A biodiversity inversion method based on multi-source remote sensing image fusion

Abstract

The present application belongs to the technical field of computer data processing, and provides a biodiversity inversion method based on multi-source remote sensing image fusion, comprising: remote sensing image data acquisition, image preprocessing, image fusion processing, multispectral resolution image data generation, spectral feature extraction, final frequency feature extraction, feature integration and target ecological variable prediction; the present application realizes differentiated fusion of high and low frequency components under different scales through wave base adaptive selection and multi-scale wavelet decomposition, taking into account spectral fidelity and spatial structure preservation, improves the complementarity of multi-source data and the quality of fused images; through spectral resolution refinement processing and multi-bandwidth scale simulation, the limitation of single resolution is broken through, and the ability to capture complex spectral features of vegetation is enhanced; through spatial neighborhood feature fusion, spatial correlation is enhanced; through the ecological variable inversion estimation model, multi-index synchronous prediction is realized, and the generality of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and in particular to a biodiversity inversion method based on multi-source remote sensing image fusion. Background Technology

[0002] Biodiversity is a crucial indicator for measuring the health and function of ecosystems, and it plays a vital role in environmental protection, resource management, and sustainable development. In recent years, with the rapid development of remote sensing technology, using remote sensing imagery for biodiversity assessment and biodiversity index inversion has become a hot topic in ecological research. Existing biodiversity remote sensing inversion methods mainly rely on hyperspectral remote sensing data, estimating indicators such as vegetation diversity indices and species richness through spectral feature extraction and model inversion.

[0003] However, existing biodiversity remote sensing inversion methods typically involve manually setting wave bases and using a single frequency component fusion strategy after wavelet decomposition, which fails to balance spectral fidelity and spatial structure, resulting in mediocre final image fusion effects. Conventional frequency domain features only include the harmonic attributes of a single pixel, neglecting spatial features in the vicinity. Furthermore, conventional models can only predict a single index, resulting in low versatility. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a biodiversity inversion method based on multi-source remote sensing image fusion, which solves the problems of unreasonable basis setting, single fusion strategy, lack of spatial features, and single prediction index in existing biodiversity remote sensing inversion methods.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A biodiversity inversion method based on multi-source remote sensing image fusion includes:

[0007] Acquire remote sensing image data of the target area; the remote sensing image data includes: hyperspectral remote sensing image and high-resolution remote sensing image;

[0008] The remote sensing image data is preprocessed; the preprocessing includes: radiometric correction, atmospheric correction, geometric correction, and cropping.

[0009] The preprocessed remote sensing image data is subjected to wavelet basis adaptive selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image.

[0010] The final fused image is subjected to spectral resolution refinement processing and multi-bandwidth scale simulation to obtain multispectral resolution image data; the multispectral resolution image includes: several sets of simulated images with different spectral resolutions;

[0011] Spectral features are extracted from the multispectral resolution image data to obtain vegetation-related spectral features;

[0012] The multispectral resolution image data is subjected to adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration to obtain the final frequency features.

[0013] The vegetation-related spectral features and the final frequency features are combined to obtain the input feature set;

[0014] The input feature set is fed into a pre-trained ecological variable inversion estimation model to generate data and obtain the predicted value of the target ecological variable.

[0015] Preferably, the preprocessed remote sensing image data undergoes wavelet basis adaptive selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image, including:

[0016] Pre-selected set of preliminary wavelet bases;

[0017] The preprocessed remote sensing image data is decomposed and fused using each candidate wavelet basis in the pre-set wavelet basis set to obtain a temporary fused image.

[0018] The spectral similarity index is obtained by performing spectral angle mapping data calculation on the temporary fused image; the calculation formula for the spectral similarity index is as follows: Wherein, SAM(p,q) is the spectral similarity index; p is the pixel spectral vector of the preprocessed hyperspectral remote sensing image; q is the pixel spectral vector of the temporary fused image; arccos(·) represents the inverse cosine function;

[0019] Structural similarity is calculated on the temporary fused image to obtain a spatial quality index; the formula for calculating the spatial quality index is as follows: Wherein, SSIM(x,y) is the spatial quality index; x and y are the image patches to be compared between the preprocessed high-resolution remote sensing image and the temporary fused image, respectively; μ x μ y These are the pixel mean values ​​for x and y, respectively; σ represents the pixel variances of x and y, respectively. xy Let x be the covariance of x and y; C1 and C2 are the first and second constant terms, respectively;

[0020] The spectral similarity index and the spatial quality index are weighted and calculated to obtain the band score; the formula for calculating the band score is as follows: Where ω1+ω2=1; Score is the band score; ω1 and ω2 are the first weight and the second weight, respectively;

[0021] The band scores are sorted in descending order to obtain a score order set, and the candidate wavelet bases corresponding to the target number of band scores in the score order set are extracted sequentially to obtain the optimal wavelet base set.

[0022] Preferably, wavelet decomposition and fusion are performed on the preprocessed remote sensing image data using each candidate wavelet basis in the pre-set wavelet basis set to obtain a temporary fused image, including:

[0023] The preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image are decomposed using each candidate wavelet basis in the pre-set wavelet basis set.

[0024] The low-frequency portion of the decomposed hyperspectral remote sensing image and the high-frequency portion of the high-resolution remote sensing image are selected to obtain temporary fused data.

[0025] The temporary fused data is subjected to a single-level wavelet inverse transform to obtain the temporary fused image.

[0026] Preferably, the preprocessed remote sensing image data undergoes wavelet basis adaptive selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image, including:

[0027] The preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image are subjected to two-level wavelet decomposition using the optimal wavelet basis set to obtain hyperspectral decomposition results and high-resolution decomposition results. The hyperspectral decomposition results include: a hyperspectral low-frequency component, a first hyperspectral high-frequency component, and a second hyperspectral high-frequency component. The high-resolution decomposition results include: a high-resolution low-frequency component, a first high-resolution high-frequency component, and a second high-resolution high-frequency component.

[0028] The local variance of each pixel in the 3×3 neighborhood of the preprocessed high-resolution remote sensing image is calculated and normalized to obtain the local fusion weights.

[0029] The high-frequency components within the hyperspectral decomposition result and the high-resolution decomposition result are weighted and fused using the local fusion weights to obtain a high-frequency fusion result.

[0030] The correlation coefficient is obtained by calculating the correlation between the hyperspectral low-frequency component and the high-resolution low-frequency component; the formula for calculating the correlation coefficient is as follows:

[0031]

[0032] Where, ρ b Let be the correlation coefficient for band b; i and j are the row and column numbers of the pixel coordinates, respectively. These represent the reflectance values ​​of the hyperspectral low-frequency component and the b-th band low-frequency component of the high-resolution low-frequency component at pixel (i,j), respectively. These represent the global mean values ​​of the hyperspectral low-frequency component and the b-th band of the high-resolution low-frequency component, respectively.

[0033] When the correlation coefficient is greater than the preset spectral threshold, the high-spectral low-frequency component is retained to obtain the low-frequency fusion result;

[0034] When the correlation coefficient is less than a preset spectral threshold, the spectral angle mapping value of the band is calculated. Based on the spectral angle mapping value, a spectral fidelity weight is calculated. Then, the spectral fidelity weight is used to perform weighted fusion of the hyperspectral low-frequency component and the high-resolution low-frequency component to obtain the low-frequency fusion result. The formula for calculating the low-frequency fusion result is as follows:

[0035]

[0036] in, The low-frequency fusion result; α b The spectral fidelity weights are SAM. b The spectral angle mapping value; min SAM SAM in all bands b The minimum value;

[0037] The high-frequency fusion result and the low-frequency fusion result are subjected to a two-level wavelet inverse transform to obtain the final fused image.

[0038] Preferably, the final fused image is subjected to spectral resolution refinement processing and multi-bandwidth scale simulation to obtain multispectral resolution image data, including:

[0039] Construct a Gaussian spectral response function; the expression of the Gaussian spectral response function is:

[0040]

[0041] Among them, R a (i,j) represents the reflectance value of the a-th band at pixel (i,j); A a μ is the amplitude of the a-th band; a σ is the center wavelength of the a-th band; a The standard deviation parameter is mentioned above;

[0042] Select the bandwidth range;

[0043] Calculate the standard deviation parameter corresponding to each bandwidth within the bandwidth range; the formula for calculating the standard deviation parameter is: Where W is the target bandwidth;

[0044] The standard deviation parameter is updated in the Gaussian spectral response function, and spectral reflectance data under the bandwidth corresponding to the standard deviation parameter is generated using the updated Gaussian spectral response function based on the final fused image to obtain the multispectral resolution image data.

[0045] Preferably, spectral feature extraction is performed on the multispectral resolution image data to obtain vegetation-related spectral features, including:

[0046] The vegetation-related spectral features are obtained by calculating the normalized vegetation index, green vegetation index, differential vegetation index, improved soil-adjusted vegetation index, green normalized differential vegetation index, soil-adjusted vegetation index, spectral coefficient of variation, and spectral convex hull volume of the multispectral resolution image data.

[0047] Preferably, the multispectral resolution image data is subjected to adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration to obtain the final frequency features, including:

[0048] Gaussian low-pass filtering is applied to the spectral sequence of each pixel in the multispectral resolution image data, and band data with a signal-to-noise ratio of less than 20dB are removed to obtain a denoised spectral sequence.

[0049] Perform a Fourier transform on the denoised spectral sequence, calculate the spectral reconstruction error at each order step by step, and record the order at which the spectral reconstruction error first becomes less than 5%, thus obtaining the optimal order.

[0050] Harmonic coefficients and power spectra corresponding to each order from the first order to the optimal order are extracted, and the amplitude and phase are calculated based on the harmonic coefficients to obtain the harmonic decomposition feature vector;

[0051] Calculate the energy entropy from the first order to the optimal order, the maximum frequency energy ratio, and the cross power spectrum of adjacent harmonics to obtain the fundamental frequency eigenvector.

[0052] The mean and standard deviation of the harmonic coefficients from the first order to the optimal order of all pixels within the multispectral resolution image data window are calculated using a 3×3 sliding window to obtain the spatial enhancement feature vector;

[0053] The final frequency feature is obtained by merging the harmonic decomposition feature vector, the fundamental frequency feature vector, and the spatial enhancement feature vector.

[0054] Preferably, the training process of the ecological variable inversion estimation model includes:

[0055] An improved DNN network is obtained by replacing the two-dimensional convolution of the basic DNN network with three-dimensional convolution and replacing the output layer of the basic DNN network with a multi-task output head.

[0056] Collect remote sensing image data and several biological indicator data from different sub-regions and time periods of the study area, extract the input feature set of the remote sensing image data, and combine the input feature set and all the biological indicator data in pairs to obtain a model training dataset including several sets of training data pairs.

[0057] The model training dataset is input into the improved DNN network for forward propagation to obtain the network output;

[0058] The loss value of each task output head is calculated using a loss function to obtain the single-index error. All single-index errors are then summed to obtain the overall error. The expression for the loss function is:

[0059]

[0060] in, The single index error of the kth task output header; The predicted value output by the k-th task output header; y k This refers to the actual indicator data corresponding to the kth task output header;

[0061] The parameters of the task output head are optimized by backpropagation using the single index error, and the hidden layer parameters of the improved DNN network are optimized by backpropagation using the overall error. Then, the process returns to the step "input the model training dataset into the improved DNN network for forward propagation to obtain the network output", thus obtaining the trained ecological variable inversion estimation model.

[0062] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned biodiversity inversion method based on multi-source remote sensing image fusion.

[0063] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned biodiversity inversion method based on multi-source remote sensing image fusion.

[0064] The present invention discloses the following technical effects:

[0065] This invention provides a biodiversity inversion method based on multi-source remote sensing image fusion. Through adaptive basis selection and multi-scale wavelet decomposition, it addresses the problem of existing methods failing to balance spectral fidelity and spatial structure preservation, achieving differentiated fusion of high and low frequency components at different scales. By refining spectral resolution and performing multi-bandwidth scale simulation, it overcomes the limitation of single resolution, enhancing the ability to capture complex spectral features of vegetation. Through spatial neighborhood feature fusion, it addresses the lack of neighboring pixel influence in conventional methods, enhancing spatial correlation. Finally, through an ecological variable inversion estimation model, it overcomes the limitation of existing methods in predicting only a single indicator, enabling simultaneous prediction of multiple indicators. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A schematic diagram of the biodiversity inversion process based on multi-source remote sensing image fusion provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram illustrating the process of determining the optimal wavelet basis set provided in an embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of the final fused image acquisition process provided in an embodiment of the present invention;

[0070] Figure 4 This is a schematic diagram of the multispectral resolution image data generation process provided in an embodiment of the present invention;

[0071] Figure 5 This is a schematic diagram of the final frequency feature extraction process provided in an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The purpose of this invention is to provide a biodiversity inversion method based on multi-source remote sensing image fusion, which solves the problems of unreasonable basis setting, single fusion strategy, lack of spatial features and single prediction index in existing biodiversity remote sensing inversion methods.

[0074] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] Figure 1 This is a schematic diagram of the biodiversity inversion process based on multi-source remote sensing image fusion provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a biodiversity inversion method based on multi-source remote sensing image fusion, comprising:

[0076] Step 100: Acquire remote sensing image data of the target area; the remote sensing image data includes: hyperspectral remote sensing image and high-resolution remote sensing image;

[0077] Step 200: Preprocess the remote sensing image data; the preprocessing includes: radiometric correction, atmospheric correction, geometric correction, and cropping;

[0078] Step 300: Perform wavelet basis adaptive selection, multi-scale wavelet decomposition, low-frequency component and high-frequency component fusion, and wavelet inverse transform reconstruction on the preprocessed remote sensing image data to obtain the final fused image;

[0079] Step 400: Perform spectral resolution refinement processing and multi-bandwidth scale simulation on the final fused image to obtain multispectral resolution image data; the multispectral resolution image includes: several sets of simulated images with different spectral resolutions;

[0080] Step 500: Extract spectral features from the multispectral resolution image data to obtain vegetation-related spectral features;

[0081] Step 600: Perform adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration on the multispectral resolution image data to obtain the final frequency features;

[0082] Step 700: Combine the vegetation-related spectral features and the final frequency features to obtain the input feature set;

[0083] Step 800: Input the input feature set into the pre-trained ecological variable inversion estimation model to generate data and obtain the predicted value of the target ecological variable.

[0084] refer to Figure 2The preprocessed remote sensing image data undergoes wavelet basis adaptive selection, multi-scale wavelet decomposition, low-frequency and high-frequency component fusion, and inverse wavelet transform reconstruction to obtain the final fused image, including:

[0085] Step 301: Pre-select the set of preliminary wavelet bases;

[0086] Step 302: Perform wavelet decomposition and fusion on the preprocessed remote sensing image data using each candidate wavelet basis in the preparatory wavelet basis set to obtain a temporary fused image;

[0087] Step 303: Calculate the spectral angle mapping data of the temporary fused image to obtain a spectral similarity index; the calculation formula for the spectral similarity index is: Wherein, SAM(p,q) is the spectral similarity index; p is the pixel spectral vector of the preprocessed hyperspectral remote sensing image; q is the pixel spectral vector of the temporary fused image; arccos(·) represents the inverse cosine function;

[0088] Step 304: Calculate the structural similarity of the temporary fused image to obtain a spatial quality index; the formula for calculating the spatial quality index is: Wherein, SSIM(x,y) is the spatial quality index; x and y are the image patches to be compared between the preprocessed high-resolution remote sensing image and the temporary fused image, respectively; μ x μ y These are the pixel mean values ​​for x and y, respectively; σ represents the pixel variances of x and y, respectively. xy Let x be the covariance of x and y; C1 and C2 are the first and second constant terms, respectively;

[0089] Step 305: Perform a weighted calculation on the spectral similarity index and the spatial quality index to obtain the band score; the calculation formula for the band score is: Where ω1+ω2=1; Score is the band score; ω1 and ω2 are the first weight and the second weight, respectively;

[0090] Step 306: Sort the band scores in descending order to obtain a score order set, and sequentially extract the candidate wavelet bases corresponding to the target number of band scores in the score order set to obtain the optimal wavelet base set.

[0091] Specifically, wavelet decomposition and fusion are performed on the preprocessed remote sensing image data using each candidate wavelet basis in the pre-set wavelet basis set to obtain a temporary fused image, including:

[0092] The preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image are decomposed using each candidate wavelet basis in the pre-set wavelet basis set.

[0093] The low-frequency portion of the decomposed hyperspectral remote sensing image and the high-frequency portion of the high-resolution remote sensing image are selected to obtain temporary fused data.

[0094] The temporary fused data is subjected to a single-level wavelet inverse transform to obtain the temporary fused image.

[0095] refer to Figure 3 The preprocessed remote sensing image data undergoes wavelet basis adaptive selection, multi-scale wavelet decomposition, low-frequency and high-frequency component fusion, and inverse wavelet transform reconstruction to obtain the final fused image, including:

[0096] Step 307: Perform two-level wavelet decomposition on the preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image using the optimal wavelet basis set to obtain hyperspectral decomposition results and high-resolution decomposition results; the hyperspectral decomposition results include: hyperspectral low-frequency components, first hyperspectral high-frequency components and second hyperspectral high-frequency components; the high-resolution decomposition results include: high-resolution low-frequency components, first high-resolution high-frequency components and second high-resolution high-frequency components.

[0097] Step 308: Calculate the local variance of each pixel in the 3×3 neighborhood of the preprocessed high-resolution remote sensing image and perform normalization processing to obtain the local fusion weight;

[0098] Step 309: Use the local fusion weights to perform weighted fusion of the high-frequency components in the hyperspectral decomposition result and the high-resolution decomposition result to obtain the high-frequency fusion result;

[0099] Step 310: Calculate the correlation between the hyperspectral low-frequency component and the high-resolution low-frequency component to obtain the correlation coefficient; the formula for calculating the correlation coefficient is:

[0100]

[0101] Where, ρ b Let be the correlation coefficient for band b; i and j are the row and column numbers of the pixel coordinates, respectively. These represent the reflectance values ​​of the hyperspectral low-frequency component and the b-th band low-frequency component of the high-resolution low-frequency component at pixel (i,j), respectively. These represent the global mean values ​​of the hyperspectral low-frequency component and the b-th band of the high-resolution low-frequency component, respectively.

[0102] Step 311: When the correlation coefficient is greater than the preset spectral threshold, retain the high-spectral low-frequency component to obtain the low-frequency fusion result;

[0103] Step 312: When the correlation coefficient is less than a preset spectral threshold, calculate the spectral angle mapping value of the band, calculate the spectral fidelity weight based on the spectral angle mapping value, and use the spectral fidelity weight to perform weighted fusion of the hyperspectral low-frequency component and the high-resolution low-frequency component to obtain the low-frequency fusion result; the calculation formula for the low-frequency fusion result is:

[0104]

[0105] in, The low-frequency fusion result; α b The spectral fidelity weights are SAM. b The spectral angle mapping value; min SAM SAM in all bands b The minimum value;

[0106] Step 313: Perform a two-level wavelet inverse transform on the high-frequency fusion result and the low-frequency fusion result to obtain the final fused image.

[0107] refer to Figure 4 The final fused image is subjected to spectral resolution refinement processing and multi-bandwidth scale simulation to obtain multispectral resolution image data, including:

[0108] Step 401: Construct the Gaussian spectral response function; the expression for the Gaussian spectral response function is:

[0109]

[0110] Among them, R a (i,j) represents the reflectance value of the a-th band at pixel (i,j); A a μ is the amplitude of the a-th band; a σ is the center wavelength of the a-th band; a The standard deviation parameter is mentioned above;

[0111] Step 402: Select the bandwidth range;

[0112] Step 403: Calculate the standard deviation parameter corresponding to each bandwidth within the bandwidth range; the formula for calculating the standard deviation parameter is: Where W is the target bandwidth;

[0113] Step 404: Update the standard deviation parameter to the Gaussian spectral response function, and generate spectral reflectance data under the bandwidth corresponding to the standard deviation parameter using the updated Gaussian spectral response function based on the final fused image, to obtain the multispectral resolution image data.

[0114] Optionally, spectral feature extraction is performed on the multispectral resolution image data to obtain vegetation-related spectral features, including:

[0115] The vegetation-related spectral features are obtained by calculating the normalized vegetation index, green vegetation index, differential vegetation index, improved soil-adjusted vegetation index, green normalized differential vegetation index, soil-adjusted vegetation index, spectral coefficient of variation, and spectral convex hull volume of the multispectral resolution image data.

[0116] refer to Figure 5 The multispectral resolution image data is subjected to adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration to obtain the final frequency features, including:

[0117] Step 601: Perform Gaussian low-pass filtering on the spectral sequence of each pixel in the multispectral resolution image data, and remove band data with a signal-to-noise ratio of less than 20dB to obtain a denoised spectral sequence.

[0118] Step 602: Perform Fourier transform on the denoised spectral sequence, calculate the spectral reconstruction error at each order step by step, and record the order at which the spectral reconstruction error first becomes less than 5%, thus obtaining the optimal order;

[0119] Step 603: Extract the harmonic coefficients and power spectrum corresponding to each order from the first order to the optimal order, and calculate the amplitude and phase based on the harmonic coefficients to obtain the harmonic decomposition feature vector;

[0120] Step 604: Calculate the energy entropy from the first order to the optimal order, the maximum frequency energy ratio, and the cross power spectrum of adjacent harmonics to obtain the fundamental frequency eigenvector;

[0121] Step 605: Calculate the mean and standard deviation of the harmonic coefficients from the first order to the optimal order of all pixels in the multispectral resolution image data window using a 3×3 sliding window to obtain the spatial enhancement feature vector;

[0122] Step 606: Merge the harmonic decomposition feature vector, the fundamental frequency feature vector, and the spatial enhancement feature vector to obtain the final frequency feature.

[0123] Specifically, the training process of the ecological variable inversion estimation model includes:

[0124] An improved DNN network is obtained by replacing the two-dimensional convolution of the basic DNN network with three-dimensional convolution and replacing the output layer of the basic DNN network with a multi-task output head.

[0125] Collect remote sensing image data and several biological indicator data from different sub-regions and time periods of the study area, extract the input feature set of the remote sensing image data, and combine the input feature set and all the biological indicator data in pairs to obtain a model training dataset including several sets of training data pairs.

[0126] The model training dataset is input into the improved DNN network for forward propagation to obtain the network output;

[0127] The loss value of each task output head is calculated using a loss function to obtain the single-index error. All single-index errors are then summed to obtain the overall error. The expression for the loss function is:

[0128]

[0129] in, The single index error of the kth task output header; The predicted value output by the k-th task output header; y k This refers to the actual indicator data corresponding to the kth task output header;

[0130] The parameters of the task output head are optimized by backpropagation using the single index error, and the hidden layer parameters of the improved DNN network are optimized by backpropagation using the overall error. Then, the process returns to the step "input the model training dataset into the improved DNN network for forward propagation to obtain the network output", thus obtaining the trained ecological variable inversion estimation model.

[0131] Specifically, remote sensing image data of the target area is acquired, including hyperspectral remote sensing data and high-resolution remote sensing images. Preprocessing operations are performed on the images, including radiometric correction, atmospheric correction, geometric correction, and cropping, to ensure image quality and unify spatial extent. Commonly used wavelet bases such as Daubechies, Symlets, and Coiflets are selected. For each candidate wavelet base, wavelet decomposition and fusion are performed on the preprocessed images. Specific steps are as follows:

[0132] Perform a one-level wavelet decomposition on each band of the hyperspectral image using the current candidate wavelet basis to obtain the low-frequency components. High-frequency components: horizontal components Vertical component diagonal components Using a single-level decomposition can reduce computational load and quickly evaluate wavelet basis performance;

[0133] For each RGB band of the high-resolution image, perform a single-level wavelet decomposition using the same wavelet basis to obtain the horizontal component. Vertical component diagonal components The advantage of high-resolution images lies in their spatial detail. However, they have less spectral information in their low-frequency components. Therefore, only the high-frequency components are extracted for fusion to enhance the spatial resolution of the fused image.

[0134] Temporary fusion strategy: Directly use the low-frequency components of the hyperspectral image, and replace the corresponding high-frequency components of the high-resolution image with the high-frequency components of the hyperspectral image. Repeat the above operation for all bands of the hyperspectral image to ensure that the low-frequency components of each band come from the hyperspectral image and the high-frequency components come from the high-resolution image. Perform a one-level inverse wavelet transform on the fused components of each band to reconstruct the temporarily fused image.

[0135] Furthermore, to quantitatively evaluate the fusion performance, spectral angle mapping and structural similarity were used, and their calculation formulas are as follows:

[0136]

[0137] The optimal basis is selected by fusing spectral angle mapping and structural similarity, and the wavelet basis with the highest number of target values ​​is selected as the wavelet basis for subsequent decomposition and fusion.

[0138] Preferably, the preprocessed hyperspectral image and high-resolution image are subjected to two-level wavelet decomposition using a selected wavelet basis. The hyperspectral image decomposition result includes: low-frequency components. Layer 1 high-frequency components: horizontal components Vertical component diagonal components Second layer high-frequency components: horizontal components Vertical component diagonal components High-resolution image decomposition results include: low-frequency components High-frequency components of layer 1: Second layer high-frequency components:

[0139] Furthermore, the local variance of the 3×3 neighborhood at each pixel of each high-frequency sub-band of the high-resolution image is calculated and normalized weights are applied; the high-frequency information of the high-resolution image and the hyperspectral image are fused according to the weights for each layer of high-frequency components.

[0140]

[0141] in, For the high-frequency information obtained through fusion; The normalized weights For the high-frequency components of the q-th decomposition layer, corresponding to and

[0142] Calculate the correlation coefficient between the hyperspectral low-frequency component and the high-resolution low-frequency component in band b. If the correlation coefficient is greater than 0.8, the hyperspectral low-frequency component is directly retained. If the correlation coefficient is less than 0.8, the spectral angle mapping value of that band is calculated, and the spectral fidelity weight is dynamically calculated. This weight is then used for weighted fusion.

[0143]

[0144] The processed low-frequency components and the weighted high-frequency components are subjected to a two-level wavelet inverse transform to generate a fused image.

[0145] Preferably, a Gaussian spectral response function is used to fit the spectral curve of each pixel in the fused image to generate spectral response data at different spectral bandwidth scales. The Gaussian function that achieves the above function is constructed as follows:

[0146]

[0147] The spectral curve of each pixel in the fused image generated above is fitted band by band to obtain the spectral response characteristics at different bandwidth scales.

[0148] The formula for determining the standard deviation in a Gaussian function is:

[0149]

[0150] Specifically, based on the aforementioned images of different spectral resolutions, operations are performed on each band to calculate vegetation-related spectral indices, including: Normalized Difference Vegetation Index (NDVI), Green Vegetation Index (GCVI), Differential Vegetation Index (DVI), Improved Soil-Adjusted Vegetation Index (MSAVI), Green Normalized Differential Vegetation Index (GNDVI), and Soil-Adjusted Vegetation Index (SAVI), as well as spectral coefficient of variation (CV) and spectral convex hull volume (CHV).

[0151] Further, the spectral sequence of each pixel in the images with different spectral resolutions obtained above is extracted. Gaussian low-pass filtering is used to remove high-frequency noise, and invalid bands with a signal-to-noise ratio less than 20dB are then removed to obtain a denoised spectral sequence. A Fourier transform is performed on the spectral sequence, gradually increasing the harmonic order from the first order. The spectral reconstruction error at each order is calculated, and the optimal order is determined when the error increase is less than 5%. The corresponding harmonic coefficients, amplitudes, phases, and power spectra are extracted. Energy entropy, the dominant frequency energy percentage, and the cross-power spectrum of adjacent harmonics are calculated to construct a fundamental frequency feature vector. A 3×3 sliding window is used, centered on the current pixel, to calculate the mean and standard deviation of the harmonic coefficients of all pixels within the window to generate spatial enhancement features. The fundamental frequency features and spatial enhancement features are merged to form the final frequency feature vector, which is then combined with the spectral features to form the input feature set.

[0152] Preferably, a 3D convolutional kernel is used to replace the 2D convolution in the traditional DNN network, and an independent output head is set for each biological indicator to obtain an improved DNN network. A batch of input data is used, and forward propagation is performed to obtain the predicted values ​​for each task. The loss for each task and the total loss are calculated. The output head parameters are individually optimized using the loss for each task, and the shared parameters are integrated and optimized using the total loss until the model accuracy meets the preset requirements and the model is output.

[0153] The beneficial effects of this invention are as follows:

[0154] This invention balances spectral fidelity and spatial structure preservation through adaptive basis selection and multi-scale wavelet decomposition, achieving differentiated fusion of high and low frequency components at different scales and improving the complementarity of multi-source data and the quality of fused images. By refining spectral resolution and simulating at multiple bandwidth scales, it overcomes the limitations of single resolution, enhancing the ability to capture complex spectral features of vegetation. Spatial correlation is enhanced through spatial neighborhood feature fusion. Finally, the ecological variable inversion estimation model enables simultaneous prediction of multiple indicators, improving the model's versatility.

[0155] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0156] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A biodiversity inversion method based on multi-source remote sensing image fusion, characterized in that, include: Acquire remote sensing image data of the target area; The remote sensing image data includes: hyperspectral remote sensing images and high-resolution remote sensing images; The remote sensing image data is preprocessed; the preprocessing includes: radiometric correction, atmospheric correction, geometric correction, and cropping. The preprocessed remote sensing image data is subjected to wavelet basis adaptive selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image. The final fused image is subjected to spectral resolution refinement processing and multi-bandwidth scale simulation to obtain multispectral resolution image data; the multispectral resolution image includes: several sets of simulated images with different spectral resolutions; Spectral features are extracted from the multispectral resolution image data to obtain vegetation-related spectral features; The multispectral resolution image data is subjected to adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration to obtain the final frequency features. The vegetation-related spectral features and the final frequency features are combined to obtain the input feature set; The input feature set is fed into a pre-trained ecological variable inversion estimation model to generate data and obtain the predicted value of the target ecological variable.

2. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 1, characterized in that, The preprocessed remote sensing image data undergoes adaptive basis selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image, including: Pre-selected set of preliminary wavelet bases; The preprocessed remote sensing image data is decomposed and fused using each candidate wavelet basis in the pre-set wavelet basis set to obtain a temporary fused image. The spectral similarity index is obtained by performing spectral angle mapping data calculation on the temporary fused image; the calculation formula for the spectral similarity index is as follows: Wherein, SAM(p,q) is the spectral similarity index; p is the pixel spectral vector of the preprocessed hyperspectral remote sensing image; q is the pixel spectral vector of the temporary fused image; arccos(·) represents the inverse cosine function; Structural similarity is calculated on the temporary fused image to obtain a spatial quality index; the formula for calculating the spatial quality index is as follows: Wherein, SSIM(x,y) is the spatial quality index; x and y are the image patches to be compared between the preprocessed high-resolution remote sensing image and the temporary fused image, respectively; μ x μ y These are the pixel mean values ​​for x and y, respectively; σ represents the pixel variances of x and y, respectively. xy Let x be the covariance of x and y; C1 and C2 are the first and second constant terms, respectively; The spectral similarity index and the spatial quality index are weighted and calculated to obtain the band score; the formula for calculating the band score is as follows: Where ω1+ω2=1; Score is the band score; ω1 and ω2 are the first weight and the second weight, respectively; The band scores are sorted in descending order to obtain a score order set, and the candidate wavelet bases corresponding to the target number of band scores in the score order set are extracted sequentially to obtain the optimal wavelet base set.

3. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 2, characterized in that, Using each candidate wavelet basis in the pre-selected wavelet basis set, wavelet decomposition and fusion are performed on the preprocessed remote sensing image data to obtain a temporary fused image, including: The preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image are decomposed using each candidate wavelet basis in the pre-set wavelet basis set. The low-frequency portion of the decomposed hyperspectral remote sensing image and the high-frequency portion of the high-resolution remote sensing image are selected to obtain temporary fused data. The temporary fused data is subjected to a single-level wavelet inverse transform to obtain the temporary fused image.

4. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 2, characterized in that, The preprocessed remote sensing image data undergoes adaptive basis selection, multi-scale wavelet decomposition, fusion of low-frequency and high-frequency components, and inverse wavelet transform reconstruction to obtain the final fused image, including: The preprocessed hyperspectral remote sensing image and the high-resolution remote sensing image are subjected to two-level wavelet decomposition using the optimal wavelet basis set to obtain hyperspectral decomposition results and high-resolution decomposition results. The hyperspectral decomposition results include: a hyperspectral low-frequency component, a first hyperspectral high-frequency component, and a second hyperspectral high-frequency component. The high-resolution decomposition results include: a high-resolution low-frequency component, a first high-resolution high-frequency component, and a second high-resolution high-frequency component. The local variance of each pixel in the 3×3 neighborhood of the preprocessed high-resolution remote sensing image is calculated and normalized to obtain the local fusion weights. The high-frequency components within the hyperspectral decomposition result and the high-resolution decomposition result are weighted and fused using the local fusion weights to obtain a high-frequency fusion result. The correlation coefficient is obtained by calculating the correlation between the hyperspectral low-frequency component and the high-resolution low-frequency component; the formula for calculating the correlation coefficient is as follows: Where, ρ b Let be the correlation coefficient for band b; i and j are the row and column numbers of the pixel coordinates, respectively. These represent the reflectance values ​​of the hyperspectral low-frequency component and the b-th band low-frequency component of the high-resolution low-frequency component at pixel (i,j), respectively. These represent the global mean values ​​of the hyperspectral low-frequency component and the b-th band of the high-resolution low-frequency component, respectively. When the correlation coefficient is greater than the preset spectral threshold, the high-spectral low-frequency component is retained to obtain the low-frequency fusion result; When the correlation coefficient is less than a preset spectral threshold, the spectral angle mapping value of the band is calculated. Based on the spectral angle mapping value, a spectral fidelity weight is calculated. Then, the spectral fidelity weight is used to perform weighted fusion of the hyperspectral low-frequency component and the high-resolution low-frequency component to obtain the low-frequency fusion result. The formula for calculating the low-frequency fusion result is as follows: in, The low-frequency fusion result; α b The spectral fidelity weights are SAM. b The spectral angle mapping value; min SAM SAM in all bands b The minimum value; The high-frequency fusion result and the low-frequency fusion result are subjected to a two-level wavelet inverse transform to obtain the final fused image.

5. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 4, characterized in that, The final fused image is subjected to spectral resolution refinement processing and multi-bandwidth scale simulation to obtain multispectral resolution image data, including: Construct a Gaussian spectral response function; the expression of the Gaussian spectral response function is: Among them, R a (i,j) represents the reflectance value of the a-th band at pixel (i,j); A a μ is the amplitude of the a-th band; a σ is the center wavelength of the a-th band; a The standard deviation parameter; Select the bandwidth range; Calculate the standard deviation parameter corresponding to each bandwidth within the bandwidth range; the formula for calculating the standard deviation parameter is: Where W is the target bandwidth; The standard deviation parameter is updated in the Gaussian spectral response function, and spectral reflectance data under the bandwidth corresponding to the standard deviation parameter is generated using the updated Gaussian spectral response function based on the final fused image to obtain the multispectral resolution image data.

6. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 5, characterized in that, Spectral features are extracted from the multispectral resolution image data to obtain vegetation-related spectral features, including: The vegetation-related spectral features are obtained by calculating the normalized vegetation index, green vegetation index, differential vegetation index, improved soil-adjusted vegetation index, green normalized differential vegetation index, soil-adjusted vegetation index, spectral coefficient of variation, and spectral convex hull volume of the multispectral resolution image data.

7. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 6, characterized in that, Adaptive harmonic decomposition, extended frequency domain feature calculation, spatial neighborhood feature fusion, and feature integration are performed on the multispectral resolution image data to obtain the final frequency features, including: Gaussian low-pass filtering is applied to the spectral sequence of each pixel in the multispectral resolution image data, and band data with a signal-to-noise ratio of less than 20dB are removed to obtain a denoised spectral sequence. Perform a Fourier transform on the denoised spectral sequence, calculate the spectral reconstruction error at each order step by step, and record the order at which the spectral reconstruction error first becomes less than 5%, thus obtaining the optimal order. Harmonic coefficients and power spectra corresponding to each order from the first order to the optimal order are extracted, and the amplitude and phase are calculated based on the harmonic coefficients to obtain the harmonic decomposition feature vector; Calculate the energy entropy from the first order to the optimal order, the maximum frequency energy ratio, and the cross power spectrum of adjacent harmonics to obtain the fundamental frequency eigenvector. The mean and standard deviation of the harmonic coefficients from the first order to the optimal order of all pixels within the multispectral resolution image data window are calculated using a 3×3 sliding window to obtain the spatial enhancement feature vector; The final frequency feature is obtained by merging the harmonic decomposition feature vector, the fundamental frequency feature vector, and the spatial enhancement feature vector.

8. The biodiversity inversion method based on multi-source remote sensing image fusion according to claim 7, characterized in that, The training process of the ecological variable inversion estimation model includes: An improved DNN network is obtained by replacing the two-dimensional convolution of the basic DNN network with three-dimensional convolution and replacing the output layer of the basic DNN network with a multi-task output head. Collect remote sensing image data and several biological indicator data from different sub-regions and time periods of the study area, extract the input feature set of the remote sensing image data, and combine the input feature set and all the biological indicator data in pairs to obtain a model training dataset including several sets of training data pairs. The model training dataset is input into the improved DNN network for forward propagation to obtain the network output; The loss value of each task output head is calculated using a loss function to obtain the single-index error. All single-index errors are then summed to obtain the overall error. The expression for the loss function is: in, The single index error of the kth task output header; The predicted value output by the k-th task output header; y k This refers to the actual indicator data corresponding to the output header of the kth task. The parameters of the task output head are optimized by backpropagation using the single index error, and the hidden layer parameters of the improved DNN network are optimized by backpropagation using the overall error. Then, the process returns to the step "input the model training dataset into the improved DNN network for forward propagation to obtain the network output", thus obtaining the trained ecological variable inversion estimation model.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a biodiversity inversion method based on multi-source remote sensing image fusion as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute any one of claims 1 to 8, a biodiversity inversion method based on multi-source remote sensing image fusion.

Images