Biodiversity index extraction method, device, equipment, medium and product

Abstract

The invention discloses a biodiversity index extraction method, device and equipment, a medium and a product, and relates to the technical field of ecological protection and management. The method comprises the following steps: processing to obtain a label image of a target area and a plurality of sample images with different dimension information based on optical remote sensing image data and ecological special product data in a historical period, and then obtaining sample data based on the images; carrying out model training on a lightweight convolutional neural network model based on a ConvNeXt framework by using the sample data to obtain a label image generation model, and finally, carrying out model training on each sub-region in the target region, respectively extracting corresponding sub-region sample images from each new sample image in the target region to gather and input the sub-region sample images into a tag image generation model to obtain corresponding sub-region tag images, and gathering and splicing to generate a biodiversity index distribution image of the target region in the target period, therefore, the biodiversity index extraction precision and efficiency can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of ecological protection and management, and specifically relates to a method, device, equipment, medium and product for extracting biodiversity indicators. Background Art

[0002] Biodiversity surveys and monitoring are crucial foundations for ecological conservation and management. Traditional ground-based survey methods are time-consuming, costly, and have limited spatial coverage in complex terrain, necessitating the urgent need for efficient and automated solutions.

[0003] Remote sensing technology, as a non-contact, large-scale observation method, can overcome the spatial limitations of ground-based surveys and is increasingly being used in ecological and biodiversity monitoring and assessment. For example, Wang Shu et al. used remote sensing to assess the ecological environment of specific areas, such as Qionglai Mountain, while Li Ainong et al. implemented integrated space-based, ground-based biodiversity monitoring. Furthermore, remote sensing technology can be used to monitor and assess changes in habitat quality and human disturbance, as well as to extract biodiversity indices such as biomass and vegetation cover.

[0004] Remote sensing data typically contains a large amount of information and requires effective processing and analysis to obtain useful results. However, traditional statistical methods and models such as classical machine learning methods rely on manual feature extraction, are limited to shallow models, and have difficulty processing nonlinear data. This leads to inefficiency and information loss when processing large-scale data such as multi-source remote sensing. The combination of deep learning and remote sensing imagery has become a research hotspot in recent years. Using deep learning technology to automatically classify remote sensing images, extract features, and detect targets can greatly improve the efficiency and accuracy of remote sensing image processing, and has a wide range of applications in urban planning, environmental monitoring, and disaster assessment. However, existing research on biodiversity indicator extraction based on deep learning is still relatively limited, mainly involving the inversion of forest height and biomass and grassland biomass, resulting in problems such as difficult data labeling and insufficient model generalization.

[0005] Convolutional Neural Network (CNN) is a deep learning model widely used in fields such as image recognition and speech processing. CNN simulates the receptive field mechanism of biological neurons, uses convolutional layers to extract local features of input data, and reduces feature dimensions through pooling layers, thereby enhancing the model's perception of spatial structure. The network is usually composed of multiple convolutional layers, pooling layers, activation functions, and fully connected layers. CNN has the characteristics of parameter sharing and sparse connections, which makes it more computationally efficient when processing high-dimensional data and can effectively reduce overfitting. Therefore, how to integrate optical remote sensing image data with ecological thematic product data, and construct a deep learning model based on convolutional neural networks that fuses attribute features and spatial features to perform automatic regression extraction of biodiversity indicators, thereby improving the accuracy and efficiency of biodiversity indicator extraction and providing a new and scalable technical path for mountain biodiversity monitoring, is a topic that technicians in this field urgently need to study. Summary of the Invention

[0006] The purpose of the present invention is to provide a biodiversity indicator extraction method, apparatus, computer device, computer-readable storage medium, and computer program product to address the problems of existing biodiversity indicator extraction schemes, such as difficulty in data labeling, insufficient model generalization, and low extraction accuracy and efficiency.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, a method for extracting biodiversity indicators is provided, comprising:

[0009] Acquire historical optical remote sensing image data and ecological thematic product data, where the ecological thematic product data includes distribution images of biodiversity indicators;

[0010] Based on the optical remote sensing image data processing of the historical period, multiple sample images of the target area with different dimensional information are obtained;

[0011] Extract the biodiversity indicator distribution image of the target area from the historical ecological thematic product data, and align the biodiversity indicator distribution image with the sample image to obtain the label image of the target area;

[0012] Based on multiple sample images and label images, sample data is obtained as follows: a sub-region label image containing m×m label pixels is extracted from the label image as the model output item, and then a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image is extracted from each sample image to be summarized as the model input item, and finally the model input item and the model output item are used as a sample data, where m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, η l is the resolution of the label image and η s multiples of;

[0013] All sample data are divided into training set, validation set and test set, and the division results are used to train the lightweight convolutional neural network model based on the ConvNeXt architecture to obtain the labeled image generation model;

[0014] Based on the optical remote sensing image data processing of the target period, multiple new sample images of the target area are obtained;

[0015] For each sub-region in the target area, extract the corresponding sub-region sample image containing n×n sample pixels from each new sample image in the target area to summarize the input label image generation model, and output the corresponding sub-region label image;

[0016] Based on the sub-region label images of each sub-region, the biodiversity indicator distribution image of the target area in the target period is generated by splicing.

[0017] Based on the above invention, a new solution for extracting biodiversity indicators based on deep learning and remote sensing imagery is provided. The solution first processes optical remote sensing image data and ecological thematic product data from historical periods to obtain a labeled image of the target area and multiple sample images with different dimensional information. Sample data is then obtained based on these images, and a lightweight convolutional neural network model based on the ConvNeXt architecture is trained with these sample data to obtain a labeled image generation model. Finally, for each sub-region in the target area, corresponding sub-region sample images are extracted from each new sample image of the target area to be aggregated and input into the labeled image generation model to obtain corresponding sub-region labeled images. These images are then aggregated and spliced ​​to generate a biodiversity indicator distribution image for the target area during the target period. By integrating optical remote sensing image data with ecological thematic product data and constructing a deep learning model based on a convolutional neural network that fuses attribute features and spatial features, biodiversity indicators can be automatically regressed and extracted. This reduces the difficulty of data labeling, improves model generalization, and enhances the accuracy and efficiency of biodiversity indicator extraction. This provides a new and scalable technology path for biodiversity monitoring in mountainous areas, facilitating practical application and promotion.

[0018] In a possible design, the biodiversity indicator includes vegetation cover and / or aboveground biomass.

[0019] In one possible design, multiple sample images of the target area with different dimensional information are obtained based on historical optical remote sensing image data processing, including:

[0020] extracting a blue band spectral image, a green band spectral image, a red band spectral image, a near infrared band spectral image, and a short-wave infrared band spectral image of the target area from the optical remote sensing image data of the historical period;

[0021] Based on the red band spectral image and the near infrared band spectral image of the target area, a normalized vegetation index distribution image of the target area is constructed;

[0022] Based on the blue band spectral image, the red band spectral image and the near infrared band spectral image of the target area, an enhanced vegetation index distribution image of the target area is constructed;

[0023] Based on the near-infrared band spectral image and the red band spectral image of the target area, a ratio vegetation coefficient distribution image of the target area is constructed;

[0024] The blue band spectral image, green band spectral image, red band spectral image, near infrared band spectral image, short wave infrared band spectral image, normalized vegetation index distribution image, enhanced vegetation index distribution image and ratio vegetation coefficient distribution image of the target area are used as multiple sample images of the target area with different dimensional information.

[0025] In one possible design, the ConvNeXt architecture adopts a ConvNeXt variant architecture having the following features (A) to (B):

[0026] (A) the input layer is used to input a plurality of sub-region sample images corresponding one-to-one to the plurality of sample images, and the output layer is used to output a sub-region label image that covers the same area as the plurality of sub-region sample images and is used to reflect the distribution of biodiversity indicators;

[0027] (B) The network structure is composed of multiple stacked ConvNeXt blocks, and the internal structure of each ConvNeXt block includes a depth-wise separable convolution module, a layer normalization module, a GELU activation function module, a Dropout layer module and a residual connection module. The depth-wise separable convolution module uses a depth-wise separable convolution kernel of size 7×7, and first performs convolution operations independently in each image channel, and then realizes cross-channel information fusion through 1×1 point-by-point convolution.

[0028] In one possible design, each sub-region is obtained by performing sliding window processing on the target region according to a sliding window overlap rate of δ%, where δ represents a positive integer greater than or equal to 30 and less than or equal to 70;

[0029] Based on the sub-region label images of each sub-region, a biodiversity indicator distribution image of the target region during the target period is generated by stitching, including:

[0030] For each pixel in the biodiversity indicator distribution image to be generated, the corresponding pixel value is calculated according to the sub-region label image of each sub-region using the following formula:

[0031]

[0032] In the formula, i represents a positive integer, P i represents the pixel value of the i-th pixel in the biodiversity indicator distribution image to be generated, K i represents the total number of times the i-th pixel is covered by the sub-region label images of each sub-region, p i,k represents the pixel value of the i-th pixel in the sub-region label image when it is covered for the kth time, w i,krepresents the weight coefficient corresponding to the position of the i-th pixel in the sub-region label image used for the k-th coverage in a two-dimensional Gaussian weight matrix of size m×m, wherein the two-dimensional Gaussian weight matrix has the characteristics of having the largest weight at the center, gradually decreasing weight toward the edge, and using a standard deviation σ=1.5 to control the width of the weight distribution;

[0033] The biodiversity indicator distribution image of the target area in the target period is generated according to the pixel value of each pixel.

[0034] In one possible design, the method further includes: using a mask mechanism to eliminate invalid areas in the input image, retaining only valid pixels for model training and application.

[0035] In a second aspect, a biodiversity indicator extraction device is provided, comprising a multi-source data acquisition unit, a sample image processing unit, a label image processing unit, a sample data acquisition unit, a generation model training unit, a generation model application unit, and a distribution image stitching unit;

[0036] The multi-source data acquisition unit is used to acquire optical remote sensing image data and ecological thematic product data of historical periods, wherein the ecological thematic product data includes distribution images of biodiversity indicators;

[0037] The sample image processing unit is communicatively connected to the multi-source data acquisition unit and is used to process the optical remote sensing image data of the historical period to obtain a plurality of sample images of the target area with different dimensional information;

[0038] The label image processing unit is communicatively connected to the multi-source data acquisition unit and is used to extract a biodiversity indicator distribution image of the target area from the ecological thematic product data of the historical period, and register the biodiversity indicator distribution image with the sample image to obtain a label image of the target area;

[0039] The sample data acquisition unit is connected to the sample image processing unit and the label image processing unit in a distributed communication manner, and is used to acquire sample data based on multiple sample images and label images in the following manner: extracting a sub-region label image containing m×m label pixels from the label image as a model output item, and then extracting a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image from each sample image to summarize as a model input item, and finally using the model input item and the model output item as a piece of sample data, wherein m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, η lis the resolution of the label image and η s multiples of;

[0040] The generation model training unit is communicatively connected to the sample data acquisition unit, and is used to divide all sample data into a training set, a validation set, and a test set, and use the division results to perform model training on a lightweight convolutional neural network model based on the ConvNeXt architecture to obtain a label image generation model;

[0041] The sample image processing unit is further used to obtain a target area and multiple new sample images based on the optical remote sensing image data of the target period;

[0042] The generation model application unit is connected to the generation model training unit and the sample image processing unit in a distributed communication manner, and is used to extract, for each sub-region in the target region, a sub-region sample image containing n×n sample pixels and corresponding to each new sample image of the target region to summarize the input label image generation model, and output the corresponding sub-region label image;

[0043] The distribution image splicing unit is communicatively connected to the generation model application unit, and is used to splice and generate a biodiversity indicator distribution image of the target area in the target period based on the sub-area label images of each sub-area.

[0044] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a transceiver communicatively connected in sequence, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the biodiversity indicator extraction method as described in the first aspect or any possible design of the first aspect.

[0045] In a fourth aspect, the present invention provides a computer-readable storage medium having instructions stored thereon. When the instructions are executed on a computer, the method for extracting biodiversity indicators as described in the first aspect or any possible design of the first aspect is executed.

[0046] In a fifth aspect, the present invention provides a computer program product, comprising a computer program or instructions, which, when executed by a computer, implements the method for extracting biodiversity indicators as described in the first aspect or any possible design of the first aspect.

[0047] Beneficial effects of the above scheme:

[0048] (1) The present invention creatively provides a new solution for extracting biodiversity indicators based on deep learning and remote sensing images, namely, firstly, based on the optical remote sensing image data and ecological thematic product data of the historical period, the label image of the target area and multiple sample images with different dimensional information are processed, and then sample data are obtained based on these images, and these sample data are used to train the lightweight convolutional neural network model based on the ConvNeXt architecture to obtain a label image generation model, and finally, for each sub-area in the target area, the corresponding sub-area sample image is extracted from each new sample image of the target area to summarize the input label image generation model to obtain the corresponding sub-area label image, and the biodiversity indicator distribution image of the target area in the target period is summarized and spliced. Thus, by integrating the optical remote sensing image data and the ecological thematic product data, and constructing a deep learning model that integrates attribute features and spatial features based on the convolutional neural network, the biodiversity indicators can be automatically regressed and extracted, thereby reducing the difficulty of data labeling, improving the generalization ability of the model, and improving the accuracy and efficiency of biodiversity indicator extraction, providing a new technology path that can be promoted for mountain biodiversity monitoring, and facilitating practical application and promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A schematic diagram of the process of extracting biodiversity indicators provided in the embodiments of this application.

[0051] Figure 2 A schematic diagram of the structure of the ConvNeXt variant architecture provided in an embodiment of the present application.

[0052] Figure 3 A schematic diagram of the structure of a ConvNeXt block in a ConvNeXt variant architecture provided in an embodiment of the present application.

[0053] Figure 4 This is an example diagram of the positions of pixels provided in an embodiment of the present application when they are covered each time and in the sub-region label image used for the covering.

[0054] Figure 5 The vegetation coverage distribution image generation result of Muchuan County in the target period provided by the embodiment of this application, wherein: Figure 5(a) shows the vegetation coverage distribution image generation result of Muchuan County in 2020. Figure 5 (b) in the figure shows the vegetation coverage distribution image generation results of Muchuan County in 2021.

[0055] Figure 6 The aboveground biomass distribution image generation result of Muchuan County in the target period provided by the embodiment of this application, wherein: Figure 6 (a) shows the aboveground biomass distribution image generation result of Muchuan County in 2020. Figure 6 (b) in the figure shows the results of generating the aboveground biomass distribution image in Muchuan County in 2021.

[0056] Figure 7 The predicted value of vegetation coverage in Muchuan County in 2020 and the distribution results of the reference value of the public dataset provided in the embodiment of this application are as follows: Figure 7 (a) shows the distribution of predicted vegetation coverage values ​​in Muchuan County in 2020. Figure 7 (b) shows the distribution results of the reference values ​​of the public dataset of vegetation coverage in Muchuan County in 2020.

[0057] Figure 8 The predicted value of vegetation coverage in Muchuan County in 2021 and the distribution results of the reference value of the public dataset provided in the embodiment of this application are as follows: Figure 8 (a) shows the distribution results of the predicted vegetation coverage value in Muchuan County in 2021. Figure 8 (b) shows the distribution results of the reference values ​​of the public dataset of vegetation coverage in Muchuan County in 2021.

[0058] Figure 9 The aboveground biomass prediction value of Muchuan County in 2020 and the distribution results of the reference value of the public dataset provided in the embodiment of this application are as follows: Figure 9 (a) shows the distribution results of the predicted aboveground biomass in Muchuan County in 2020. Figure 9 (b) shows the distribution results of the reference values ​​of the aboveground biomass public dataset in Muchuan County in 2020.

[0059] Figure 10 The aboveground biomass prediction value of Muchuan County in 2021 and the reference value distribution results of the public dataset provided in the embodiment of this application are as follows: Figure 10 (a) shows the distribution results of the predicted aboveground biomass in Muchuan County in 2021. Figure 10 (b) shows the distribution results of the reference values ​​of the aboveground biomass public dataset in Muchuan County in 2021.

[0060] Figure 11This is a schematic diagram of the structure of the biodiversity indicator extraction device provided in an embodiment of the present application.

[0061] Figure 12 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention will be briefly introduced below in conjunction with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structures of the drawings is only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these embodiments without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0063] It should be understood that although the terms first, second, etc. may be used herein to describe various objects, these objects should not be limited by these terms. These terms are merely used to distinguish one object from another. For example, a first object can be referred to as a second object, and similarly, a second object can be referred to as a first object without departing from the scope of the exemplary embodiments of the present invention.

[0064] It should be understood that the term "and / or" that may appear in this document is merely a description of the association relationship between associated objects, indicating that there may be three relationships. For example, A and / or B can indicate three situations: A exists alone, B exists alone, or A and B exist at the same time. For another example, A, B and / or C can indicate the existence of any one of A, B and C or any combination of them. The term " / and" that may appear in this document describes another type of association object relationship, indicating that there may be two relationships. For example, A / and B can indicate two situations: A exists alone or A and B exist at the same time. In addition, the character " / " that may appear in this document generally indicates that the previous and next associated objects are in an "or" relationship.

[0065] Example

[0066] like Figure 1As shown, the biodiversity indicator extraction method provided in the first aspect of this embodiment can be, but is not limited to, executed by a computer device with certain computing resources, such as a cloud server, a personal computer (PC, a multi-purpose computer with a size, price, and performance suitable for personal use; desktops, laptops, small laptops, tablets, and ultrabooks are all personal computers), a smart phone, a personal digital assistant (PDA), or a wearable device. Figure 1 As shown, the biodiversity indicator extraction method may include, but is not limited to, the following steps S1 to S7 and S8.

[0067] S1. Acquire optical remote sensing image data and ecological thematic product data from a historical period, wherein the ecological thematic product data includes but is not limited to distribution images of biodiversity indicators.

[0068] In step S1, specifically, but not limited to, the Sentinel-2 series remote sensing image data and ecological parameter data (i.e., the ecological thematic product data) from 2020 to 2021 can be selected to construct a deep learning model for the automatic extraction of biodiversity indicators (i.e., the subsequent labeled image generation model). For example, the European Space Agency's Copernicus Open Data Center (https: / / scihub.copernicus.eu / dhus / # / home) can be accessed to obtain the Sentinel-2 Level-2A products taken by the S2B satellite in June 2020 and the S2A satellite in August 2021, and used as the optical remote sensing image data for the historical period. In order to ensure the clarity and availability of remote sensing images, the selected images can all be high-quality images with a cloud cover of less than 10% to avoid interference with vegetation indices and model training caused by cloud cover. Considering that there are three resolutions of 10m, 20m and 60m in the Sentinel-2 series images, in order to facilitate model training, the bilinear interpolation method can also be used to resample all band images to a resolution of 10m.

[0069] In step S1, the biodiversity indicators specifically include, but are not limited to, fractional vegetation cover (FVC) and / or above-ground biomass (AGB). The 2020–2021 semi-monthly vegetation cover product released by the National Tibetan Plateau Data Center and / or the 2020–2021 annual above-ground biomass data provided by the China Forest Above-Ground Biomass Scientific Database can be obtained and used as the ecological thematic product data for the historical period. Furthermore, the spatial distribution of biodiversity indicators is influenced by multiple ecological factors (such as topography, vegetation status, and human activities), which often have complex coupling relationships. For example, changes in elevation can affect local temperature and humidity, thereby indirectly influencing the distribution pattern of vegetation communities. Slope aspect affects vegetation productivity by regulating light conditions, and changes in vegetation cover are directly related to above-ground biomass, making it a critical parameter in biodiversity assessment. Therefore, in order to extract the biodiversity indicators more comprehensively, data selection should be based on its ecological explanatory power while fully considering current accessibility and adaptability. It is also possible to obtain terrain data with information such as elevation, slope and aspect, such as accessing a geospatial data cloud to obtain a digital elevation model, and use it as the terrain data.

[0070] S2. Processing the optical remote sensing image data of the historical period to obtain multiple sample images of the target area with different dimensional information.

[0071] In step S2, specifically, the target area can be, but is not limited to, Muchuan County, Sichuan Province. The county is located in the southeastern edge of the Sichuan Basin, with an area of ​​1,407 square kilometers. The county is mainly composed of low mountains and hills, with a forest coverage rate of more than 77%, forming rich species diversity and ecosystem diversity. As a national key ecological functional area, Muchuan County is distributed with national forest parks and provincial nature reserves and other nature reserves. It is an important area for biodiversity conservation in the upper reaches of the Yangtze River. However, its existing biodiversity indicators (such as vegetation coverage and species distribution hotspots) are mainly obtained by traditional ground surveys, which face challenges such as time-consuming and labor-intensive and delayed data updates. It is urgent to comprehensively use optical remote sensing image data and deep learning technology to automatically extract indicators such as vegetation coverage and aboveground biomass to improve the information level and accuracy of biodiversity monitoring, so as to not only directly serve Muchuan's "ecological priority" strategic needs, but also provide a portable solution for similar mountain ecological protection.

[0072] In the step S2, specifically, a plurality of sample images of the target area with different dimensional information are obtained based on the optical remote sensing image data of the historical period, including but not limited to the following steps S21 to S25.

[0073] S21. Extract blue band spectral images, green band spectral images, red band spectral images, near infrared band spectral images, and short wave infrared band spectral images of the target area from the optical remote sensing image data of the historical period.

[0074] In step S21, since the wavelength range of the blue band is 0.45-0.52 μm, it can be used for the identification and analysis of specific landforms such as water bodies, vegetation and soil; the wavelength range of the green band is 0.52-0.59 μm, which is suitable for applications such as vegetation monitoring and land cover classification; the wavelength range of the red band is 0.63-0.69 μm, which is commonly used in fields such as vegetation health assessment and land use monitoring; the wavelength range of the near-infrared band is 0.77-0.89 μm, which is suitable for analysis of vegetation growth status and land cover type; the wavelength range of the short-wave infrared band is 1.4-3 μm, which is commonly used for vegetation monitoring, land use classification and soil property analysis; therefore, these selected bands have good spectral response characteristics in vegetation and ecological environment monitoring, can effectively capture changes in ecological parameters such as vegetation growth status, moisture content and structural differences, and provide a rich spectral feature basis for model regression prediction of biodiversity indicators.

[0075] S22. Based on the red band spectral image and the near-infrared band spectral image of the target area, a normalized vegetation index distribution image of the target area is constructed.

[0076] S23. Based on the blue band spectral image, the red band spectral image and the near-infrared band spectral image of the target area, an enhanced vegetation index distribution image of the target area is constructed.

[0077] S24. Based on the near-infrared band spectral image and the red band spectral image of the target area, a ratio vegetation coefficient distribution image of the target area is constructed.

[0078] In the aforementioned steps S22 to S24, the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), and the Ratio Vegetation Index (RVI) are existing commonly used vegetation indices, which are used to supplement the information of the original band in terms of ecological response (i.e., NDVI can reflect vegetation density by using the difference between red and near-infrared reflectance; EVI can have stronger sensitivity in high-density vegetation areas, and RVI is more suitable for structural enhancement under sparse vegetation conditions) so as to enhance the model's ability to distinguish vegetation conditions.

[0079] S25. Use the blue band spectral image, green band spectral image, red band spectral image, near infrared band spectral image, short wave infrared band spectral image, normalized vegetation index distribution image, enhanced vegetation index distribution image and ratio vegetation coefficient distribution image of the target area as multiple sample images of the target area with different dimensional information.

[0080] Based on the aforementioned steps S21 to S25, eight-channel sample images with different dimensional information can be obtained. In order to cover the entire target area and ensure temporal and spatial continuity, this embodiment can also select four adjacent remote sensing image data of the target area from the optical remote sensing image data of the historical period, and after registration and splicing, the image is cropped according to the 2km buffer zone extending outward from the boundary of the target area to construct the spatial sample area required for model training; thus, this buffering strategy not only ensures the information integrity of the boundary area, but also helps the model to make stable predictions in the spatial edge area. In addition, the elevation distribution image, slope distribution image, and slope distribution image of the target area can be extracted from the terrain data and added to the multiple sample images to further enrich the sample information.

[0081] S3. Extracting a biodiversity indicator distribution image of the target area from the ecological thematic product data of the historical period, and registering the biodiversity indicator distribution image with the sample image to obtain a label image of the target area.

[0082] In step S3, the biodiversity indicator distribution image specifically includes, but is not limited to, a vegetation cover distribution image and / or an aboveground biomass distribution image. The specific process of registering the biodiversity indicator distribution image with the sample image may include, but is not limited to, using ArcGIS software to spatially crop and perform cubic spline interpolation on the biodiversity indicator distribution image, and then uniformly projecting it into the WGS1984 UTM Zone 48N coordinate system to ensure spatial consistency and matching accuracy between the sample image and the labeled image.

[0083] S4. Based on the multiple sample images and the label image, sample data is obtained in the following manner: a sub-region label image containing m×m label pixels is extracted from the label image as a model output item, and then a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image is extracted from each sample image to be aggregated as a model input item, and finally the model input item and the model output item are used as a sample data item, wherein m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, ηl is the resolution of the label image and η s multiples of .

[0084] In step S4, since sub-region sample images containing n×n sample pixels and covering the same area as the sub-region label image are extracted from each sample image and summarized as model input items, when the multiple sample images include the blue band spectral image, green band spectral image, red band spectral image, near infrared band spectral image, short wave infrared band spectral image, normalized vegetation index distribution image, enhanced vegetation index distribution image and ratio vegetation coefficient distribution image of the target area, the model input items include eight sub-region sample images located on different channels. Considering that the sample image and the label image generally have different resolutions, if one of the images is directly upsampled to the resolution of the other image (for example, the resolution of the sample image is 10m and the resolution of the label image is 30m, if the label image is directly upsampled to 10m), spatial noise and error are easily introduced. Therefore, this embodiment adopts the above-mentioned matching strategy of "unchanged labels and sample aggregation", that is, when the resolution of the sample image is 10m and the resolution of the label image is 30m, each label pixel with a resolution of 30m corresponds to 3×3 sample pixels with a resolution of 10m in its coverage area, and the samples in the area are uniformly assigned the same label value to participate in training (this method is essentially a fusion solution of "low-resolution labels and high-resolution features", which not only retains the original accuracy of the labels, but also avoids unnecessary interpolation errors, thereby improving the stability of the training data and the speed of model convergence).

[0085] In step S4, in order to maintain the spatial integrity and geographical consistency of the training samples, this embodiment adopts a seamless sliding window sampling strategy to systematically crop the entire target area to obtain a sample data set that can ensure that the model learning covers all landforms and ecological types. That is, for Muchuan County, a total of 20,236 non-overlapping sub-areas can be obtained through a seamless sliding window method, each sub-area is a rectangular plot of 300m×300m, and the resolution of the sample image is 10m. Therefore, each non-overlapping sub-area contains 900 (i.e., 30×30, n is 30) input pixels, and the corresponding label image sub-area is 100 (i.e., 10×10, m is 10) output pixels, thereby achieving a spatial pairing relationship where every 9 input pixels correspond to one output pixel.

[0086] Based on the aforementioned steps S1 to S4, a deep learning training sample set with clear structure, rich information and consistent time and space can be constructed, providing a reliable data foundation for the effective training and generalization of subsequent models.

[0087] S5. Divide all the sample data into a training set, a validation set, and a test set, and apply the division results to train a lightweight convolutional neural network model based on the ConvNeXt architecture to obtain a labeled image generation model.

[0088] In step S5, the ConvNeXt architecture is a convolutional neural network architecture proposed by the Facebook AI Research (FAIR) team in 2022. It integrates the design concept of the Transformer architecture while maintaining the advantages of convolutional neural networks in local feature modeling and computational efficiency. The architecture is based on the classic ResNet and retains its residual connection structure to ensure that the gradient in the deep network can be stably transmitted. At the same time, multiple optimizations are performed on the traditional convolution module to improve performance. In this embodiment, in order to achieve accurate regression analysis of surface biodiversity indicators in remote sensing images, a lightweight convolutional neural network model is constructed and trained: a label image generation model.

[0089] In step S5, the ConvNeXt architecture used in this embodiment is as follows: Figure 2 As shown, the architecture is improved based on the existing ConvNeXt architecture so as to combine the distribution law and spatial structure characteristics of the multi-band features of remote sensing images and specifically optimize the receptive field, nonlinear expression ability and training stability of the network. Specifically, preferably, the ConvNeXt architecture adopts a ConvNeXt variant architecture having but not limited to the following features (A) to (B): (A) the input layer is used to input multiple sub-region sample images corresponding one-to-one to the multiple sample images, and the output layer is used to output sub-region label images covering the same area as the multiple sub-region sample images and used to reflect the distribution of biodiversity indicators; (B) the network structure is composed of multiple stacked ConvNeXt blocks, and the internal structure of each ConvNeXt block includes but is not limited to a depthwise separable convolution module, a layer normalization module, a GELU activation function module, a Dropout layer module and a residual connection module, wherein the depthwise separable convolution module adopts a depthwise separable convolution kernel of size 7×7, and first performs a convolution operation independently in each image channel, and then realizes cross-channel information fusion through 1×1 point-by-point convolution. The ConvNeXtblock is the core computing unit of the entire network, and its detailed structure is as follows: Figure 3As shown, the depthwise separable convolution module uses a 7×7 depthwise separable convolution kernel and first performs independent convolution operations within each image channel, followed by cross-channel information fusion through 1×1 pointwise convolution. This improves the network's ability to model spatial structure while significantly reducing the number of parameters and computational overhead, making it suitable for large-scale image regression tasks in remote sensing scenarios. By replacing traditional batch normalization with layer normalization in the ConvNeXt block, the model's numerical stability during small-batch training is improved and its compatibility with Transformer-like networks is further enhanced. The use of the GELU (Gaussian Error Linear Unit) activation function, which has superior gradient properties, helps improve nonlinear modeling capabilities. Furthermore, the introduction of a Dropout regularization mechanism (set to 0.5) effectively mitigates overfitting. Based on the aforementioned ConvNeXt variant architecture design, the model significantly reduces the number of parameters and computational resource consumption while retaining sufficient expressive power, making it more suitable for large-scale scenarios such as regional remote sensing prediction.

[0090] In step S5, the specific input and output of the model training are: taking the multiple sub-region sample images located on different channels as input, extracting high-order semantic features through a multi-level residual convolution module and a downsampling module, and finally outputting continuous variable prediction values ​​through a 1×1 convolution: vegetation cover (FVC) distribution image and / or aboveground biomass (AGB) distribution image, so that the overall structure of the final model takes into account parameter compactness and computational efficiency while ensuring regression performance, and is suitable for rapid training and deployment under small and medium sample conditions. In order to achieve efficient training and generalization performance control of the model, this embodiment also develops a set of systematic training strategies: (1) In terms of data processing, the training, validation and test samples are divided according to a fixed ratio (for example, 6:2:2) and constructed into batch inputs through the data loader DataLoader. The training process adopts a grouping method with a batch size of 32; (2) In terms of loss function, the model optimization goal is the accurate regression of multi-output continuous variables. The mean squared error loss function (MSE) is selected as the loss function, which has good numerical stability and convexity characteristics in continuous variable regression; (3) The optimizer adopts the Adam algorithm, and the initial value of the learning rate is set to 0.0001, which has good adaptability and convergence efficiency; (4) In order to improve the adaptive adjustment ability of the learning rate during training, the ReduceLROnPlateau learning rate scheduler is introduced: when the validation set loss fails to decrease within 5 consecutive epochs, the learning rate is automatically decayed to the original 0.5 to avoid model oscillation or falling into local optimality due to excessively high learning rate; at the same time, the Early The stopping mechanism tolerates no improvement in training loss within 20 epochs, which automatically terminates the training to avoid overfitting; (5) Finally, the model parameters with the smallest error in the validation set are retained as the final model file.

[0091] S6. Processing the optical remote sensing image data of the target period to obtain the target area and the multiple new sample images.

[0092] In step S6, the target period can be any historical period or the current period. The specific acquisition and processing of the optical remote sensing image data of the target period is consistent with the above steps S1 to S2 and will not be repeated here.

[0093] S7. For each sub-region in the target region, extract a sub-region sample image containing n×n sample pixels and corresponding to each new sample image of the target region to summarize and input the label image generation model, and output the corresponding sub-region label image.

[0094] In step S7, in order to reduce edge artifacts (i.e., "blocking effects") caused by splicing, this embodiment also designs a post-processing strategy based on overlapping sliding prediction and Gaussian weighted fusion. That is, unlike the non-overlapping sampling in the training phase, each sub-region is obtained by sliding window processing the target region according to a sliding window overlap rate of δ%, where δ represents a positive integer greater than or equal to 30 and less than or equal to 70. For example, when δ is 50, for the Muchuan County area, a total of 83,737 overlapping sub-regions can be obtained by a 50% overlapping sliding window method, and for each overlapping sub-region, a sub-region sample image containing n×n sample pixels and corresponding to each new sample image is extracted from the target region and each new sample image to be summarized and input into the label image generation model, and finally the corresponding sub-region label image is obtained.

[0095] S8. Generate a biodiversity indicator distribution image of the target area during the target period by stitching the sub-area label images of the sub-areas.

[0096] In step S8, the post-processing strategy based on overlapping sliding prediction and Gaussian weighted fusion is continued. Preferably, the biodiversity indicator distribution image of the target area in the target period is generated by splicing based on the sub-area label images of the sub-areas, including but not limited to the following steps S81 to S82.

[0097] S81. For each pixel in the biodiversity indicator distribution image to be generated, the corresponding pixel value is calculated according to the sub-region label image of each sub-region according to the following formula:

[0098]

[0099] In the formula, i represents a positive integer, P i represents the pixel value of the i-th pixel in the biodiversity indicator distribution image to be generated, K i represents the total number of times the i-th pixel is covered by the sub-region label images of each sub-region, p i,k represents the pixel value of the i-th pixel in the sub-region label image when it is covered for the kth time, w i,k Represents the weight coefficient corresponding to the position of the i-th pixel in the sub-region label image when it is covered for the kth time in a two-dimensional Gaussian weight matrix of size m×m, wherein the two-dimensional Gaussian weight matrix has the characteristics that the weight is the largest at the center, the weight gradually decreases toward the edge, and the standard deviation σ=1.5 is used to control the width of the weight distribution.

[0100] In step S81, for example, if m is 10 and δ is 50, then Figure 4As shown in the figure, the i-th pixel indicated by i will be covered by the sub-region label images of the sub-regions at least 4 times (that is, K is 4): when the i-th pixel is covered for the first time and is located at the seventh row and seventh column in the sub-region label image used for the covering (at this time w i,k Take the weight coefficient located in the seventh row and seventh column in the two-dimensional Gaussian weight matrix), when the i-th pixel is covered for the second time and is located in the seventh row and second column in the sub-region label image used for the covering (at this time w i,k Take the weight coefficient located in the seventh row and second column in the two-dimensional Gaussian weight matrix), when the i-th pixel is covered for the third time and is located in the second row and seventh column in the sub-region label image used for the covering (at this time w i,k Take the weight coefficient located in the second row and seventh column in the two-dimensional Gaussian weight matrix) and the position of the i-th pixel when it is covered for the fourth time and in the sub-region label image used for this covering is the second row and second column (at this time w i,k Take the weight coefficient located in the second row and second column of the two-dimensional Gaussian weight matrix). In addition, each weight coefficient in the two-dimensional Gaussian weight matrix can be calculated according to the following formula:

[0101]

[0102] In the formula, m′ and m″ represent positive integers less than or equal to m, w m′,m″ represents the weight coefficient located at the m′th row and the m″th column in the two-dimensional Gaussian weight matrix.

[0103] S82. Generate the biodiversity indicator distribution image of the target area in the target period according to the pixel value of each pixel.

[0104] Based on the above steps S81 to S82, not only can the biodiversity indicator extraction results (i.e., the biodiversity indicator distribution image of the target area during the target period) with both spatial continuity and numerical consistency be obtained, providing reliable basic data for the ecological analysis of the target area, but it can also effectively reduce the impact of the decreased accuracy of the indicator distribution image generation in the edge area, while smoothing the transition of the indicator distribution between adjacent areas to avoid the occurrence of breaks or mutations. In addition, it is also considered that the input data is derived from multi-phase and / or multi-platform images, and there are inevitably invalid pixel areas, such as shadows, cloud occlusions, invalid edge values, or resampling holes. Therefore, if these pixels are directly used for model training and evaluation, it will seriously affect the stability of the extraction results and the interpretability of the indicators. Preferably, the method also includes but is not limited to: using a mask mechanism to eliminate invalid areas in the input image, retaining only valid pixels for model training and application. The specific implementation process is as follows: for each sample image, a four-dimensional mask tensor mask_Y is constructed, whose dimensions are consistent with the target label (i.e., number of samples × number of channels × height × width), where the value True indicates that the pixel is invalid; during the model training and prediction stage, the corresponding mask block is dynamically extracted in each batch, and valid pixels are filtered out by culling to perform point-by-point regression prediction and error evaluation on them, thereby ensuring the effectiveness and accuracy of training gradients and indicator statistics.

[0105] This embodiment also performs the following effect analysis on the biodiversity index extraction results obtained based on the aforementioned steps S1 to S8:

[0106] (1) In model evaluation, this embodiment adopts a multi-index joint evaluation method to comprehensively characterize the generation performance of the model: Mean Absolute Error (MAE), which is used to reflect the average deviation between the model prediction value and the true value. The smaller the value, the lower the overall error; Mean Squared Error (MSE), which is used to be more sensitive to large errors and emphasizes the stability of the model in extreme value prediction; Root Mean Squared Error (RMSE), which is used as the square root of MSE, which is convenient for comparison with actual units and has more intuitive meaning; Coefficient of Determination (R 2) is used to assess the model's ability to explain fluctuations in the target variable; the closer it is to 1, the better the model fit. The Pearson correlation coefficient (r) measures the linear correlation between the predicted and true values, emphasizing trend consistency and a commonly used quantitative consistency indicator in ecological remote sensing. These evaluation metrics are independently calculated on the training, validation, and test sets, and ultimately used to evaluate the model's accuracy in fitting fractional vegetation cover (FVC) and aboveground biomass (AGB) at different stages. Because the target area spans multiple eco-geomorphological types and exhibits significant spatial heterogeneity, the aforementioned multi-metric approach can more comprehensively characterize the model's robustness and generalizability, providing a scientific basis for subsequent spatial prediction and application promotion.

[0107] (2) Distribution image generation results: The vegetation coverage distribution image generation results of Muchuan County in 2020 and 2021 are as follows Figure 5 As shown in the figure, the aboveground biomass distribution image generation results of Muchuan County in 2020 and 2021 are as follows Figure 6 shown.

[0108] (3) The vegetation cover model sample set loss indicators for Muchuan County in 2020 and 2021 are shown in Table 1, and the aboveground biomass sample set loss indicators for Muchuan County in 2020 and 2021 are shown in Table 2:

[0109] Table 1. Vegetation cover model sample set loss indicators in Muchuan County in 2020 and 2021

[0110]

[0111] Table 2. Loss indicators of aboveground biomass sample sets in Muchuan County in 2020 and 2021

[0112]

[0113] Judging from the accuracy of the model evaluation in the above two years, the model losses of vegetation cover (FVC) and aboveground biomass (AGB) in the two years were relatively close, indicating that the model has good generalization and the losses in both years were low. This shows that the model has good performance in the extraction tasks of vegetation cover (FVC) and aboveground biomass (AGB) in Muchuan County, providing good data support for the extraction results, providing stable and effective extraction results for the estimation of vegetation cover (FVC) and aboveground biomass (AGB), and providing a reference for the automatic extraction of biodiversity indicator information in Muchuan County.

[0114] (4) The distribution of the predicted values ​​of fractional vegetation cover (FVC) in 2020 in the experiment and the reference values ​​of the public dataset is as follows: Figure 7 As shown in the figure, the distribution of the predicted values ​​of vegetation coverage (FVC) in 2021 and the reference values ​​of the public dataset are as follows: Figure 8 As shown in the figure, the distribution of the predicted aboveground biomass (AGB) in 2020 and the reference values ​​of the public dataset is as follows: Figure 9 As shown in the figure, the distribution of the aboveground biomass (AGB) forecast values ​​in 2021 and the reference values ​​of the public datasets are as follows: Figure 10 shown.

[0115] (4.1) Prediction results of fractional vegetation cover (FVC) in 2020 and 2021: In the fractional vegetation cover (FVC) prediction tasks in 2020 and 2021, the model performed well overall, which is specifically reflected in the various regression evaluation indicators (Table 1). The mean absolute error (MAE) is around 0.08, indicating that the average deviation of the model in the fractional vegetation cover (FVC) prediction is small, and it can be relatively stable and close to the reference value of the public dataset; the mean square error (MSE) remains at around 0.02, further indicating that the model has a strong ability to control large errors; and the root mean square error (RMSE) remains at around 0.12, confirming the reliability of the model in overall accuracy. These error indicators are all at a low level, reflecting that the model has a small prediction error for FVC at the pixel scale. In addition, the coefficient of determination R 2 It reached about 0.88, which means that the model can explain about 88% of the changes in vegetation cover (FVC), indicating that it fits the relationship between input features and vegetation cover (FVC) well. The Pearson correlation coefficient r reached 0.94, indicating that there is a very strong positive correlation between the prediction results and the reference values ​​of the public dataset, that is, the numerical trend of the prediction curve and the true curve is highly consistent. And from the visualization results, in 2020 ( Figure 7 ) and 2021 ( Figure 8 )'s vegetation cover (FVC) prediction curves can closely follow the fluctuations of the reference vegetation cover (FVC) curve of the public dataset, showing that the model has good generalization ability and temporal stability in different years, and is consistent in the overall trend. The predictions for high values ​​in 2020 and 2021 are somewhat lacking compared to the reference values ​​of the public dataset, and are more concentrated on the mean.

[0116] (4.2) Forecast results of aboveground biomass (AGB) in 2020 and 2021: In terms of the aboveground biomass (AGB) model loss index in 2020 and 2021 (Table 2), the error index of the model is generally at a low level. Specifically, the mean absolute error (MAE) is maintained at around 0.04, the mean square error (MSE) is around 0.01, and the root mean square error (RMSE) is about 0.07, which shows that the model's prediction error for aboveground biomass (AGB) at the numerical level is relatively small and has a certain degree of accuracy. From the perspective of goodness of fit, the model's coefficient of determination R 2The Pearson correlation coefficient r is around 0.2, indicating that the model has limited ability to explain the total variability of aboveground biomass (AGB). As can be seen from the visualization results, the predicted value in 2020 ( Figure 9 ) and the forecast value for 2021 ( Figure 10 ), the model lacks response to some extreme or drastic changes in regions, but its prediction curve can still roughly follow the main trend of the reference aboveground biomass (AGB) of the public dataset in most regions, especially in the medium value range where the reference aboveground biomass (AGB) of the public dataset is more densely distributed, the model's prediction values ​​are significantly concentrated and comparable to the reference values ​​of the public dataset.

[0117] Based on the biodiversity indicator extraction method described in steps S1 to S8 above, a new biodiversity indicator extraction scheme based on deep learning and remote sensing imagery is provided. Specifically, based on historical optical remote sensing image data and ecological thematic product data, a labeled image of the target area and multiple sample images with different dimensional information are processed. Sample data is then obtained based on these images, and a lightweight convolutional neural network model based on the ConvNeXt architecture is trained using these sample data to obtain a labeled image generation model. Finally, for each sub-region in the target area, corresponding sub-region sample images are extracted from each new sample image of the target area to be aggregated and input into the labeled image generation model to obtain corresponding sub-region labeled images. These images are then aggregated and spliced ​​to generate a biodiversity indicator distribution image for the target area during the target period. By integrating optical remote sensing image data and ecological thematic product data and constructing a deep learning model based on a convolutional neural network that fuses attribute features and spatial features, biodiversity indicators can be automatically regressed and extracted. This reduces the difficulty of data labeling, improves model generalization, and enhances the accuracy and efficiency of biodiversity indicator extraction. This provides a scalable new technical path for mountain biodiversity monitoring, facilitating practical application and promotion.

[0118] like Figure 11 As shown, the second aspect of this embodiment provides a virtual device for implementing the biodiversity indicator extraction method described in the first aspect, comprising a multi-source data acquisition unit, a sample image processing unit, a label image processing unit, a sample data acquisition unit, a generation model training unit, a generation model application unit, and a distribution image stitching unit;

[0119] The multi-source data acquisition unit is used to acquire optical remote sensing image data and ecological thematic product data of historical periods, wherein the ecological thematic product data includes distribution images of biodiversity indicators;

[0120] The sample image processing unit is communicatively connected to the multi-source data acquisition unit and is used to process the optical remote sensing image data of the historical period to obtain a plurality of sample images of the target area with different dimensional information;

[0121] The label image processing unit is communicatively connected to the multi-source data acquisition unit and is used to extract a biodiversity indicator distribution image of the target area from the ecological thematic product data of the historical period, and register the biodiversity indicator distribution image with the sample image to obtain a label image of the target area;

[0122] The sample data acquisition unit is connected to the sample image processing unit and the label image processing unit in a distributed communication manner, and is used to acquire sample data based on multiple sample images and label images in the following manner: extracting a sub-region label image containing m×m label pixels from the label image as a model output item, and then extracting a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image from each sample image to summarize as a model input item, and finally using the model input item and the model output item as a piece of sample data, wherein m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, η l is the resolution of the label image and η s multiples of;

[0123] The generation model training unit is communicatively connected to the sample data acquisition unit, and is used to divide all sample data into a training set, a validation set, and a test set, and use the division results to perform model training on a lightweight convolutional neural network model based on the ConvNeXt architecture to obtain a label image generation model;

[0124] The sample image processing unit is further used to obtain a target area and multiple new sample images based on the optical remote sensing image data of the target period;

[0125] The generation model application unit is connected to the generation model training unit and the sample image processing unit in a distributed communication manner, and is used to extract, for each sub-region in the target region, a sub-region sample image containing n×n sample pixels and corresponding to each new sample image of the target region to summarize the input label image generation model, and output the corresponding sub-region label image;

[0126] The distribution image splicing unit is communicatively connected to the generation model application unit, and is used to splice and generate a biodiversity indicator distribution image of the target area in the target period based on the sub-area label images of each sub-area.

[0127] The working process, working details and technical effects of the aforementioned device provided in the second aspect of this embodiment can be found in the biodiversity indicator extraction method described in the first aspect, and will not be described in detail here.

[0128] like Figure 12 As shown, a third aspect of this embodiment provides a computer device for executing the biodiversity indicator extraction method described in the first aspect, comprising a memory, a processor, and a transceiver communicatively connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program to execute the biodiversity indicator extraction method described in the first aspect or the first embodiment. Specifically, the memory may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), flash memory, first-input first-output (FIFO), and / or first-input last-output (FILO) memory; the processor may include, but is not limited to, a microprocessor from the STM32F105 series. Furthermore, the computer device may include, but is not limited to, a power module, a display screen, and other necessary components.

[0129] The working process, working details and technical effects of the aforementioned computer device provided in the third aspect of this embodiment can be found in the biodiversity indicator extraction method described in the first aspect, and will not be described in detail here.

[0130] A fourth aspect of this embodiment provides a computer-readable storage medium storing instructions including the method for extracting biodiversity indicators as described in the first aspect. Specifically, the computer-readable storage medium stores instructions that, when executed on a computer, execute the method for extracting biodiversity indicators as described in the first aspect. The computer-readable storage medium refers to a medium for storing data, and may include, but is not limited to, a floppy disk, a CD, a hard disk, a flash memory, a USB flash drive, and / or a memory stick. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0131] The working process, working details and technical effects of the aforementioned computer-readable storage medium provided in the fourth aspect of this embodiment can be found in the biodiversity indicator extraction method described in the first aspect, and will not be described in detail here.

[0132] A fifth aspect of this embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, implements the method for extracting biodiversity indicators as described in the first aspect. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0133] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for extracting biodiversity indicators, characterized in that: include: Acquire historical optical remote sensing image data and ecological thematic product data, where the ecological thematic product data includes distribution images of biodiversity indicators; Based on the optical remote sensing image data processing of the historical period, multiple sample images of the target area with different dimensional information are obtained; Extract the biodiversity indicator distribution image of the target area from the historical ecological thematic product data, and align the biodiversity indicator distribution image with the sample image to obtain the label image of the target area; Based on multiple sample images and label images, sample data is obtained as follows: a sub-region label image containing m×m label pixels is extracted from the label image as the model output item, and then a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image is extracted from each sample image to be summarized as the model input item, and finally the model input item and the model output item are used as a sample data, where m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, η l is the resolution of the label image and η s multiples of; All sample data are divided into training set, validation set and test set, and the division results are used to train the lightweight convolutional neural network model based on the ConvNeXt architecture to obtain the labeled image generation model; Based on the optical remote sensing image data processing of the target period, multiple new sample images of the target area are obtained; For each sub-region in the target area, extract the corresponding sub-region sample image containing n×n sample pixels from each new sample image in the target area to summarize the input label image generation model, and output the corresponding sub-region label image; Based on the sub-region label images of each sub-region, the biodiversity indicator distribution image of the target area in the target period is generated by splicing.

2. The method for extracting biodiversity indicators according to claim 1, characterized in that: The biodiversity indicators include vegetation cover and / or aboveground biomass.

3. The method for extracting biodiversity indicators according to claim 1, wherein: Based on the historical optical remote sensing image data processing, multiple sample images of the target area with different dimensional information are obtained, including: extracting a blue band spectral image, a green band spectral image, a red band spectral image, a near infrared band spectral image, and a short-wave infrared band spectral image of the target area from the optical remote sensing image data of the historical period; Based on the red band spectral image and the near infrared band spectral image of the target area, a normalized vegetation index distribution image of the target area is constructed; Based on the blue band spectral image, the red band spectral image and the near infrared band spectral image of the target area, an enhanced vegetation index distribution image of the target area is constructed; Based on the near-infrared band spectral image and the red band spectral image of the target area, a ratio vegetation coefficient distribution image of the target area is constructed; The blue band spectral image, green band spectral image, red band spectral image, near infrared band spectral image, short wave infrared band spectral image, normalized vegetation index distribution image, enhanced vegetation index distribution image and ratio vegetation coefficient distribution image of the target area are used as multiple sample images of the target area with different dimensional information.

4. The method for extracting biodiversity indicators according to claim 1, wherein: The ConvNeXt architecture adopts a ConvNeXt variant architecture with the following features (A) to (B): (A) the input layer is used to input a plurality of sub-region sample images corresponding one-to-one to the plurality of sample images, and the output layer is used to output a sub-region label image that covers the same area as the plurality of sub-region sample images and is used to reflect the distribution of biodiversity indicators; (B) The network structure is composed of multiple stacked ConvNeXt blocks, and the internal structure of each ConvNeXt block includes a depth-wise separable convolution module, a layer normalization module, a GELU activation function module, a Dropout layer module and a residual connection module. The depth-wise separable convolution module uses a depth-wise separable convolution kernel of size 7×7, and first performs convolution operations independently in each image channel, and then realizes cross-channel information fusion through 1×1 point-by-point convolution.

5. The method for extracting biodiversity indicators according to claim 1, wherein: Each sub-region is obtained by performing sliding window processing on the target region according to a sliding window overlap rate of δ%, where δ represents a positive integer greater than or equal to 30 and less than or equal to 70; Based on the sub-region label images of each sub-region, a biodiversity indicator distribution image of the target region during the target period is generated by stitching, including: For each pixel in the biodiversity indicator distribution image to be generated, the corresponding pixel value is calculated according to the sub-region label image of each sub-region using the following formula: In the formula, i represents a positive integer, P i represents the pixel value of the i-th pixel in the biodiversity indicator distribution image to be generated, K i represents the total number of times the i-th pixel is covered by the sub-region label images of each sub-region, p i,k represents the pixel value of the i-th pixel in the sub-region label image when it is covered for the kth time, w i,k represents the weight coefficient corresponding to the position of the i-th pixel in the sub-region label image used for the k-th coverage in a two-dimensional Gaussian weight matrix of size m×m, wherein the two-dimensional Gaussian weight matrix has the characteristics of having the largest weight at the center, gradually decreasing weight toward the edge, and using a standard deviation σ=1.5 to control the width of the weight distribution; The biodiversity indicator distribution image of the target area in the target period is generated according to the pixel value of each pixel.

6. The method for extracting biodiversity indicators according to claim 1, characterized in that: The method also includes: using a mask mechanism to eliminate invalid areas in the input image, retaining only valid pixels for model training and application.

7. A biodiversity index extraction device, characterized in that: It includes a multi-source data acquisition unit, a sample image processing unit, a label image processing unit, a sample data acquisition unit, a generation model training unit, a generation model application unit and a distribution image stitching unit; The multi-source data acquisition unit is used to acquire optical remote sensing image data and ecological thematic product data of historical periods, wherein the ecological thematic product data includes distribution images of biodiversity indicators; The sample image processing unit is communicatively connected to the multi-source data acquisition unit and is used to process the optical remote sensing image data of the historical period to obtain a plurality of sample images of the target area with different dimensional information; The label image processing unit is communicatively connected to the multi-source data acquisition unit and is used to extract a biodiversity indicator distribution image of the target area from the ecological thematic product data of the historical period, and register the biodiversity indicator distribution image with the sample image to obtain a label image of the target area; The sample data acquisition unit is connected to the sample image processing unit and the label image processing unit in a distributed communication manner, and is used to acquire sample data based on multiple sample images and label images in the following manner: extracting a sub-region label image containing m×m label pixels from the label image as a model output item, and then extracting a sub-region sample image containing n×n sample pixels and covering the same area as the sub-region label image from each sample image to summarize as a model input item, and finally using the model input item and the model output item as a piece of sample data, wherein m is a positive integer, n=m×(η l ÷η s ), η s is the resolution of the sample image, η l is the resolution of the label image and η s multiples of; The generation model training unit is communicatively connected to the sample data acquisition unit, and is used to divide all sample data into a training set, a validation set, and a test set, and use the division results to perform model training on a lightweight convolutional neural network model based on the ConvNeXt architecture to obtain a label image generation model; The sample image processing unit is further used to obtain a target area and multiple new sample images based on the optical remote sensing image data of the target period; The generation model application unit is connected to the generation model training unit and the sample image processing unit in a distributed communication manner, and is used to extract, for each sub-region in the target region, a sub-region sample image containing n×n sample pixels and corresponding to each new sample image of the target region to summarize the input label image generation model, and output the corresponding sub-region label image; The distribution image splicing unit is communicatively connected to the generation model application unit, and is used to splice and generate a biodiversity indicator distribution image of the target area in the target period based on the sub-area label images of each sub-area.

8. A computer device, characterized in that: The method comprises a memory, a processor and a transceiver which are communicatively connected in sequence, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the biodiversity indicator extraction method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the biodiversity indicator extraction method according to any one of claims 1 to 6 is executed.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or the instruction is executed by a computer, the method for extracting biodiversity indicators according to any one of claims 1 to 6 is implemented.

Images