlapped patch embedding.

[0103] As shown in Figure 4 An entity structure diagram of an electronic device is shown, which can include a processor 610, a communications interface 620, a memory 630 and a communications bus 640, wherein the processor 610, the communications interface 620 and the memory 630 complete mutual communication through the communications bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the above-mentioned one kind of stress vegetation indicative information extraction method, including: S1: building a stress vegetation indicative information extraction model based on the ViT framework, including: FOPE module, Transformer encoder and Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder; S2: making source domain samples and target domain samples; S3: dividing the source domain samples into a training set and a validation set as input data of the model; S4: pre-processing the input data to obtain pre-processed data; S5: processing the pre-processed data through the FOPE module to obtain fragmented data, and zero padding the feature map; S6: obtaining source domain coding data after the fragmented data passes through the Transformer encoder; S7: taking the target domain sample as the input data of the model, repeating steps S4-S6 to obtain the target domain coding data; S8: calculating the difference between the source domain coding data and the target domain coding data through the adaptive layer, adjusting the model to reduce the difference, and when the difference is less than a set threshold, the training is ended, and the target domain recognition result is output through the Transformer decoder.

[0104] Moreover, the logic instructions in the memory 630 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0105] In another aspect, the embodiments of the present application also provide a storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the above-described stress vegetation indicative information extraction method, including: S1: building a stress vegetation indicative information extraction model based on a ViT framework, including: a FOPE module, a Transformer encoder, and a Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder; S2: preparing source domain samples and target domain samples; S3: dividing the source domain samples into a training set and a validation set as input data of the model; S4: performing a preprocessing operation on the input data to obtain preprocessed data; S5: processing the preprocessed data through the FOPE module to obtain fragmented data, and performing zero value padding on the feature map; S6: obtaining source domain coding data after the fragmented data passes through the Transformer encoder; S7: taking the target domain samples as input data of the model, repeating steps S4-S6 to obtain target domain coding data; S8: calculating the difference between the source domain coding data and the target domain coding data through the adaptive layer, adjusting the model to reduce the difference, and when the difference is less than a set threshold, the training is ended, and the target domain recognition result is output through the Transformer decoder.

[0106] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles, or systems that include a series of elements not only include those elements, but also include other elements not explicitly listed, or inherent to such processes, methods, articles, or systems. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or system that includes the element.

[0107] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. In the unit claims in which several devices are listed, several of the devices can be embodied by the same hardware item. The use of the words first, second, and third does not represent any order, and the words can be interpreted as identifiers.

[0108] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method of extracting information indicative of stress vegetation, characterized by, The method comprises the following steps: S1: building a stress vegetation indicative information extraction model based on a ViT framework, comprising a FOPE module, a Transformer encoder and a Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder; S2: preparing source domain samples and target domain samples; S3: dividing the source domain samples into a training set and a validation set as input data of the model; S4: performing a preprocessing operation on the input data to obtain preprocessed data; S5: processing the preprocessed data through the FOPE module to obtain sliced data, and performing zero value padding on the feature map; S6: obtaining source domain encoding data after the sliced data passes through the Transformer encoder; S7: taking the target domain samples as input data of the model, repeating steps S4-S6 to obtain target domain encoding data; S8: calculating the difference between the source domain encoding data and the target domain encoding data through the adaptive layer, adjusting the model to reduce the difference, and when the difference is less than a set threshold, the training is completed, and the target domain recognition result is output through the Transformer decoder; The steps of preparing the source domain samples comprise: In a high vegetation coverage research area, selecting hyperspectral images at the bottom of the stressed vegetation canopy according to hyperspectral data and laser radar point cloud data; Cutting the hyperspectral images at the bottom of the stressed vegetation canopy into images with a size of 512*512 pixels, and preparing the source domain samples; The steps of preparing the target domain samples comprise: In a high vegetation coverage research area, cutting the hyperspectral data into images with a size of 512*512 pixels, and taking all of them as target domain samples; The FOPE module comprises a convolution layer, a max pooling layer and an Overlapped patch embedding connected in sequence.

2. The stress vegetation index information extraction method according to claim 1, characterized by, The steps of dividing the source domain samples into a training set and a validation set comprise: Dividing the source domain samples into a training set and a validation set according to a ratio of 8:

2.

3. The stress vegetation index information extraction method according to claim 1, characterized by, The steps of performing a preprocessing operation on the input data to obtain preprocessed data comprise: Uniformly standardizing the image size, color space and channel number of the input data; the hyperspectral data has multiple spectral bands in addition to red, green and blue; the dimensionality is reduced to 10 dimensions through principal component analysis, and the channel number of the model is also set to 10 to obtain more spectral information data.

4. The stress vegetation index information extraction method according to claim 3, characterized by, The steps of processing the preprocessed data through the FOPE module to obtain sliced data and performing zero value padding on the feature map comprise: Using multiple convolution kernels to perform convolution operation on the preprocessed data to obtain local features; the number of convolution kernels is consistent with the channel number of the model, so that the dimensionality of the data does not change; Performing max pooling processing on the obtained local features; Performing zero value padding on the feature map, and then obtaining sliced data through overlapped patch embedding.

5. A stress vegetation index information extracting apparatus for implementing the method according to any one of claims 1 to 4, characterized by, The method comprises the following modules: The stress vegetation indicative information extraction model based on the ViT framework is built, and the model includes a FOPE module, a Transformer encoder and a Transformer decoder; an adaptive layer is added to the last layer of the Transformer encoder; The source domain samples and the target domain samples are made; The source domain samples are divided into a training set and a validation set as input data of the model; The input data is preprocessed to obtain preprocessed data; The preprocessed data is processed by the FOPE module to obtain sliced data, and the feature map is zero-padded; The sliced data is input into the Transformer encoder to obtain source domain encoding data; The target domain samples are input into the model as input data, and the preprocessing, slicing and zero-padding operations are repeated to obtain target domain encoding data; The training module is used to calculate the difference between the source domain encoding data and the target domain encoding data through the adaptive layer, adjust the model to reduce the difference, and when the difference is less than a set threshold, the training is completed, and the target domain recognition result is output through the Transformer decoder; The steps of making the source domain samples include: In a high vegetation coverage research area, high spectral images of the bottom of the stressed vegetation canopy are selected according to high spectral data and laser radar point cloud data; The high spectral images of the bottom of the stressed vegetation canopy are cropped into images with a size of 512*512 pixels and made into source domain samples; The steps of making the target domain samples include: In a high vegetation coverage research area, high spectral data is cropped into images with a size of 512*512 pixels, and all of them are used as target domain samples; The FOPE module includes a convolution layer, a max pooling layer and an Overlapped patch embedding connected in sequence.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the stress vegetation indicative information extraction method according to any one of claims 1-4.

7. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the stress vegetation indicative information extraction method according to any one of claims 1-4.

Images