Biodiversity monitoring method for natural reserve
By combining hyperspectral imaging sensors with illumination correction and depth feature modeling, the problems of spectral distortion and classification accuracy in biodiversity monitoring in nature reserves have been solved, enabling refined classification and dynamic monitoring of spectrally similar species.
Patent Information
- Application Number
- CN202511488706.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-01-09
AI Technical Summary
Existing technologies for biodiversity monitoring in nature reserves suffer from problems such as spectral distortion caused by changes in light conditions, band redundancy and decreased classification accuracy, as well as insufficient ability to model complex biological spectral phenomena, especially the difficulty in accurately distinguishing species with similar spectra.
Data acquisition is performed using a hyperspectral imaging sensor. Combined with illumination correction, band selection, and depth feature modeling methods, an illumination invariance correction function, a biospectral fingerprint matrix, and a multi-scale convolutional network are constructed. By fusing biological attribute matrices, feature extraction and classification are performed. Refined species classification is achieved by utilizing adaptive illumination correction, band selection, and biological knowledge constraints.
It effectively eliminates spectral distortion, preserves key biological characteristics, improves classification accuracy, and can accurately distinguish spectrally similar species, enabling long-term dynamic biodiversity monitoring in nature reserves.
Smart Images
Figure CN121305352A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and biological monitoring technology, and in particular to a method for monitoring biodiversity in nature reserves. Background Technology
[0002] With the increasing emphasis on ecological and environmental protection, biodiversity monitoring in nature reserves has become a core research direction in ecology and remote sensing. Conventional monitoring methods mainly rely on manual quadrat surveys and species statistics. While these methods provide relatively accurate species information, their spatial coverage is limited, and they are time-consuming and labor-intensive, making it difficult to achieve large-scale, continuous dynamic monitoring. In recent years, the development of hyperspectral remote sensing technology has provided a new means for biodiversity monitoring. Hyperspectral imagery has high spectral resolution and can capture the reflectance characteristics of ground features across hundreds of consecutive bands, making it widely used for vegetation classification and species identification.
[0003] Existing technologies have the following shortcomings in practical applications: changes in lighting conditions (such as differences between shady and sunny slopes, and the influence of atmospheric scattering) can cause significant distortions in the spectral responses of similar ground features under different environments, and conventional normalization methods are unable to eliminate this nonlinear bias; hyperspectral data has numerous bands and is highly redundant, and commonly used dimensionality reduction methods such as principal component analysis often destroy key structures unique to biological spectra, such as absorption peaks and red edges, leading to a decrease in classification accuracy; conventional convolutional neural networks are limited to single-scale convolution in hyperspectral data processing, making it difficult to take into account both local details and global spectral features, and their ability to model complex biological spectral phenomena is insufficient; existing methods generally rely solely on spectral differences for classification, which can easily lead to confusion for closely related species with highly similar spectra, and lacks a mechanism for discrimination based on prior biological knowledge.
[0004] Therefore, this invention proposes a biodiversity monitoring method for nature reserves to address the above-mentioned problems. Summary of the Invention
[0005] This invention addresses the shortcomings of existing technologies by developing a biodiversity monitoring method for nature reserves. By proposing a comprehensive approach that takes into account illumination correction, band selection, depth feature modeling, and biological knowledge constraints, this invention enables refined classification of species in nature reserves and long-term dynamic biodiversity monitoring.
[0006] The technical solution of this invention to solve the technical problem is a method for monitoring biodiversity in nature reserves, comprising the following steps: S1. Use a hyperspectral imaging sensor to collect remote sensing data of nature reserves and label the collected data with biological species types; S2. Construct an illumination invariance correction function, and use a solar irradiance and shadow compensation mechanism to dynamically adjust the collected data and generate a correction spectral vector; S3. Based on the collected and calibrated data, construct a biospectral fingerprint database matrix, calculate the curvature characteristics and mutual information of each band, fuse the two to calculate the comprehensive discrimination score, select the band with the highest score to form a discriminative band subset, and then obtain the band selection spectral data according to the band number index. S4. Construct a deep feature extraction model to predict species types. Extract local and global features through multi-scale convolution and dilated convolution and fuse them. Generate knowledge correction terms based on the biological attribute matrix, calculate the loss function, and update the model parameters through an optimizer to obtain a trained model. S5. After performing steps S2 and S3 on the newly collected data, input it into the trained model to obtain the final classification result, then generate a species distribution map and calculate the diversity index, and monitor it regularly to analyze the species change trend.
[0007] S1 is as follows: (1) Data collection: Hyperspectral imaging sensor types include airborne or spaceborne hyperspectral cameras; The data collection process was conducted under clear, cloudless weather conditions, during periods of high sun altitude. The acquisition method involves using a flight platform or satellite platform to cover the entire protected area along a preset flight path or orbit, acquiring high spatial resolution hyperspectral image data, with each pixel containing spectral information of hundreds of consecutive bands; (2) Data labeling: The collected data is radiometrically calibrated and geometrically corrected to convert the raw digital values into surface reflectance and register them to a geographic coordinate system. The annotation process is based on ground field surveys and expert knowledge. By setting up quadrats and transects, ground true data is collected within the protected area, the land cover or species category of each sampling point is recorded, and GPS positioning is used to match hyperspectral image pixels. The labeling categories include the main biological species types and abiotic features within the protected area. The labeling data is stored in the form of vector layers or raster labels for model training and validation.
[0008] S2 is as follows: An illumination invariance correction function is constructed. Based on the imaging time index, latitude, and longitude, solar irradiance is calculated using an atmospheric radiative transfer model as a quantitative indicator of illumination intensity. By calculating the difference between the mean values of the near-infrared band and the mean values of the visible band, a shadow compensation term is obtained to correct the spectral shift caused by shadows. The original spectral vector is corrected by using an illumination invariance correction function. Specifically, the original spectral vector is multiplied element-wise by the logarithm of solar irradiance, and a shadow compensation term is added to obtain the illumination-corrected spectral vector.
[0009] S3 is as follows: Based on the collected data of known biological species categories, the calculation of all samples in each category at the [missing information] rank [missing information]. to The average reflectance across the spectral bands is used to obtain the average spectral response vector, forming a category spectral template, which is then used to construct a biospectral fingerprint matrix. This represents the total number of original bands; For the spectral curve of each category in the biospectral fingerprint database, calculate the absolute value of the second derivative curvature at each band, capture key morphological features such as spectral absorption peaks and reflection peaks, and obtain the band curvature characteristics of each category at each band. Calculate the mutual information between the spectral response vector and the category label vector for each band, quantify the contribution of the band to class distinction, and obtain the mutual information of each band. By integrating the curvature characteristics and mutual information of each band, a comprehensive discrimination score is calculated, and the band with the highest score is selected to form a discriminative band subset. Based on the band index in the discriminative band subset, the characteristic values of the corresponding bands are filtered to obtain the band-selected spectral data.
[0010] S4 is as follows: The deep feature extraction model includes a multi-scale feature extraction module and a bioinformatics correction module; The multi-scale feature extraction module extracts local detail features by using small-scale convolution kernels in parallel and extracts global spectral features by combining large-scale dilated convolution kernels. The local feature matrix and the global feature matrix are concatenated and then fused by a weight matrix to obtain the fused feature matrix. The bioinformatics correction module defines the quantitative biological attributes of each category by constructing a biological attribute matrix, extracts the original classification logic value from the fusion feature matrix, generates biological knowledge correction items by using the autocorrelation operation of the biological attribute matrix, and generates the final classification prediction by adding the original classification logic value and the correction item and then using the Softmax function. Then, the cross-entropy between the predicted probability and the true label is used as the basic loss. The attribute projection matrix is used to map features and the difference between the actual biological attributes is calculated as the attribute consistency loss. For easily confused species pairs, the feature distance decay value is calculated as the separation loss. The three types of losses are combined and balanced by the weight coefficients to form the total loss function. Finally, the parameters of the deep feature extraction model are updated, and the model is trained using the backpropagation algorithm combined with an optimizer to minimize the total loss function, thus obtaining the trained model.
[0011] The operation of the multi-scale feature extraction module is as follows: Small-scale convolution kernels are used to extract local detail features from the input spectral data, capturing subtle spectral changes between adjacent bands to obtain a local feature matrix. Large-scale convolution kernels are used and dilated convolution mechanism is introduced to extract global spectral features of input spectral data, expand the receptive field to capture global features, and obtain global feature matrix; The local feature matrix and the global feature matrix are concatenated along the feature dimension and then fused using a feature fusion weight matrix to obtain the fused feature matrix.
[0012] The operation of the bioinformatics correction module is as follows: Based on domain knowledge, a set of quantifiable biological attribute features are defined for each species category, and a biological attribute matrix is constructed, in which each row represents a biological attribute vector of a category. Classification information is extracted from the fused feature matrix through a multilayer perceptron network to obtain the original classification logical value vector, which serves as the category score without biological knowledge correction. A biological attribute similarity matrix is constructed by autocorrelation operation of the biological attribute matrix, and then combined with the one-hot encoding vector of the real label to generate biological knowledge correction items; The original classification logic value vector is added to the biological knowledge correction term, and the final classification prediction probability vector is generated by the softmax function.
[0013] The process of calculating the loss function is as follows: Calculate the cross-entropy between the predicted probability vector and the one-hot encoded vector of the true label, and use it as the basic classification loss; The fused feature matrix is mapped to the biological attribute space using the attribute projection matrix, and the difference between the Frobenius norm and the real biological attribute vector is calculated as the biological attribute consistency loss. For a set of easily confused species pairs, the exponential decay value of the Euclidean distance between their feature vectors is calculated, and the separation loss of confused species is obtained by using biological attribute similarity as weight. The basic classification loss, biological attribute consistency loss, and confused species separation loss are combined and balanced by weight coefficients to obtain the total loss function, which serves as the overall optimization objective for model training.
[0014] The process of updating model parameters is as follows: The parameter update process trains the model based on the collected data, calculates the fused feature matrix and classification prediction probability vector through forward propagation, and then calculates the total loss function; During the training process of the model, the chain rule is used to calculate the gradient of the loss function with respect to each parameter of the model. The optimizer adopts the adaptive moment estimation algorithm, sets the initial learning rate, first moment and second moment decay factor, updates the parameter values according to the gradient, and dynamically adjusts the learning rate. During training, batch gradient descent is used. Each iteration inputs a batch of training samples, calculates the average loss and gradient, and updates the parameters. A condition is set to stop the model training iteration. The model terminates training when the condition is met, and the trained model is obtained.
[0015] S5 is detailed below: The newly acquired hyperspectral data undergoes the same preprocessing steps as during the training phase, including adaptive illumination correction and band selection based on biospectral characteristics, to obtain corrected band-selected spectral data. The processed data is then input into the trained model for multi-scale feature extraction and bioinformatics correction, and the classification prediction probability vector of each pixel is output to obtain the species category prediction result. Based on the prediction results, a species distribution map of the protected area is generated and the diversity index is calculated; Regular monitoring is conducted to compare species distribution and diversity indices at different times, monitor species change trends, and identify species invasions, habitat degradation, or conservation effectiveness.
[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention, through an adaptive illumination correction method based on the logarithmic reciprocal of solar irradiance and shadow compensation, can effectively solve the problem that linear normalization methods such as Min-Max standardization in the prior art cannot eliminate spectral distortion between shady and sunny slopes, thus enabling the same type of land cover to maintain spectral consistency under different illumination conditions. This invention employs a discriminative band selection method that fuses curvature features with band mutual information. This method reduces data dimensionality while preserving key biological characteristics, thus avoiding the shortcomings of conventional dimensionality reduction methods such as PCA in existing technologies, which destroy biological spectral absorption peaks and red-edge features, leading to a decrease in classification accuracy. This invention constructs a deep feature extraction module that combines multi-scale convolution and dilated convolution, simultaneously capturing local details and global spectral features, overcoming the problems of existing single-scale convolutional networks being unable to model long-range dependencies and lacking the ability to characterize complex spectral changes. This invention employs a biological attribute matrix and a knowledge correction mechanism, combined with a composite loss function consisting of cross-entropy loss, attribute consistency loss, and confusion species separation loss, to effectively improve the distinguishability of closely related species. This overcomes the shortcomings of existing technologies that rely solely on spectral features, are prone to confusing spectrally similar species, and lack biological prior knowledge constraints.
[0017] In summary, this invention proposes a comprehensive method that takes into account illumination correction, band selection, depth feature modeling, and biological knowledge constraints, which can realize the refined classification of species in nature reserves and long-term dynamic biodiversity monitoring. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0019] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0020] Figure 2 This is a schematic diagram illustrating the effect of spectral curve correction on the same type of land cover on shady / sunny slopes.
[0021] Figure 3 A comparison chart of the comprehensive discrimination scores for different band selection methods.
[0022] Figure 4 A comparison chart showing the classification accuracy results of different methods selected for different bands.
[0023] Figure 5 This is a distribution map of the four tree species in the feature space of the conventional model.
[0024] Figure 6 This is a distribution diagram of the four tree species in the feature space of the model of this invention. Detailed Implementation
[0025] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.
[0026] Example 1 like Figure 1 As shown, a method for monitoring biodiversity in nature reserves includes the following steps: S1. Use a hyperspectral imaging sensor to collect remote sensing data of nature reserves and label the collected data with biological species types; S2. Construct an illumination invariance correction function, and use a solar irradiance and shadow compensation mechanism to dynamically adjust the collected data and generate a correction spectral vector; S3. Based on the collected and calibrated data, construct a biospectral fingerprint database matrix, calculate the curvature characteristics and mutual information of each band, fuse the two to calculate the comprehensive discrimination score, select the band with the highest score to form a discriminative band subset, and then obtain the band selection spectral data according to the band number index. S4. Construct a deep feature extraction model to predict species types. Extract local and global features through multi-scale convolution and dilated convolution and fuse them. Generate knowledge correction terms based on the biological attribute matrix, calculate the loss function, and update the model parameters through an optimizer to obtain a trained model. S5. After performing steps S2 and S3 on the newly collected data, input it into the trained model to obtain the final classification result, then generate a species distribution map and calculate the diversity index, and monitor it regularly to analyze the species change trend.
[0027] In a specific implementation, S1 is as follows: During the hyperspectral data acquisition phase, hyperspectral imaging sensors are used to collect remote sensing data of nature reserves. Sensor types include airborne or spaceborne hyperspectral cameras, such as AVIRIS or Hyperion sensors, covering the visible to near-infrared spectral range (400–2500 nm). The data acquisition process is conducted under clear, cloudless weather conditions during periods of high solar altitude (e.g., 10:00 to 14:00 local time) to minimize atmospheric scattering and cloud interference. The acquisition method utilizes a flight platform (such as a drone or airplane) or satellite platform, covering the entire protected area along a pre-defined flight path or orbit to obtain high-resolution hyperspectral image data, with each pixel containing spectral information across hundreds of consecutive bands. After data acquisition, radiometric calibration and geometric correction are performed to convert the raw digital values into surface reflectance and register them to the geographic coordinate system. The annotation process is based on ground field surveys and expert knowledge. By setting up quadrats and transects, ground true data is collected within the protected area, the land cover or species category of each sampling point is recorded, and GPS positioning is used to match hyperspectral image pixels. The labeling categories include the main biological species types (such as specific tree species, shrubs, and herbaceous plants) and abiotic features (such as water bodies, soil, and rocks) within the protected area. The labeled data is stored in the form of vector layers or raster labels for model training and validation.
[0028] In specific implementations, open-source hyperspectral datasets can also be used to validate algorithms and conduct research. For example, the Indian Pines dataset from Indiana, USA, can be used. This dataset is a classic and widely used dataset in the field of hyperspectral image applications. It was taken in 1992 by an airborne AVIRIS sensor in an agricultural / forest area in Indiana, USA. It contains 224 bands (usually 200 are used after removing noise and water absorption bands), with a wavelength range of 400-2500nm and a spatial resolution of 20 meters / pixel. Furthermore, this dataset contains 16 categories, mainly different types of crops (such as corn and soybeans), forests (pine trees of different ages and deciduous forests), and some other land features (haystacks and pastures).
[0029] In a specific implementation, S2 is as follows: Because hyperspectral remote sensing data are affected by factors such as solar elevation angle and atmospheric scattering, the same type of land cover will exhibit nonlinear distortion of spectral response under different lighting conditions. Conventional linear normalization methods such as Min-Max standardization cannot effectively eliminate the proportional distortion between bands caused by differences in lighting, making it difficult to ensure the consistency of the spectrum of the same type of land cover under different lighting conditions such as shady slopes and sunny slopes.
[0030] Therefore, this invention constructs an illumination invariance correction function and employs a solar irradiance and shadow compensation mechanism to dynamically adjust the spectral response. The specific steps are as follows: 1) Calculate solar irradiance Based on the imaging time index, latitude, and longitude, solar irradiance is calculated using an atmospheric radiative transfer model as a quantitative indicator of light intensity for subsequent illumination correction, expressed as: In the formula, Solar irradiance, measured in W / m², is used to quantify light intensity. This serves as an imaging time index, representing the specific moment of data acquisition. Indicates the latitude of the imaging area, used to locate geographic coordinates; Indicates the longitude of the imaging area, used to locate geographic coordinates; This refers to atmospheric radiative transfer models, such as the MODTRAN model, used to accurately calculate solar radiation intensity based on imaging conditions.
[0031] It should be noted that the MODTRAN model (Moderate Resolution Atmospheric Radiance and Transmittance Model) is a medium-resolution atmospheric radiative transfer model used to calculate the transmittance, reflectance, and radiance of electromagnetic waves in the atmosphere. It is widely used in remote sensing, meteorology, and environmental science. In one embodiment, the MODTRAN model, based on the specific time, latitude, and longitude of the input data acquisition, and optionally, altitude, atmospheric composition (water vapor, ozone), aerosol parameters, etc., solves for solar irradiance using the radiative transfer equation. For example, when... , , At that time, the MODTRAN model output This indicates a clear, cloudless day.
[0032] 2) Calculate the shadow compensation term By calculating the difference between the mean values of the near-infrared band and the visible light band, a shadow compensation term is obtained to correct the spectral shift caused by shadows, and is expressed as: In the formula, This is a shadow compensation term used to correct spectral shifts caused by shadows; This represents the mean value in the near-infrared band, reflecting the average reflectance intensity in the near-infrared region, specifically derived from the original spectral vector. The mean value of all pixels is obtained by extracting the reflectance of the 700–1000 nm near-infrared band; This represents the mean value in the visible light band, reflecting the average reflectance intensity in the visible light region, specifically derived from the original spectral vector. The average value of all pixels is obtained by extracting the reflectance in the 400–700nm visible light band.
[0033] 3) Perform illumination correction The original spectral vector is multiplied element-wise by the reciprocal of the logarithm of solar irradiance, and a shading compensation term is added to obtain the illumination-corrected spectral vector. This eliminates spectral distortion caused by illumination variations and improves the spectral consistency of the same ground cover under different illumination conditions, expressed as: In the formula, To correct the spectral vector and characterize the spectral data after illumination invariance processing, the number of bands is B; The original spectral vector represents the uncorrected input spectral data, with B bands. This indicates an element-wise multiplication operation, used to adjust irradiance independently for each band; This is the surface reflectance correction factor, typically taken as 0.85. This item is used to adjust the intensity of the effect of irradiance on reflectivity; It is a logarithmic function with a default base of natural constant, used to compress the dynamic range of irradiance; This represents the total number of original bands, equivalent to the number of rows in the biospectral fingerprint database matrix.
[0034] It should be noted that the difference in illumination between shady and sunny slopes causes nonlinear distortion of the spectrum of the same type of land cover, such as overall reflectance shift and waveform distortion. High reflectivity under strong light compression The term characterizes the logarithm of solar irradiance, dynamically compressing the difference in illumination to avoid the scaling distortion of linear normalization. Simultaneously, a shadow compensation term is added. Eliminating the nonlinear shift caused by terrain shadows improves the spectral similarity of the same type of land feature on shady / sunny slopes.
[0035] In a specific implementation, S3 is as follows: Because hyperspectral data contains a large number of redundant bands, conventional dimensionality reduction methods such as principal component analysis can reduce the data dimensionality, but they can destroy the unique spectral structure of biological features, such as absorption peaks and reflection thresholds, which can easily lead to the loss of discriminative information and reduce classification accuracy.
[0036] Therefore, this invention constructs a biospectral fingerprint database matrix, calculates the band curvature characteristics and band mutual information of the spectral curves of each category, fuses the two to calculate a comprehensive discrimination score, and selects the band with the highest score to form a subset. This reduces data dimensionality while retaining key biological features. The specific steps are as follows: 1) Constructing a biospectral fingerprint library Based on training samples of known categories, the average spectral response vector of each category across all bands is calculated to form a category spectral template. This template is then used to construct a biospectral fingerprint matrix for subsequent band selection, as shown below: In the formula, This is a biospectral fingerprint database matrix with dimension 1. , where each row represents the average spectral response vector of a class; This represents the total number of categories of land features or species, used to determine the number of columns in the fingerprint database, and is equivalent to the number of categories. Indicates the first All samples of class 1 in the 1st to The average spectral response vector over the band is used to construct the spectral template for this category, by applying the first... The mean reflectance of each band is calculated from all training samples of the class, and is expressed as: ; For the first Number of samples in each class; For sample index; For the first The spectral reflectance value of the nth sample in the first band, i.e., the nth The correction spectral vector corresponding to each sample In the band Reflectance at that location; For the first The spectral reflectance value of the sample in the B band, i.e., the spectral reflectance value of the sample in the B band. The correction spectral vector corresponding to each sample In the band Reflectance at that location; A category index for land features or species; Indicates the first All samples of class 1 in the 1st to The average spectral response vector over the band; This represents the matrix transpose operation.
[0037] 2) Calculate the band curvature characteristics For each category of spectral curve in the biospectral fingerprint database, the absolute value of the second derivative curvature at each band is calculated to capture key morphological features such as spectral absorption peaks and reflection peaks, which are represented as follows: In the formula, The absolute value of the curvature of the c-th category at the b-th band is used to quantify the degree of curvature of the spectral curve in that band. This represents the band number index, with a value range from 1 to B; Indicates the first in the biospectral fingerprint database The spectral response value of class in the b-th band is the element value in the c-th row and b-th column of the biospectral fingerprint matrix F; Indicates the band number The second derivative is calculated in practice using the discrete difference method to approximate its value, and is used to extract spectral morphology features.
[0038] 3) Calculate the mutual information of the bands Calculate the mutual information between the spectral response vector and the class label vector for each band, quantifying the contribution of that band to class distinction, expressed as: In the formula, Let be the mutual information of the b-th band, used to measure the statistical correlation between the b-th band and the category label; The vector representing the spectral response values of all categories in the b-th band is the biospectral fingerprint matrix. The A column vector containing the average spectral response values of all categories in the b-th band; This is a category label vector used to provide category information; This represents the mutual information calculation function, used to measure the statistical dependency between two variables. In specific calculations, a vector... and category label vectors It needs to be a discretized vector, and the vector needs to be... Discretize into 10 intervals based on quantiles, category label vector It is itself a discretized vector.
[0039] 4) Select a subset of discriminative bands By integrating the curvature features and mutual information of each band, a comprehensive discrimination score is calculated. The bands with the highest scores are selected to form a discriminative band subset, which reduces data dimensionality while preserving key biometric features. This is represented as follows: In the formula, The comprehensive discrimination score for the b-th band is used to evaluate the band's morphological and statistical discrimination ability. Indicates the selection that makes Largest band number A set of bands used to determine the optimal subset of bands; For the discriminative band subset, It is the selected set of band indexes, whose dimensionality can be reduced to about 30% of the original number of bands, while retaining more than 95% of the classification accuracy, such as .
[0040] Based on discriminative band subsets By using the band index in the spectrum, and filtering the characteristic values of the corresponding bands, the band selection spectral data can be obtained. , dimension Characterizing the spectral data after band selection, This indicates the number of bands in the discriminative band subset.
[0041] In a specific implementation, S4 is as follows: S401, Multi-scale Feature Extraction Module Conventional one-dimensional convolutional neural networks struggle to effectively model long-range dependencies between bands when processing hyperspectral data, particularly for spectral phenomena involving multiple bands, such as red-edge shift. Single-scale convolutional kernels cannot simultaneously capture both local detail features and global spectral shape features, resulting in limited representation capabilities for complex spectral variations. Therefore, this invention enhances the ability to model long-range dependencies between bands by using small-scale convolutional kernels in parallel to extract local detail features and combining them with large-scale dilated convolutional kernels to extract global spectral shape features. The local and global feature matrices are then concatenated and fused using a weight matrix, resulting in improved modeling capabilities. The specific steps are as follows: 1) Local feature extraction Small-scale convolution kernels are used to extract local detail features from the input spectral data, capturing subtle spectral changes between adjacent bands, as shown below: In the formula, The local feature matrix has a dimension of . Each row corresponds to a local feature of a band; The number of feature maps, i.e. the number of output channels, is preferably 64. Each convolutional kernel generates one feature map, and the number of feature maps is the same as the number of convolutional kernels. This represents the modified linear unit activation function; This represents a one-dimensional convolution operation used to extract local band features; The input channel number represents the feature dimension of each band; For local convolution kernels, the size is defined as follows: This is used to extract features from adjacent bands; Indicates the expansion rate, set Characterizes standard convolution.
[0042] 2) Global Feature Extraction Large-scale convolution kernels are used, and a dilated convolution mechanism is introduced to extract global spectral shape features from the input spectral data, expanding the receptive field to capture global features such as red-edge shift, as follows: In the formula, The global feature matrix has a dimension of . Characterizes global spectral features, such as red-edge shift and broad absorption peaks; For global convolution kernels, the size is defined as follows: It is used to cover important spectral regions such as the red edge region; Indicates the expansion rate, set This means expanding the receptive field through interval sampling without increasing the number of parameters.
[0043] 3) Feature splicing and fusion The local and global feature matrices are concatenated along their feature dimensions and then fused using a feature fusion weight matrix to obtain a fused feature matrix. This fused feature matrix contains an optimized combination of local and global information, and is represented as follows: In the formula, To concatenate the feature matrix, the dimension is This represents the concatenation result of local and global features; This indicates a feature concatenation operation, used to merge local and global features; The feature fusion weight matrix has a dimension of . , are trainable parameters used to integrate local and global information; To fuse the feature matrix, the dimension is It includes an optimized combination of local and global information.
[0044] S402, Bioinformatics Correction Module In hyperspectral biological species classification tasks, conventional deep learning models rely solely on spectral features, easily confusing morphologically similar but different species. These species are extremely similar in spectral features but possess different biological attributes such as leaf shape and flowering time, making accurate differentiation difficult using only spectral data. Therefore, this invention constructs a biological attribute matrix to define quantified biological attributes for each category, extracts original classification logical values from the fused feature matrix, generates biological knowledge correction terms using autocorrelation operations on the biological attribute matrix, adds the original classification logical values and correction terms, and then uses a Softmax function to generate the final classification prediction. This approach combines spectral features and biological attribute differences to enhance discriminative ability. The specific steps are as follows: 1) Constructing a biological attribute matrix Based on domain knowledge, a set of quantifiable biological attribute features is defined for each species category, and a biological attribute matrix is constructed, where each row represents a biological attribute vector for a category, as shown below: In the formula, This is a biological attribute matrix with dimension 1. Each row represents a biological attribute vector for a category, and the element in the c-th row and p-th column of the biological attribute matrix is... ; Indicates the first The species in the first Quantitative values for biological attributes, such as, This represents the quantitative value of the second species in the first biological attribute; Indicates the first The species in the first Quantitative values for biological attributes; The number of biological attributes is equivalent to the number of columns in the biological attribute matrix. Biological attributes include morphology (leaf shape, tree height), physiology (chlorophyll, water content), phenology (flowering period, fruiting period), etc. Determined by expert knowledge, such as, ; This is a biological attribute index, with values ranging from 1 to P.
[0045] In practical implementation, quantified values Based on domain knowledge bases or expert annotations, for example, for the leaf shape attribute, it is quantified according to leaf shape classification codes. For example, an ovate leaf is assigned a value of lanceolate leaves are assigned the value Tree species The leaves are ovate, then For example, regarding the tree height attribute, based on the average height measured in the field, the tree species is set... The average tree height is Rice, then .
[0046] 2) Calculate the original classification logical value Classification information is extracted from the fused feature matrix using a multilayer perceptron network, represented as follows: This yields the original classification logical value vector. As a category score that has not been corrected for biological knowledge; in, This is the original classification logical value vector, i.e., the category scores without biological knowledge correction, such as... ; This represents a multilayer perceptron network, which contains multiple fully connected layers and activation functions to extract classification features. The activation function used is the modified linear unit activation function.
[0047] 3) Calculate biological knowledge correction items A biological attribute similarity matrix is constructed by performing autocorrelation operations on the biological attribute matrix, and then combined with the one-hot encoded vector of the true label to generate a biological knowledge correction term, which is used to adjust the classification score according to the biological attribute similarity, as shown below: In the formula, This is a biological knowledge correction item used to adjust classification scores based on the similarity of biological attributes; Biological attribute matrix The transpose of the matrix, The calculation results represent the biological attribute similarity matrix, with dimensions of Its element values reflect the degree of similarity in biological attributes between different species; This is the one-hot encoded vector of the true label of the input sample, used to provide true category information. , Represents all possible A set of one-hot encoded vectors, where only one element is 1 and the rest are 0, such as if... If the true category is category 2, then ; To correct the intensity coefficient and control the intensity of the influence of biological knowledge, a value of 0.3 is preferred.
[0048] It should be noted that, The term represents the species similarity matrix, i.e. The element in the i-th row and j-th column is , indicating the first The species and the first Biological similarity among species For the first species index, For the second species index, and The category score is enhanced to resemble that of real species, and the problem of spectral confusion is solved by combining prior knowledge.
[0049] 4) Generate final classification prediction The original classification logical value is added to the biological knowledge correction term, and the final classification prediction probability vector is generated through the softmax function, taking into account both spectral features and biological prior knowledge. This is expressed as: In the formula, This is the classification prediction probability vector, which represents the probability distribution of an input sample belonging to each class, with dimensions of . For example, when the total number of categories hour, Indicates that the sample belongs to the category The probability is ,category The probability is ,category The probability is ; This represents the Softmax normalization function, which transforms the input vector into a probability distribution. The corrected classification logic value is characterized by considering both spectral features and biological prior knowledge.
[0050] S403, Calculate the loss function Conventional cross-entropy loss only considers the matching degree between classification probability and true label, which is insufficient to address misclassification problems caused by spectral confusion of closely related species and correlation of biological attributes in hyperspectral data. Therefore, this invention uses the cross-entropy between predicted probability and true label as the basic loss, utilizes the attribute projection matrix to map features and calculates the difference with the true biological attributes as the attribute consistency loss, and calculates the feature distance decay value for easily confused species pairs as the separation loss. The three types of losses are combined and balanced by weight coefficients to form the total loss function, enhancing the model's adaptability to ecological characteristics. The specific steps are as follows: 1) Classification cross-entropy loss The cross-entropy between the predicted probability vector and the one-hot encoded vector of the true label is calculated and used as the basic classification loss to measure the classification error, expressed as: In the formula, One-hot encoded vector of the true label of the input sample The Dimensional elements; It is based on classification loss, and its calculation method is the same as that of cross-entropy loss.
[0051] For classification prediction probability vector The Dimension element, indicating that the sample belongs to the th dimension. The probability of each category.
[0052] 2) Loss of consistency in biological attributes The fused feature matrix is mapped to the biological attribute space using the attribute projection matrix, and the Frobenius norm difference between the fused feature matrix and the real biological attribute vector is calculated as the biological attribute consistency loss. This forces similar samples to project similarity in the attribute space, as expressed below: In the formula, The biological attribute consistency loss forces similar projections of samples of the same type into the attribute space. The attribute projection matrix is a trainable parameter that maps spectral features to the biological attribute space. This represents the Frobenius norm, used to measure the difference between the feature projection and the true attribute vector.
[0053] It should be noted that, Extracting the biological attribute vectors corresponding to the true categories, such as one-hot encoded vectors. At that time, return the biological attribute matrix. The second line.
[0054] 3) Confusion about species separation loss For a set of easily confused species pairs, the exponential decay value of the Euclidean distance between their feature vectors is calculated, and biological attribute similarity is used as the weight to obtain the separation loss of confused species, which enhances the feature discriminativeness of confused species, as expressed as: In the formula, To mitigate species separation loss, mutually exclusive constraints are used to enhance the distinguishability of easily confused species pairs with similar spectra. For the set of easily confused species pairs, the species similarity matrix is used. Before sorting the upper triangular elements Components; For the third species index; For the fourth species index; For element indexing, e.g., to get the total number of categories. 20%; For the first The average eigenvector of each category in the fused feature matrix; For the first The average eigenvector of each category in the fused feature matrix; Indicates the first The species and the first The similarity of biological attributes between species is expressed as follows: the higher the value, the easier it is to confuse them. ; To act as a smoothing factor and avoid a denominator of zero, for example, .
[0055] It should be noted that the use of biological attribute similarity As a penalty weight, it can force the model to assign a larger feature distance to highly similar species.
[0056] 4) Integration of total loss function The combined basic classification loss, biological attribute consistency loss, and confused species separation loss are weighted and balanced to obtain the total loss function, which serves as the overall optimization objective for model training, and is expressed as: In the formula, The total loss function is the overall optimization objective of the model training process; The attribute consistency weight controls the strength of biological attribute constraints, and is preferably set to 0.2; To obfuscate the separation weights, the spatial distribution of features is optimized for closely related species, with a preferred setting of 0.5.
[0057] S404, Deep Feature Extraction Model Parameter Update In the parameter update stage of the deep feature extraction model, a backpropagation algorithm combined with an optimizer is used for model training to minimize the total loss function. The parameter update process is based on the training dataset. The fused feature matrix and classification prediction probability vector are calculated through forward propagation, and then the total loss function is calculated, including classification cross-entropy loss, biological attribute consistency loss, and confusion species separation loss. The chain rule is used to calculate the gradient of the loss function with respect to each parameter of the model, including the convolution kernel weights in the multi-scale feature extraction module, the feature fusion weight matrix, the multilayer perceptron weights in the bioinformatics correction module, and the attribute projection matrix, etc. The optimizer uses an adaptive moment estimation algorithm, such as the Adam optimizer, to set the initial learning rate, first moment, and second moment decay factor. The parameter values are updated according to the gradient. The learning rate can be dynamically adjusted according to the training progress, such as by using a learning rate decay strategy. During training, batch gradient descent is used. In each iteration, a batch of training samples is input, the average loss and gradient are calculated, and the parameters are updated.
[0058] The decision to stop model training iterations is based on validation set performance monitoring. When the classification accuracy on the validation set no longer improves or the validation loss no longer decreases in multiple consecutive iterations, the early stopping mechanism is triggered to terminate training. Alternatively, training can be stopped when the number of training iterations reaches the preset maximum number of iterations.
[0059] The model parameters are updated until convergence, and the optimal model parameters are obtained and saved as a model weight file for subsequent biodiversity monitoring applications.
[0060] In a specific implementation, S5 is as follows: During the biodiversity monitoring phase, the trained deep feature extraction model is used to process new hyperspectral data to achieve species classification and diversity assessment.
[0061] First, the newly acquired hyperspectral data undergoes the same preprocessing steps as in the training phase, including adaptive illumination correction and band selection based on biological spectral characteristics, to obtain corrected band-selected spectral data. The processed data is then input into the trained model for multi-scale feature extraction and bioinformatics correction, and the classification prediction probability vector of each pixel is output to obtain the species category prediction result. Based on the prediction results, a species distribution map of the protected area is generated, spatially displaying the species category of each cell, and calculating the species abundance, diversity indices (such as the Shannon Wiener index and the Simpson index) and species evenness within the area.
[0062] The monitoring process is conducted regularly (e.g., quarterly or annually) to monitor species change trends and identify species invasions, habitat degradation, or conservation effectiveness by comparing species distribution and diversity indices at different times.
[0063] Example 2 like Figure 2 As shown, a comparative analysis of the adaptive illumination correction effect was conducted to verify the effectiveness of the adaptive illumination correction method proposed in this invention in eliminating spectral differences between the same type of land cover on shady and sunny slopes. The experiment simulated hyperspectral data of the same vegetation in a nature reserve under shady and sunny slope conditions, comparing the changes in spectral curves before and after correction. In the figure, the horizontal axis represents wavelength in nanometers, ranging from 400 nm to 2500 nm, covering the visible to near-infrared spectral region; the vertical axis represents reflectance, dimensionless, reflecting the land cover's ability to reflect light. Experimental results show that in the uncorrected original spectrum, the spectral curve of the sunny slope has a higher overall reflectance, and the red edge feature (around 720 nm) exhibits a blue shift, while the spectral curve of the shady slope has lower reflectance, and the red edge feature shows a slight red shift, indicating that illumination differences cause nonlinear distortion in the spectral response of the same type of land cover. After correction by this invention, the spectral curves of the sunny and shady slopes almost overlap, the reflectance values tend to be consistent, and the red edge position returns to normal, indicating that the correction method successfully eliminated the influence of illumination. The experimental results demonstrate the advantages of this technology. By dynamically adjusting solar irradiance and shadow compensation, it effectively improves the spectral consistency of the same ground cover under different lighting conditions, providing a more reliable data foundation for subsequent species classification.
[0064] Example 3 like Figure 3 and Figure 4As shown, a comparison of band selection effects is conducted to evaluate the advantages of the proposed band selection method in preserving key biological features and improving classification accuracy. The experiment compares the proposed method, principal component analysis, and random selection method, analyzing the results across two dimensions: band score and classification accuracy. In the figure, the horizontal axis represents wavelength in nanometers, and the vertical axis represents the comprehensive discrimination score, which integrates band curvature features and mutual information to measure the discrimination ability of each band. Experimental results show that... Figure 3 The wavebands selected by the method of this invention are concentrated in regions with high comprehensive discrimination scores. These wavebands typically correspond to key morphological features such as absorption and reflection peaks in the spectral curve, indicating that this invention can effectively capture biologically specific information. In contrast, the wavebands selected by the random selection method are more evenly distributed, while those selected by the principal component analysis method are scattered in regions with lower scores. Figure 4 As shown, in the comparison of classification accuracy, the accuracy of the method of this invention is significantly higher than that of the other two methods, which proves that by fusing curvature features and mutual information, this invention can reduce data dimensionality while retaining discriminative bands, thereby improving classification performance and avoiding the feature loss problem caused by conventional dimensionality reduction methods.
[0065] Example 4 like Figure 5 and Figure 6 As shown, the advantages of the model are analyzed through feature space distribution maps. The experimental configuration uses four typical tree species from the Indian-Pines dataset: pine (green dots), oak (purple dots), maple (yellow dots), and birch (blue dots), with 100 sample points for each category. Figure 5 As shown, the conventional model uses a standard convolutional neural network as a comparison method. In the feature space of the conventional model, the boundaries of the four categories overlap, leading to classification confusion. Figure 6 As shown, in the feature space of the model of this invention, samples of the same type are highly clustered, and the distance between different categories is significantly increased. At the same time, this invention forms richer discriminative dimensions in the feature space through parallel small-scale convolutional kernels (capturing local features such as leaf vein texture) and large-scale dilated convolutions (capturing global features such as tree crown morphology).
[0066] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.
Claims
1. A method for monitoring biodiversity in nature reserves, characterized in that, Includes the following steps: S1. Use a hyperspectral imaging sensor to collect remote sensing data of nature reserves and label the collected data with biological species types; S2. Construct an illumination invariance correction function, and use a solar irradiance and shadow compensation mechanism to dynamically adjust the collected data and generate a correction spectral vector; S3. Based on the collected and calibrated data, construct a biospectral fingerprint database matrix, calculate the curvature characteristics and mutual information of each band, fuse the two to calculate the comprehensive discrimination score, select the band with the highest score to form a discriminative band subset, and then obtain the band selection spectral data according to the band number index. S4. Construct a deep feature extraction model to predict species types. Extract local and global features through multi-scale convolution and dilated convolution and fuse them. Generate knowledge correction terms based on the biological attribute matrix, calculate the loss function, and update the model parameters through an optimizer to obtain a trained model. S5. After performing steps S2 and S3 on the newly collected data, input it into the trained model to obtain the final classification result, then generate a species distribution map and calculate the diversity index, and monitor it regularly to analyze the species change trend.
2. The method for monitoring biodiversity in nature reserves according to claim 1, characterized in that, S2 is as follows: An illumination invariance correction function is constructed, and solar irradiance is calculated using an atmospheric radiative transfer model based on imaging time index, latitude, and longitude. The shadow compensation term is obtained by calculating the difference between the mean values in the near-infrared band and the mean values in the visible light band. The original spectral vector is corrected by using an illumination invariance correction function. Specifically, the original spectral vector is multiplied element-wise by the logarithm of solar irradiance, and a shadow compensation term is added to obtain the illumination-corrected spectral vector.
3. The biodiversity monitoring method for nature reserves according to claim 1, characterized in that, S3 is as follows: Based on the collected data of known biological species categories, the calculation of all samples in each category at the [missing information] rank [missing information]. to The average reflectance across the spectral bands is used to obtain the average spectral response vector, forming a category spectral template, which is then used to construct a biospectral fingerprint matrix. This represents the total number of original bands; For the spectral curve of each category in the biospectral fingerprint database, calculate the absolute value of the second derivative curvature at each band to obtain the band curvature characteristics of each category at each band. Calculate the mutual information between the spectral response vector and the category label vector for each band, quantify the contribution of the band to class distinction, and obtain the mutual information of each band. By integrating the curvature characteristics and mutual information of each band, a comprehensive discrimination score is calculated, and the band with the highest score is selected to form a discriminative band subset. Based on the band index in the discriminative band subset, the characteristic values of the corresponding bands are selected to obtain the band-selected spectral data.
4. The biodiversity monitoring method for nature reserves according to claim 1, characterized in that, S4 is as follows: The deep feature extraction model includes a multi-scale feature extraction module and a bioinformatics correction module; The multi-scale feature extraction module extracts local detail features by using small-scale convolution kernels in parallel and extracts global spectral features by combining large-scale dilated convolution kernels. The local feature matrix and the global feature matrix are concatenated and then fused by a weight matrix to obtain the fused feature matrix. The bioinformatics correction module defines the quantitative biological attributes of each category by constructing a biological attribute matrix, extracts the original classification logic value from the fusion feature matrix, generates biological knowledge correction items by using the autocorrelation operation of the biological attribute matrix, and generates the final classification prediction by adding the original classification logic value and the correction item and then using the Softmax function. Then, the cross-entropy between the predicted probability and the true label is used as the basic loss. The attribute projection matrix is used to map features and the difference between the actual biological attributes is calculated as the attribute consistency loss. For easily confused species pairs, the feature distance decay value is calculated as the separation loss. The three types of losses are combined and balanced by the weight coefficients to form the total loss function. Finally, the parameters of the deep feature extraction model are updated, and the model is trained using the backpropagation algorithm combined with an optimizer to minimize the total loss function, thus obtaining the trained model.
5. A method for monitoring biodiversity in a nature reserve according to claim 4, characterized in that, The operation of the multi-scale feature extraction module is as follows: Small-scale convolution kernels are used to extract local detail features from the input spectral data, resulting in a local feature matrix; A global feature matrix is obtained by using large-scale convolution kernels and introducing dilated convolution mechanism to extract global spectral features from the input spectral data. The local feature matrix and the global feature matrix are concatenated along the feature dimension and then fused using a feature fusion weight matrix to obtain the fused feature matrix.
6. A method for monitoring biodiversity in a nature reserve according to claim 4, characterized in that, The operation of the bioinformatics correction module is as follows: Based on domain knowledge, a set of quantifiable biological attribute features are defined for each species category, and a biological attribute matrix is constructed. The classification information is extracted from the fused feature matrix through a multilayer perceptron network to obtain the original classification logical value vector; A biological attribute similarity matrix is constructed by autocorrelation operation of the biological attribute matrix, and then combined with the one-hot encoding vector of the real label to generate biological knowledge correction items; The original classification logic value vector is added to the biological knowledge correction term, and the final classification prediction probability vector is generated by the softmax function.
7. A method for monitoring biodiversity in nature reserves according to claim 4, characterized in that, The process of calculating the loss function is as follows: Calculate the cross-entropy between the predicted probability vector and the one-hot encoded vector of the true label, and use it as the basic classification loss; The fused feature matrix is mapped to the biological attribute space using the attribute projection matrix, and the difference between the Frobenius norm and the real biological attribute vector is calculated as the biological attribute consistency loss. For a set of easily confused species pairs, the exponential decay value of the Euclidean distance between their feature vectors is calculated, and the separation loss of confused species is obtained by using biological attribute similarity as weight. The basic classification loss, biological attribute consistency loss, and confused species separation loss are combined and balanced by weight coefficients to obtain the total loss function, which serves as the overall optimization objective for model training.
8. A method for monitoring biodiversity in a nature reserve according to claim 4, characterized in that, The process of updating model parameters is as follows: The parameter update process trains the model based on the collected data, calculates the fused feature matrix and classification prediction probability vector through forward propagation, and then calculates the total loss function; During the training process of the model, the chain rule is used to calculate the gradient of the loss function with respect to each parameter of the model. The optimizer adopts the adaptive moment estimation algorithm, sets the initial learning rate, first moment and second moment decay factor, updates the parameter values according to the gradient, and dynamically adjusts the learning rate. During training, batch gradient descent is used. In each iteration, a batch of training samples is input, the average loss and gradient are calculated, and the parameters are updated. A condition is set to stop the model training iteration. The model will terminate training when the condition is met, and a trained model will be obtained.
9. A method for monitoring biodiversity in a nature reserve according to claim 1, characterized in that, S5 is detailed below: The newly acquired hyperspectral data undergoes the same preprocessing steps as during the training phase, including adaptive illumination correction and band selection based on biospectral characteristics, to obtain corrected band-selected spectral data. The processed data is then input into the trained model for multi-scale feature extraction and bioinformatics correction, and the classification prediction probability vector of each pixel is output to obtain the species category prediction result. Based on the prediction results, a species distribution map of the protected area is generated and the diversity index is calculated.
10. A method for monitoring biodiversity in a nature reserve according to claim 1, characterized in that, S1 is as follows: (1) Data collection: Hyperspectral imaging sensor types include airborne or spaceborne hyperspectral cameras; The data collection process was conducted under clear, cloudless weather conditions, during periods of high sun altitude. The acquisition method involves using a flight platform or satellite platform to cover the entire protected area along a preset flight path or orbit, acquiring high spatial resolution hyperspectral image data, with each pixel containing spectral information of hundreds of consecutive bands; (2) Data labeling: The collected data is radiometrically calibrated and geometrically corrected to convert the raw digital values into surface reflectance and register them to a geographic coordinate system. The annotation process is based on ground field surveys and expert knowledge. By setting up quadrats and transects, ground true data is collected within the protected area, the land cover or species category of each sampling point is recorded, and GPS positioning is used to match hyperspectral image pixels. The labeling categories include the main biological species types and abiotic features within the protected area, and the labeling data is stored in the form of vector layers or raster labels.