Method for identifying occurrence degree of dendrolimus punctatus walker based on spectral reconstruction

Abstract

The present application relates to remote sensing image processing technical field, especially to a kind of based on spectral reconstruction's fall leaf larch moth pest occurrence degree identification method, obtain the real multispectral image of forest area disturbed by fall leaf larch moth, and the ground hyperspectral data of the local area of this forest area;The real multispectral image and ground hyperspectral data obtained in step are consistent, and the simulated multispectral curve of local area is obtained;Based on simulated multispectral curve, the real multispectral image is interpolated and reconstructed, and the simulated hyperspectral image of all forest area is obtained;Sensitive band and sensitive spectral feature sensitive to the disturbance of fall leaf larch moth are extracted from simulated hyperspectral image;Using ensemble learning strategy, the degree of infestation of all areas of forest is analyzed using sensitive band and sensitive spectral feature.The present application combines spectral reconstruction and ensemble learning strategy to identify the degree of infestation of fall leaf larch moth, and improves the quality of spectral reconstruction and pest identification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and in particular relates to a method for identifying the degree of occurrence of larch caterpillar infestation based on spectral reconstruction. Background Technology

[0002] Forest pest monitoring is a crucial technological link in forest ecological management and pest disaster early warning. Remote sensing spectral information is widely used for pest severity identification because it can reflect changes in vegetation physiological and biochemical characteristics. Hyperspectral sensors can record the full-spectrum reflectance characteristics of vegetation over continuous narrow bands and are highly sensitive to early pest physiological stresses such as chlorophyll content, flavonoid changes, and water content decline. However, their application in large-scale pest monitoring is limited by spatial resolution and data acquisition costs. In contrast, multispectral sensors (such as Sentinel-2) achieve regional-scale monitoring with higher spatial resolution and more frequent revisit cycles. However, due to their limited discrete bands, their spectral resolution is insufficient, making it difficult to detect subtle spectral perturbations caused by early pests. Therefore, leveraging the advantages of ground-based hyperspectral sensors to enhance the pest identification capabilities of multispectral remote sensing has become a key technological means for high-precision identification of pest severity.

[0003] Previous spectral reconstruction techniques have largely focused on modeling spectral relationships between sensors, pre-simulating remote sensing images, and analyzing product consistency, with limited optimization for the specific application of pest monitoring. Furthermore, existing spectral reconstruction methods often rely on statistical models or fixed convolution kernels, which are insufficient for describing spectral continuity and making it difficult to accurately capture subtle spectral changes in the early stages of pest infestation. Therefore, developing a simulated spectral and pest identification technology that combines physical consistency and spectral sensitivity remains of significant research value.

[0004] Current methods for identifying larch caterpillar pests primarily rely on single data sources such as satellite multispectral data and terrestrial hyperspectral data. Multispectral data is readily available and has broad coverage, making it widely used for regional-scale pest monitoring. Researchers typically differentiate pest severity based on several visible, red-edge, and near-infrared bands using vegetation indices, texture features, or traditional machine learning methods such as random forests and support vector machines. However, due to the limited number of multispectral bands and low spectral resolution, multispectral images lack sensitivity to biochemical changes in lightly affected trees (such as chlorophyll decline and red-edge drift), hindering early pest identification. Furthermore, the sparsity of multispectral imagery limits its ability to support detailed classification and grading of forest stand conditions over large areas. Terrestrial hyperspectral data offers higher spectral resolution and can accurately capture the physiological information of affected trees, thus finding wider application in refined pest identification. Existing techniques often use hyperspectral reflectance, derivative spectra, or characteristic spectral indices, employing models such as support vector machines, neural networks, and random forests to determine the pest severity of samples. However, ground-based hyperspectral data can only cover a limited number of sample plots, making it difficult to promote on a large regional scale. Furthermore, the collection cost is high and the time span is limited, making it unsuitable for the needs of rapid, continuous, and wide-area pest monitoring and early warning. Although both types of data can be used for pest identification, their insufficient spectral and spatial resolution makes it difficult to meet the needs of large-scale forest areas for high-precision, low-cost, and continuous identification of pest occurrence levels. Summary of the Invention

[0005] In view of this, the present invention aims to provide a method for identifying the severity of larch caterpillar infestation based on spectral reconstruction. Spectral reconstruction based on ground hyperspectral and multispectral remote sensing data can obtain continuous and high-resolution simulated spectra, which is significantly superior to the band sparsity features of traditional multispectral methods, reducing spectral peak misalignment and shape deviation, thereby improving the sensitivity to spectral differences in affected areas. In addition, by combining spectral reconstruction with ensemble learning strategies to identify the severity of larch caterpillar infestation, the quality of spectral reconstruction and the accuracy of pest identification are improved.

[0006] To achieve the above objectives, the technical solution created by this invention is implemented as follows:

[0007] A method for identifying the severity of larch caterpillar infestation based on spectral reconstruction includes:

[0008] S1: Acquire real multispectral images of forest areas infested by larch caterpillars, as well as ground hyperspectral data of local areas within these forest areas;

[0009] S2: Perform consistency processing on the real multispectral image and ground hyperspectral data obtained in step S1 to obtain the simulated multispectral curve of the local area;

[0010] S3: Based on the simulated multispectral curves obtained in step S2, interpolate and reconstruct the real multispectral images to obtain simulated hyperspectral images of the entire forest area.

[0011] S4: Extract the sensitive bands and sensitive spectral features that are sensitive to larch caterpillar infestation from the simulated hyperspectral image obtained in step S3.

[0012] S5: Using the sensitive bands and sensitive spectral characteristics obtained in step S4, analyze the degree of insect pest occurrence in all areas of the forest.

[0013] Furthermore, step S1 also includes: preprocessing the ground hyperspectral data, including noise filtering, removal of abnormal spectral curves, and resampling.

[0014] Furthermore, step S2 includes:

[0015] S21: Based on the spectral response function of real multispectral images, energy redistribution is performed on ground hyperspectral data to obtain simulated multispectral reflectance;

[0016] S22: Perform partial least squares regression on the real multispectral image and the simulated multispectral reflectance obtained in step S21 to capture the linear and weakly nonlinear relationship between the simulated multispectral reflectance and the real multispectral image in each band, and obtain the simulated multispectral curve.

[0017] Furthermore, in step S21, the simulated multispectral reflectance is obtained using the following formula:

[0018] ;

[0019] Where, η i f represents the simulated multispectral reflectance of the i-th band. i Let ρ(λ) represent the spectral response function of the i-th band, and let ρ(λ) represent the ground hyperspectral reflectance at wavelength λ. max and λ min These represent the upper and lower wavelength limits of the spectral value range of the real multispectral image, respectively.

[0020] Furthermore, step S3 includes: inputting the simulated multispectral curve and the real multispectral image into a one-dimensional convolutional model. The one-dimensional convolutional model captures the local spectral variation trend between multispectral bands based on the simulated multispectral curve, and performs interpolation reconstruction on the real multispectral image to obtain the simulated hyperspectral image.

[0021] Furthermore, the one-dimensional convolutional model includes, along the data processing direction: an input layer for inputting simulated multispectral curves and real multispectral images; a first adaptive convolutional layer for automatically adjusting the receptive field based on the internal correlation of different bands in the simulated multispectral curve, capturing local spectral variation trends between multispectral bands in the simulated multispectral curve, and obtaining high-dimensional spectral features; a fully connected layer for globally fusing the high-dimensional spectral features and extracting comprehensive spectral features containing local variations, trend variations, and material absorption features from the fused features; a second adaptive convolutional layer for further extracting detailed spectral features from the comprehensive spectral features; multiple fully connected layers for interpolating and fusing the detailed spectral features with the real multispectral images to complete the interpolation reconstruction of the real multispectral images; and an output layer for outputting a high-resolution simulated hyperspectral image.

[0022] Furthermore, in step S4: the average reflectance of different pest infestation levels and multiple spectral indices are extracted from the simulated hyperspectral image; the correlation coefficient between the average reflectance of each spectral band and the larch leaf loss rate is calculated, and the band with the high correlation coefficient is selected as the sensitive band; the sensitive spectral index is selected from multiple spectral indices using analysis of variance.

[0023] Furthermore, in step S5, a soft-voting ensemble model is used to output the degree of disaster in the affected area based on the input sensitive bands and sensitive spectral characteristics.

[0024] Furthermore, the soft-voting ensemble model includes a Naive Bayes classifier and a random forest classifier. The process of the soft-voting ensemble model outputting the degree of disaster includes: the Naive Bayes classifier and the random forest classifier make predictions based on sensitive band and sensitive spectral features, respectively, to obtain the category probability distribution of their respective degrees of disaster; the category probability distributions output by the two classifiers are weighted and averaged through soft voting, and the category with the highest weighted probability value is taken as the final classification result of the degree of disaster.

[0025] Furthermore, during the training of the Naive Bayes classifier and the Random Forest classifier: following step S4, sensitive bands and sensitive spectral features are extracted from hyperspectral images of different disaster-stricken areas with known disaster severity categories; a dataset is constructed using the sensitive bands and sensitive spectral features, and the corresponding disaster severity categories of the disaster-stricken areas; the dataset is stratified according to the proportion of pest infestation levels and divided into five subsets in equal proportion; during each training process, four subsets are selected as the training set, and the remaining subset is selected as the validation set. The two classifiers are trained using the training set, and the two classifiers obtained from the current training are validated using the validation set; the above training operation is repeated until all five subsets have been used as validation sets, at which point the training of the two classifiers is completed.

[0026] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0027] This invention presents a method for identifying the severity of larch caterpillar infestation based on spectral reconstruction. It eliminates the need for expensive and spatiotemporally limited satellite hyperspectral data, instead constructing a high-precision simulated hyperspectral image at the regional scale using only terrestrial hyperspectral data and readily available multispectral imagery. This significantly reduces monitoring costs and improves the applicability of the technology in different forest areas and seasons. Furthermore, the invention utilizes a soft-voting ensemble learning model to significantly advance the timing of pest identification, providing forestry departments with more time for prevention and control, and achieving efficient early warning. By dynamically extracting simulated hyperspectral data and sensitive features, and combining this with hierarchical five-fold cross-validation to enhance the robustness of the identification model, this invention can monitor different levels of pest occurrence in forests ranging from healthy to severely affected areas, accurately reflecting pest spread trends and providing support for the scientific development of control measures. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0029] Figure 1 A schematic flowchart of the method for identifying the degree of larch caterpillar infestation based on spectral reconstruction as described in the embodiments of the present invention;

[0030] Figure 2 This is a schematic diagram of the one-dimensional convolution model described in the embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0032] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] like Figure 1 As shown in the embodiments of the present invention, the method for identifying the degree of larch caterpillar infestation based on spectral reconstruction includes:

[0034] S1: Acquire real multispectral images of forest areas infested by larch caterpillars, as well as ground hyperspectral data of local areas within these forest areas.

[0035] In some embodiments, step S1 further includes: preprocessing the ground hyperspectral data, including noise filtering, outlier spectral curve removal, and resampling.

[0036] In this embodiment of the invention, a ground-based spectrometer is used to sample local areas of forest areas affected by larch caterpillars to obtain ground hyperspectral data; the real multispectral image is a multispectral image captured by a multispectral sensor (such as Sentinel-2) and preprocessing operations such as radiometric calibration, atmospheric correction, resampling and band synthesis are also performed on the real multispectral image.

[0037] S2: Perform consistency processing on the real multispectral image and ground hyperspectral data obtained in step S1 to obtain the simulated multispectral curve of the local area. By performing consistency processing on the real multispectral image and ground hyperspectral data, the real multispectral image and ground hyperspectral data become comparable, thus completing the correction of the ground hyperspectral data.

[0038] In some embodiments, step S2 includes:

[0039] S21: Based on the spectral response function of the real multispectral image, the energy of the ground hyperspectral data is redistributed using the following formula to ensure that the spectral characteristics of the ground hyperspectral data are consistent with the spectral response of the real multispectral image, avoiding spectral shift or energy mismatch caused by differences in imaging systems, thus obtaining the simulated multispectral reflectance:

[0040] ;

[0041] Where, η i f represents the simulated multispectral reflectance of the i-th band. i Let ρ(λ) represent the spectral response function of the i-th band, and let ρ(λ) represent the ground hyperspectral reflectance at wavelength λ. max and λ min These represent the upper and lower wavelength boundaries of the spectral value range of the real multispectral image, respectively.

[0042] S22: Perform partial least squares regression on the real multispectral image and the simulated multispectral reflectance obtained in step S21 to capture the linear and weakly nonlinear relationship between the simulated multispectral reflectance and the real multispectral image in each band, and obtain the simulated multispectral curve.

[0043] S3: Based on the simulated multispectral curves obtained in step S2, interpolate and reconstruct the real multispectral images to obtain simulated hyperspectral images of the entire forest area.

[0044] In some embodiments, step S3 includes: inputting the simulated multispectral curve and the real multispectral image into a one-dimensional convolutional model, wherein the one-dimensional convolutional model captures the local spectral variation trend between multispectral bands based on the simulated multispectral curve, and performs interpolation reconstruction on the real multispectral image to obtain the simulated hyperspectral image.

[0045] This invention constructs a one-dimensional convolution model with an adaptive convolution kernel to map simulated multispectral reflectance to simulated hyperspectral images, achieving regional-scale spectral reconstruction. Specifically, in some embodiments, the one-dimensional convolution model is as follows: Figure 2 As shown, the model includes an input layer, a first adaptive convolutional layer, a fully connected layer, a second adaptive convolutional layer, multiple fully connected layers, and an output layer along the data processing direction. The input layer takes simulated multispectral curves and real multispectral images as input. The first adaptive convolutional layer automatically adjusts the receptive field based on the internal correlations of different bands in the simulated multispectral curve, capturing local spectral variation trends between multispectral bands to obtain high-dimensional spectral features. The fully connected layer globally fuses these high-dimensional spectral features and extracts comprehensive spectral features from the fused features, including local variations, trend changes, and material absorption characteristics. The second adaptive convolutional layer further extracts detailed spectral features from the comprehensive spectral features. The multiple fully connected layers interpolate and fuse the detailed spectral features with the real multispectral image, completing the interpolation reconstruction of the real multispectral image. The output layer outputs a high-resolution simulated hyperspectral image. The second adaptive convolutional layer enables the model to learn higher-order spectral variation relationships, especially subtle differences in absorption bands related to physiological indicators (such as the red edge and near-infrared plateau).

[0046] In this embodiment of the invention, the multi-layer fully connected layer is specifically configured as a two-layer fully connected layer. This two-layer fully connected layer further fuses detailed spectral features with real multispectral images to adapt to the continuous reconstruction of the 350-2500nm hyperspectral range. Furthermore, this embodiment of the invention also includes a flattening layer between the first adaptive convolutional layer and the fully connected layer, and between the second adaptive convolutional layer and the first fully connected layer of the multi-layer fully connected layer. The flattening layer expands the multidimensional tensor output of the adaptive convolutional layer into a one-dimensional vector before inputting it into the fully connected layer. This flattening layer improves the trainability of the model and provides a unified input format for high-dimensional feature fusion. Specifically, this embodiment of the invention employs reflectance curve fitting loss, and the Adam optimization method is used for model training. The network gradually learns the mapping relationship from multispectral to hyperspectral.

[0047] In other embodiments, step S3 can also employ a spectral interpolation method based on a physical radiative transfer model, a principal component inversion model, a sparse coding reconstruction method, or a dictionary learning method using a prior spectral library to generate simulated hyperspectral images. Specifically: in the spectral interpolation method based on a physical radiative transfer model, real multispectral images are used as input, combined with imaging geometry conditions and prior parameters of the land surface or vegetation, and the radiative transfer model is used to perform physically consistent spectral interpolation and expansion of the multispectral discrete bands, outputting hyperspectral reflectance data for continuous bands; in the spectral reconstruction method based on a principal component inversion model, real multispectral images are used as input, combined with imaging geometry conditions and prior parameters of the land surface or vegetation, and the radiative transfer model is used to perform physically consistent spectral interpolation and expansion of the multispectral discrete bands, outputting hyperspectral reflectance data for continuous bands; Using hyperspectral training samples as input, principal component basis functions are constructed, corresponding principal component coefficients are inverted, and simulated hyperspectral images with continuous spectra are generated through principal component reconstruction. In the sparse coding-based hyperspectral reconstruction method, real multispectral images are used as input, combined with a pre-built hyperspectral dictionary, to solve for spectral representation coefficients under sparse constraints, and then reconstruct the hyperspectral reflectance of continuous bands. In the dictionary learning method based on a priori spectral database, real multispectral images and a priori spectral database are used as input, and simulated hyperspectral images are generated under the condition of satisfying multispectral observation constraints through dictionary learning and linear combination reconstruction. These methods can also acquire continuous spectral information without the need for satellite hyperspectral data, which can be used to express pest susceptibility features.

[0048] S4: Extract the sensitive bands and sensitive spectral features that are sensitive to larch caterpillar infestation from the simulated hyperspectral image obtained in step S3.

[0049] In some embodiments, step S4 includes:

[0050] S41: From the simulated hyperspectral image, the average reflectance at different levels of insect infestation, as well as multiple spectral indices, are extracted. In this embodiment of the invention, 34 insect-related spectral indices, such as the Normalized Difference Vegetation Index (NDVI), Normalized Burn Index (NBR), and Enhanced Vegetation Index (EVI), are specifically extracted, along with the first derivative of the average reflectance.

[0051] S42: Calculate the correlation coefficient between the average reflectance of each spectral band in step S41 and the leaf loss rate of larch, and select the band with the high average reflectance as the sensitive band.

[0052] This invention selects the larch leaf loss rate as an indicator for observing the severity of larch caterpillar infestation and incorporates it into the calculation of correlation coefficients. The larch leaf loss rate is calculated using the number of damaged leaves and the number of healthy leaves within the spectral measurement range, specifically obtained by the following formula:

[0053] ;

[0054] Where LLR represents the larch leaf loss rate, n represents the total number of larch branches, and NLi and N Hi These represent the number of damaged leaves and the number of healthy leaves on the i-th branch of a larch tree, respectively. In this embodiment of the invention, the average leaf loss rate of the upper, middle, and lower layers of branches in the current larch tree is calculated as the current leaf loss rate of the larch tree. If the leaf loss rate (LLR) is less than 25%, the current larch tree is considered healthy; if the LLR is between 25% and 50%, the current larch tree is considered slightly damaged; if the LLR is between 50% and 75%, the current larch tree is considered moderately damaged; and if the LLR is greater than 75%, the current larch tree is considered severely damaged.

[0055] The correlation coefficient between the average reflectance and leaf loss rate of each spectral band is calculated. When the absolute value of the correlation coefficient is greater than the preset value r (set to 0.6 in this embodiment, i.e. r≥0.6), it is considered that the band has a strong correlation with the degree of pest occurrence, and the band is defined as a sensitive band.

[0056] S43: Using analysis of variance, select the sensitive spectral index from the multiple spectral indices obtained in step S41.

[0057] In this embodiment of the invention, the mean differences of different disaster levels across different bands are tested using one-way ANOVA. The spectral indices with significant levels of severity are determined to be sensitive features with significant discriminative power. Specifically, this includes: extracting 34 spectral indices from simulated hyperspectral images, using different pest damage levels as grouping conditions, and employing one-way ANOVA to test the mean differences of each spectral index under different disaster levels, as shown in the following formula:

[0058] ;

[0059] ;

[0060] Among them, SS between SS represents the between-group squares. within The square within the group represents the degree of disaster, k represents the degree of disaster, and n represents the degree of disaster. i This represents the number of samples in the i-th group. Let x represent the mean of the i-th sample group. ij This represents the sample value of the j-th sample in the i-th sample group;

[0061] The following formula is used to compare between-group and within-group differences to assess its discriminative power and obtain the significance level:

[0062] ;

[0063] p=P(f (k-1,N-k) ≥F);

[0064] Where N represents the total number of samples, F represents the difference parameter, and f (k-1,N-k) This represents the probability distribution that satisfies the difference parameter F when the numerator has k-1 degrees of freedom and the denominator has Nk degrees of freedom, where P represents probability and p represents significance level.

[0065] When the significance level p < 0.05, the difference in the mean values ​​of the corresponding spectral indices for different levels of damage is considered statistically significant. This embodiment of the invention ultimately selected eight insect-sensitive spectral features for subsequent analysis.

[0066] S5: Using an ensemble learning strategy, the degree of pest occurrence in all areas of the forest is analyzed by utilizing the sensitive bands and sensitive spectral features obtained in step S4.

[0067] In some embodiments, a soft-voting ensemble model is used as an ensemble learning strategy to output the degree of disaster in the affected area based on the input sensitive band and sensitive spectral features. Specifically, the soft-voting ensemble model includes a Naive Bayes classifier and a random forest classifier. The process of the soft-voting ensemble model outputting the degree of disaster includes: the Naive Bayes classifier and the random forest classifier predict the degree of disaster based on the sensitive band and sensitive spectral features respectively, obtaining their respective class probability distributions; the class probability distributions output by the two classifiers are weighted and averaged using a soft-voting method, and the class with the highest weighted probability value is taken as the final classification result of the degree of disaster.

[0068] In this embodiment of the invention, the main parameter var_smoothing of the Naive Bayes classifier is set to the default value of 10. -9 This is used to add a minimum value when calculating the feature variance to prevent the denominator from being zero. The main parameters of the random forest classifier include ntree being 500 and mtry being 10. In addition, the weight parameters of the soft voting ensemble model need to be set to allocate the contribution of each base model in the ensemble process. Specifically, the random forest classifier is assigned a weight of 0.7 and the random forest classifier is assigned a weight of 0.3 to enhance the stability of the ensemble results and the classification performance.

[0069] In some embodiments, a hierarchical five-fold cross-validation method is used to train the two classifiers to ensure that the final soft-voting ensemble model's performance analysis results are more stable. Specifically: during the training of the Naive Bayes classifier and the Random Forest classifier:

[0070] According to step S4, sensitive bands and sensitive spectral features are extracted from hyperspectral images of different disaster-stricken areas with known disaster severity categories.

[0071] A dataset is constructed based on sensitive bands and sensitive spectral features, as well as the corresponding disaster severity categories of the affected areas;

[0072] The dataset was stratified according to the proportion of pest severity and then divided into five subsets in equal proportion;

[0073] In each training process, four subsets are selected as the training set and the remaining subset is selected as the validation set. The training set is used to train the two classifiers, and the validation set is used to validate the two classifiers obtained in the current training.

[0074] Repeat the above training steps until all five subsets have been validated once. At this point, the training of the two classifiers is complete.

[0075] In other embodiments, other ensemble learning strategies such as stacked ensemble, weighted voting, gradient boosting decision trees, random forests, lightweight Transformers, or one-dimensional convolutional neural networks can be used to analyze the degree of pest infestation in all areas of the forest. The input and output of these methods are consistent with the soft-voting ensemble model: the input is sensitive bands and sensitive spectral features, and the output is the probability distribution of the severity of the infestation. These methods can also analyze the degree of pest infestation and classify forest stands into healthy, lightly, moderately, and severely affected areas.

[0076] This invention overcomes the limitations of traditional remote sensing pest monitoring, which relies on high-cost satellite hyperspectral data. It achieves the generation of simulated hyperspectral data at a regional scale through ground-based hyperspectral-driven multispectral reconstruction technology. This method utilizes spectral response function correction, multi-source spectral matching, and continuous spectral reconstruction techniques to ensure that the results closely approximate real hyperspectral data in terms of spectral morphology, energy response, and feature sensitivity, providing high-dimensional, continuous, and physically consistent spectral information for pest identification.

[0077] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0078] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for identifying the degree of larch caterpillar infestation based on spectral reconstruction, characterized in that, include: S1: Acquire real multispectral images of forest areas infested by larch caterpillars, as well as ground hyperspectral data of local areas within these forest areas; S2: Perform consistency processing on the real multispectral image and ground hyperspectral data obtained in step S1 to obtain the simulated multispectral curve of the local area; step S2 includes: S21: Based on the spectral response function of real multispectral images, energy redistribution is performed on ground hyperspectral data to obtain simulated multispectral reflectance; S22: Perform partial least squares regression on the real multispectral image and the simulated multispectral reflectance obtained in step S21 to capture the linear and weakly nonlinear relationship between the simulated multispectral reflectance and the real multispectral image in each band, and obtain the simulated multispectral curve. S3: Based on the simulated multispectral curves obtained in step S2, interpolate and reconstruct the real multispectral images to obtain simulated hyperspectral images of the entire forest area. S4: From the simulated hyperspectral image obtained in step S3, extract the sensitive bands and sensitive spectral features that are sensitive to larch caterpillar infestation; In step S4: Extract the average reflectance and multiple spectral indices for different infestation levels from the simulated hyperspectral image; Calculate the correlation coefficient between the average reflectance of each spectral band and the larch leaf loss rate, and select the band with the high correlation coefficient as the sensitive band; Use analysis of variance to screen out the sensitive spectral indices from multiple spectral indices; S5: Using an ensemble learning strategy, the degree of pest occurrence in all areas of the forest is analyzed by utilizing the sensitive bands and sensitive spectral features obtained in step S4.

2. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 1, characterized in that, Step S1 also includes: preprocessing the ground hyperspectral data, including noise filtering, removal of abnormal spectral curves, and resampling.

3. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 1, characterized in that, In step S21, the simulated multispectral reflectance is obtained using the following formula: ； Where, η i f represents the simulated multispectral reflectance of the i-th band. i Let ρ(λ) represent the spectral response function of the i-th band, and let ρ(λ) represent the ground hyperspectral reflectance at wavelength λ. max and λ min These represent the upper and lower wavelength limits of the spectral value range of the real multispectral image, respectively.

4. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 1, characterized in that, Step S3 includes: The simulated multispectral curve and the real multispectral image are input into the one-dimensional convolutional model. The one-dimensional convolutional model captures the local spectral variation trend between multispectral bands based on the simulated multispectral curve, and performs interpolation and reconstruction on the real multispectral image to obtain the simulated hyperspectral image.

5. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 4, characterized in that, One-dimensional convolutional models include those along the data processing direction: The input layer is used to input simulated multispectral curves and real multispectral images; The first adaptive convolutional layer is used to automatically adjust the receptive field according to the internal correlation of different bands in the simulated multispectral curve, capture the local spectral variation trend between multispectral bands in the simulated multispectral curve, and obtain high-dimensional spectral features. The fully connected layer globally fuses high-dimensional spectral features and extracts comprehensive spectral features from the fused features, including local variations, trend variations, and material absorption characteristics. The second adaptive convolutional layer further extracts detailed spectral features from the comprehensive spectral features; Multi-layer fully connected layers interpolate and fuse detailed spectral features with real multispectral images to complete the interpolation reconstruction of real multispectral images; The output layer outputs high-resolution analog hyperspectral images.

6. In step S5, a soft-voting ensemble model is used as an ensemble learning strategy to output the degree of disaster in the affected area based on the input sensitive band and sensitive spectral characteristics.

7. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 6, characterized in that, The soft voting ensemble model includes a Naive Bayes classifier and a random forest classifier. The process by which the soft voting ensemble model outputs the degree of disaster includes: The Naive Bayes classifier and the Random Forest classifier predict the disaster severity based on the sensitive band and sensitive spectral features, respectively, and obtain their respective category probability distributions. The probability distributions of the two classifiers are weighted and averaged using a soft voting method. The class with the highest weighted probability value is then used as the final classification result for the degree of disaster.

8. The method for identifying the degree of larch caterpillar infestation based on spectral reconstruction according to claim 7, characterized in that, During the training of the Naive Bayes classifier and the Random Forest classifier: According to step S4, sensitive bands and sensitive spectral features are extracted from hyperspectral images of different disaster-stricken areas with known disaster severity categories. A dataset is constructed based on sensitive bands and sensitive spectral features, as well as the corresponding disaster severity categories of the affected areas; The dataset was stratified according to the proportion of pest severity and then divided into five subsets in equal proportion; In each training process, four subsets are selected as the training set and the remaining subset is selected as the validation set. The training set is used to train the two classifiers, and the validation set is used to validate the two classifiers obtained in the current training. Repeat the above training steps until all five subsets have been validated once. At this point, the training of the two classifiers is complete.

Images