A forest resource prediction method based on remote sensing image analysis

Abstract

The application relates to the technical field of image processing, and discloses a forest resource prediction method based on remote sensing image analysis, which comprises the following steps: acquiring multi-source remote sensing images, performing radiation and geometric correction, and then obtaining registered multi-source remote sensing images through image registration; performing wavelet decomposition on the registered multi-source remote sensing images, performing weighted fusion on low-frequency components, performing region energy-based fusion on high-frequency components, performing wavelet inverse transformation, and generating fused remote sensing images; extracting spectral features of the fused remote sensing images to generate a spectral feature matrix; taking the spectral feature matrix as a characteristic variable, taking actual observation values of forest resources as a prediction target variable, calculating mutual information values between each characteristic variable and the prediction target variable to screen optimal spectral features, calculating Gini indexes and information gains, constructing an improved random forest model, and finally realizing forest resource prediction; and the application improves the precision of forest resource prediction results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a forest resource prediction method based on remote sensing image analysis. BACKGROUND

[0002] Forest resources are important ecological resources on the earth, and play an important role in maintaining ecological balance, providing timber and other forest products, and regulating climate. Accurate prediction of the change of forest resources is of great significance for the scientific management, sustainable utilization and ecological environment protection of forest resources.

[0003] Traditional forest resource prediction methods mainly rely on ground surveys. Although this method can obtain relatively accurate local data, it has the disadvantages of time-consuming, labor-intensive, high cost, limited coverage, etc., and is difficult to meet the needs of large-scale, rapid and dynamic forest resource prediction.

[0004] With the development of remote sensing technology, remote sensing images have been widely used in forest resource monitoring due to their wide coverage, short acquisition cycle and rich information. However, most of the current forest resource prediction methods based on remote sensing images are based on a single data source, which has the problems of incomplete data information and low prediction accuracy. At the same time, the existing methods are difficult to fully mine the effective information in the multi-source remote sensing image data, resulting in unsatisfactory prediction results. SUMMARY

[0005] In view of the above problems in the prior art, the present application provides a forest resource prediction method based on remote sensing image analysis, which solves the problem of low forest resource prediction accuracy caused by the single data source, incomplete information and poor method processing effect of the existing forest resource prediction methods.

[0006] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows:

[0007] A forest resource prediction method based on remote sensing image analysis, comprising the following steps:

[0008] S1, obtaining optical remote sensing images and radar remote sensing images of a target forest area, performing data screening, and obtaining screened multi-source remote sensing images;

[0009] S2, after the screened multi-source remote sensing images are radiometrically and geometrically corrected, image registration is performed to generate registered multi-source remote sensing images;

[0010] S3, based on the registered multi-source remote sensing image, wavelet decomposition is carried out on each remote sensing image, low-frequency components and high-frequency components are extracted, the low-frequency components are weighted and fused, and the high-frequency components are fused based on regional energy, after the fused low-frequency components and high-frequency components are generated, inverse wavelet transform is carried out, and the fused remote sensing image is generated;

[0011] S4, based on the fused remote sensing image, the spectral features of forest resources are extracted, and a spectral feature matrix is generated;

[0012] S5, the spectral feature matrix is taken as a characteristic variable, the actual observation value of the forest resource is taken as a prediction target variable, the mutual information value between each characteristic variable and the prediction target variable is calculated, the characteristic variable with the mutual information value greater than the mutual information threshold is screened, and the optimized spectral feature is generated;

[0013] S6, the optimized spectral feature is taken as input, and an improved multiple decision tree is constructed by calculating Gini index and information gain, and an improved random forest model is generated;

[0014] S7, the spectral feature of the target forest area is reacquired and input into the improved random forest model, and a forest resource prediction result is generated.

[0015] The present application has the following beneficial effects:

[0016] 1. The forest resource prediction method based on remote sensing image analysis provided by the present application comprehensively utilizes the advantages of different data sources by collecting multi-source remote sensing image data, provides more comprehensive and rich information, and improves the accuracy of forest resource prediction.

[0017] 2. After data processing of the multi-source remote sensing image, image registration and data fusion are carried out after radiation correction and geometric correction, which not only effectively eliminates the error and redundant features in the data, but also improves the quality and consistency of the data.

[0018] 3. The improved random forest model is constructed, the performance and prediction accuracy of the improved random forest model are improved by optimizing feature selection and decision tree construction, and the improved random forest model can better adapt to the complex scene of forest resource prediction. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the forest resource prediction method based on remote sensing image analysis provided by the present application. DETAILED DESCRIPTION

[0020] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, all the inventions utilizing the concept of the present application are within the scope of the present application as long as various changes are obvious within the spirit and scope of the present application defined and determined by the appended claims.

[0021] As Figure 1 shown, a forest resource prediction method based on remote sensing image analysis comprises the following steps S1-S7:

[0022] S1, obtaining optical remote sensing images and radar remote sensing images of the target forest area, performing data screening, and obtaining screened multi-source remote sensing images.

[0023] Specifically, step S1 specifically comprises S11-S13:

[0024] S11, determining the latitude and longitude range of the target forest area and the time span of data acquisition.

[0025] S12, obtaining optical remote sensing images and radar remote sensing images of the target forest area.

[0026] S13, performing data screening on the optical remote sensing images and the radar remote sensing images, eliminating remote sensing images with large-area point cloud coverage and data damage, and generating screened multi-source remote sensing images.

[0027] In this embodiment, this step is the process of obtaining multi-source remote sensing images, specifically: determining the latitude and longitude range of the target forest area, and determining the time span of data acquisition, such as once every quarter in the past 5 years; obtaining optical remote sensing images through a remote sensing satellite data platform, such as a Landsat series or a Sentinel series satellite data platform; obtaining radar remote sensing images through a radar satellite data platform; screening the obtained data to eliminate images with large-area cloud coverage, data damage, and other problems, and ensuring that the obtained multi-source remote sensing image data completely covers the target forest area in space and forms a continuous sequence in time, so as to build a high-quality, multi-dimensional data basis for subsequent forest resource prediction.

[0028] S2, after radiometric correction and geometric correction of the screened multi-source remote sensing images, image registration is performed to generate registered multi-source remote sensing images.

[0029] Specifically, step S2 specifically comprises S21-S22:

[0030] S21, performing radiometric correction and geometric correction on the screened multi-source remote sensing images to generate corrected multi-source remote sensing images.

[0031] In this embodiment, the purpose of the radiation correction and geometric correction of the screened multi-source remote sensing image is to eliminate the interference error of the multi-source remote sensing image, and to provide accurate and consistent data basis for subsequent forest resource related analysis (such as registration, feature extraction, and prediction).

[0032] In the process of radiation correction, for optical remote sensing images, atmospheric parameters during image acquisition are collected, including atmospheric aerosol concentration, water vapor content, etc., a suitable atmospheric model is selected, the digital quantization value (DN value) of the optical remote sensing image is input into the atmospheric transmission model, and the reflectivity of the surface feature is calculated to complete the radiation correction; for radar remote sensing images, according to the parameters of the radar sensor, including radar wavelength, incident angle, etc., the radar echo intensity is converted into backscattering coefficient by using the radiation calibration formula to realize the radiation correction; therefore, the radiation correction can remove the interference of non-target factors such as atmospheric scattering and sensor error on the image gray value, and restore the real radiation information of the ground object (such as forest vegetation), so as to ensure the consistency of the gray features of the same ground object in different images, and provide reliable basis for subsequent extraction of key features such as forest spectrum or structure (supporting resource prediction).

[0033] In the process of geometric correction, on the remote sensing image after radiation correction, not less than 20 ground control points are uniformly selected, and the actual geographic coordinates, i.e. latitude and longitude coordinates, of these control points are obtained; a quadratic polynomial fitting method is used to establish the mathematical relationship between the pixel coordinates of the remote sensing image and the actual geographic coordinates, and the polynomial coefficients are solved by iterative calculation; the remote sensing image is resampled (interpolated) by using the solved polynomial coefficients to obtain the geometric corrected remote sensing image; therefore, the geometric correction can eliminate the spatial position deviation of the image caused by satellite attitude, terrain fluctuation, etc., ensure the accurate alignment of remote sensing images of different sources and different time phases in spatial coordinates, and avoid the influence of position misalignment on the accurate identification of forest area range and the feature comparison across images.

[0034] S22, the corrected multi-source remote sensing image is registered to generate a registered multi-source remote sensing image, specifically:

[0035] S221, from the corrected multi-source remote sensing image, a remote sensing image with high definition and resolution is randomly selected as a reference image, and the remaining images are taken as images to be registered.

[0036] In this embodiment, the purpose of selecting a remote sensing image with high definition and resolution is to make the selected reference image have high-quality, clear ground feature and accurate geometric positioning, so as to improve the registration accuracy of the image.

[0037] S222, after the reference image and the image to be registered are grayed, the feature points of the reference image and the image to be registered are extracted by using the scale invariant feature transformation method.

[0038] S223, based on the reference image and the feature points of the image to be registered, a fast nearest neighbor search method is used to perform approximate nearest neighbor matching between the feature points of the reference image and the image to be registered, to obtain the feature points in the image to be registered that are most similar to the feature points of the reference image, and to generate a preliminary matched feature point pair.

[0039] In this embodiment, for the feature point set of the reference image, a fast nearest neighbor search index is constructed, and for each feature point of the image to be registered, the nearest neighbor and the second nearest neighbor feature points in the index of the reference image are searched, and then a Lowe ratio test is used to screen the matching pairs, that is, if the ratio of the distance of the nearest feature point to the second nearest feature point is less than a set value (0.75), the matching pair is retained, and finally a preliminary matched feature point pair is obtained.

[0040] S224, based on the preliminary matched feature point pair, a random sample consensus method is used for detection, and by constructing an affine transformation model, the false matched feature point pairs are removed, and an optimized feature point pair is generated.

[0041] In this embodiment, according to the preliminary matched feature point pair, 3 matching pairs are randomly selected from the preliminary matching pairs, and an affine transformation matrix is calculated, that is:

[0042]

[0043] wherein, is the homogeneous coordinates of the feature points of the reference image, , are the horizontal coordinates and the vertical coordinates of the feature points of the reference image in the reference image, respectively, and element 1 is the normalized form of the homogeneous coordinates, which is used to uniformly process linear transformations such as translation, rotation, scaling, etc. is the homogeneous coordinates of the feature points of the image to be registered, , are the horizontal coordinates and the vertical coordinates of the feature points of the image to be registered in the image to be registered, respectively, and 1 is also a normalized element of the homogeneous coordinates; is the affine transformation matrix, which is a 3x3 matrix, and the form is which contains the translation, rotation, scaling, shearing and other transformation information of the image to be registered to the reference image, and through the matrix, the points in the image to be registered can be mapped to the corresponding positions of the reference image, , , , , , all represent the elements of the affine transformation matrix, which together determine the specific affine transformation mode, which is calculated by the randomly selected matching pairs;

[0044] Then, all the matching pairs are transformed by the calculated affine transformation matrix, and the Euclidean distance between the transformed predicted coordinates and the actual coordinates is calculated That is,

[0045]

[0046] wherein, , are the predicted horizontal and vertical coordinates of the feature points of the image to be registered in the reference image after being transformed by the affine transformation matrix; if the Euclidean distance is less than a set threshold (3 pixels are selected in the present application), the matching pair is an inlier (correct matching point pair), otherwise, it is an outlier, i.e. an incorrect matching point pair; in addition, iterative optimization is required, i.e. repeated random sampling and the above inlier judgment process is continued, and when the maximum number of iterations (1000 times are set in the present application) is reached, the affine transformation matrix with the largest number of inliers and the corresponding inlier set are retained, and the inlier set is the optimized feature point pair.

[0047] S225, based on the optimized feature point pair, an affine transformation matrix of the image to be registered to the reference image is calculated.

[0048] S226, the spatial transformation of each pixel of the image to be registered is performed by using the affine transformation matrix, new coordinates of the image to be registered in the reference image coordinate system are obtained, and finally a registered multi-source remote sensing image is generated.

[0049] In the present embodiment, after the spatial transformation, the transformed pixel coordinates may not be integers, and therefore in actual application, the new coordinates generated after the spatial transformation need to be resampled, such as selecting a bilinear difference method to interpolate the new coordinates generated after the transformation, and finally a registered multi-source remote sensing image is obtained.

[0050] S3, based on the registered multi-source remote sensing image, wavelet decomposition is performed on each remote sensing image, low-frequency components and high-frequency components are extracted, the low-frequency components are weighted and fused, and the high-frequency components are fused based on regional energy, after the fused low-frequency components and high-frequency components are generated, inverse wavelet transformation is performed, and a fused remote sensing image is generated.

[0051] In this embodiment, multi-source remote sensing images (optical and radar images) have different information advantages. Optical images have rich spectral information, while radar images can penetrate clouds and fog and reflect terrain and landform structures well. Therefore, by wavelet decomposition, the low-frequency (reflecting the overall image overview, main energy and contour information) and high-frequency (reflecting image details, edges, textures, etc.) components of different images are fused separately. This can integrate the advantages of multi-source images, so that the fused image retains the overall clear outline and has rich detailed features, and enhances the ability to express information about forest resources (such as vegetation type, tree distribution, growth status, etc.). At the same time, the fused remote sensing image integrates the effective information of multi-source images, and has higher quality and more comprehensive information. This provides a more accurate and richer data foundation for subsequent forest resource prediction based on remote sensing images (such as vegetation cover monitoring), which helps to improve the accuracy and reliability of prediction.

[0052] Specifically, step S3 includes S31-S34:

[0053] S31. Set the wavelet function and the number of decomposition levels, and perform wavelet decomposition on each registered remote sensing image to obtain the low-frequency and high-frequency components of each remote sensing image.

[0054] In this embodiment, the wavelet function is the db4 wavelet function, and the decomposition level is 3.

[0055] S32. Based on the low-frequency components, calculate the sharpness index of the low-frequency components of each remote sensing image, determine the weight of the low-frequency components, and generate the fused low-frequency components through weighted fusion, specifically as follows:

[0056] S321. Calculate the average gradient of the low-frequency components of each remote sensing image and use it as an indicator of sharpness, i.e.:

[0057]

[0058] in, Indicates the first The average gradient of the low-frequency components in a remote sensing image measures the sharpness of the image region corresponding to that low-frequency component; a larger average gradient indicates a sharper image. , These represent the number of pixels along the horizontal and vertical directions, respectively, for the low-frequency components. Indicates the first Low-frequency components of a remote sensing image at pixel points grayscale value at that location Indicates the first Low-frequency components of a remote sensing image at pixel points grayscale value at that location Indicates the first Low-frequency components of a remote sensing image at pixel points The grayscale value at that location.

[0059] S322, sort the average gradients of the low-frequency components of each remote sensing image from large to small, the greater the average gradient of the remote sensing image, the greater the weight setting, and the sum of the weights of each remote sensing image is 1, to obtain the weight corresponding to the average gradient of the low-frequency component of each remote sensing image.

[0060] S323, the average gradient of each remote sensing image low-frequency component is weighted and averaged with its corresponding weight to obtain the fused low-frequency component, that is:

[0061]

[0062] wherein, indicates the fused low-frequency component, indicates the number of remote sensing images, indicates the weight of the i-th remote sensing image.

[0063] In this embodiment, the step generates the fused low-frequency component by performing weighted fusion based on the average gradient of the low-frequency components of the multi-source remote sensing images, thereby not only retaining the main information of the clear image, but also optimizing the fusion quality, that is, using the average gradient to reflect the definition of the image, giving greater weight to the low-frequency component of the remote sensing image with greater average gradient (i.e. clearer image), in the fusion process, the key information such as the main contour and overall structure contained in the clear image is retained more in the fused low-frequency component, ensuring that the main features of the fused image are clear and identifiable, and at the same time, the low-frequency components of multiple remote sensing images are integrated by weighted averaging, the information of multi-source images is integrated, the information loss or blur problem of a single image is avoided, the fused low-frequency component is more accurate and complete in overall appearance, laying a foundation for subsequent generation of high-quality fused remote sensing images, and further improving the accuracy of subsequent forest resource prediction based on the fused image.

[0064] S33, the high-frequency component is fused based on the energy of the region to generate a fused high-frequency component, specifically:

[0065] S331, based on the high-frequency component, the energy value of the n x n region around each pixel point of each remote sensing image is calculated, that is:

[0066]

[0067] wherein, indicates the energy value of the n x n region around the pixel point of the high-frequency component of the i-th remote sensing image, indicates the energy value of the n x n region around the pixel point of the high-frequency component of the j-th remote sensing image. ​​a high frequency coefficient of the pixel point, which reflects high frequency information of the image at the pixel point, such as edge, texture and other detailed features, represents a size of a region around the pixel point, represents a down operation.

[0068] In this embodiment, n is 3, but in practical application, n can be determined according to the image resolution and the detailed scale, for example, when the resolution of a remote sensing image is 10 m, n can be an odd number such as 3 or 5, so as to ensure the center symmetry of the field.

[0069] S332, for each pixel point, comparing energy values of all remote sensing images at the pixel point, selecting a remote sensing image with the maximum energy value, taking a high frequency coefficient of the remote sensing image as an optimal high frequency coefficient of the pixel point, and finally splicing optimal high frequency coefficients of all pixel points to generate a fused high frequency component, that is:

[0070]

[0071]

[0072] wherein, represents a coefficient of the fused high frequency component at the pixel point , represents a high frequency coefficient of a remote sensing image with the maximum energy value at the pixel point , represents a remote sensing image with the maximum energy value at the pixel point , represents an independent variable obtained when a function takes a maximum value, that is, a remote sensing image .

[0073] In this embodiment, this step generates the fused high-frequency component by fusing the high-frequency components based on the regional energy, which not only accurately retains the detailed information and enhances the image detail expression capability, but also improves the reliability of subsequent analysis, that is, the high-frequency component carries the information of details, edges, textures, and the like of the image, and when fused based on the regional energy, the high-frequency coefficient of the remote sensing image with the maximum regional energy around each pixel point is selected to accurately retain the part with the most detailed expression in the multi-source remote sensing image, so that the fused high-frequency component contains rich and clear detailed features such as the outline of trees in the forest and the subtle distribution difference of vegetation; at the same time, by integrating the part with the maximum regional energy in the high-frequency component of the multi-source remote sensing image, the fused high-frequency component is enhanced in the richness and clarity of details, so that the subsequently generated fused remote sensing image is more visually layered and the details are more prominent, which helps to more accurately identify and analyze the subtle features of forest resources such as the growth state of trees and the texture change of vegetation caused by pests and diseases; in addition, the high-quality high-frequency component fusion provides a better detailed data basis for subsequent forest resource prediction and other work based on the fused remote sensing image, so that related analysis such as forest stock volume estimation and vegetation coverage monitoring can more accurately capture key detailed information, thereby improving the reliability and accuracy of analysis and prediction results.

[0074] S34, inverse wavelet transform is performed on the fused low-frequency component and the high-frequency component to generate a fused remote sensing image.

[0075] S4, based on the fused remote sensing image, a spectral feature of the forest resource is extracted to generate a spectral feature matrix.

[0076] Specifically, step S4 specifically includes S41-S42:

[0077] S41, based on the fused remote sensing image, the reflectance values of the near-infrared band and the red band are extracted, and the normalized vegetation index value of each pixel point is calculated as the spectral feature, that is:

[0078]

[0079] wherein, the normalized vegetation index value of each pixel point, the reflectance of the near-infrared band, the reflectance of the red band.

[0080] In this embodiment, the reflectance calculation of the infrared band and the red band can effectively reflect the coverage and growth state of the vegetation, which can be used as the spectral feature of the forest resource for subsequent forest resource prediction.

[0081] S42, splicing the normalized vegetation index value of each pixel point to generate a spectral feature matrix.

[0082] In this embodiment, the characteristics of NDVI effectively reflect the vegetation coverage and growth vigor, and provide basic data for subsequent calculation of mutual information value between the characteristic variable (NDVI value) and the prediction target variable (such as forest coverage), so as to screen out spectral features more valuable for forest coverage prediction, and lay a key feature data foundation for constructing a random forest model and accurately predicting forest resources.

[0083] S5, taking the spectral feature matrix as a characteristic variable and the actual observation value of the forest resource as a prediction target variable, calculating the mutual information value between each characteristic variable and the prediction target variable, screening the characteristic variables with mutual information values greater than a mutual information threshold, and generating an optimized spectral feature.

[0084] Specifically, the actual observation value of the forest resource includes forest coverage.

[0085] Specifically, the formula for calculating the mutual information value between each characteristic variable and the prediction target variable in step S5 is:

[0086]

[0087] wherein, represents the mutual information value between the characteristic variable and the prediction target variable , which quantifies how much information the characteristic variable and the prediction target variable share, and the greater the value, the stronger the correlation between the two, so that the mutual information value is used to screen the spectral features, which can retain the most effective features for forest resource prediction, such as forest coverage prediction, represents the characteristic variable, represents the prediction target variable, represents the joint probability distribution of the characteristic variable and the prediction target variable , represents the logarithmic function, represents the marginal probability distribution of the characteristic variable , represents the marginal probability distribution of the prediction target variable .

[0088] In this embodiment, the mutual information threshold is determined by an unsupervised method, i.e. taking the median of the mutual information as the threshold, thereby filtering low correlation features.

[0089] In summary, this step, by calculating the mutual information value between spectral feature variables and actual forest resource observations and then filtering features, not only accurately selects effective features but also improves the prediction accuracy and efficiency of subsequent models, while enhancing feature interpretability. Specifically, mutual information quantifies the degree of information sharing between spectral features and actual forest resource observations, filtering out feature variables with mutual information values ​​greater than a threshold. These features are more strongly correlated with actual forest resource observations, accurately retaining the most valuable spectral information for predicting actual forest resource observations and avoiding irrelevant or weakly correlated features from interfering with subsequent model construction. Furthermore, removing redundant and irrelevant spectral features reduces the number of input features for subsequent random forest models, lowering the computational complexity of model training and improving training efficiency. Building models based on more effective features helps improve the accuracy and reliability of forest resource predictions. In addition, the features selected through mutual information have a clearer correlation with actual forest resource observations, enhancing the interpretability of prediction results in subsequent random forest models built based on these features and facilitating the analysis of the influence mechanism of spectral features on actual forest resource observations.

[0090] S6. Using the optimized spectral features as input, and by calculating the Gini index and information gain, construct multiple improved decision trees to generate an improved random forest model.

[0091] In this embodiment, this step improves the performance of individual decision trees and constructs an improved random forest model, thereby enhancing the predictive power of the random forest model. Specifically, during the node splitting process of the decision tree, the Gini index (measuring node purity) and information gain (measuring the contribution of features to classification) are comprehensively considered, and a weighted average is calculated using the analytic hierarchy process (AHP) to select the optimal splitting feature. This makes the feature selection of individual decision trees more accurate, better capturing the complex relationship between spectral features and forest resources, and improving the classification or regression performance of individual decision trees. Simultaneously, the random forest model, composed of multiple improved decision trees, retains the ensemble learning advantages of random forests (such as reducing overfitting risk and improving model stability). Furthermore, due to the improved performance of each decision tree, the entire random forest model has higher accuracy and generalization ability when predicting forest resources (such as forest cover), enabling it to more reliably complete forest resource prediction tasks. The process of generating the improved random forest model is as follows:

[0092] Specifically, step S6 includes S61-S63:

[0093] S61. Randomly sample the optimized spectral features to construct multiple decision trees.

[0094] S62. Perform feature evaluation on each decision tree and construct multiple improved decision trees, specifically as follows:

[0095] S621. For each node of each decision tree, calculate the weighted Gini index and information gain for each candidate feature, i.e.:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] in, Indicates the current node The Gini index, This represents the total number of categories of samples in the current node. Indicates the first The proportion of class samples in the current node. Indicates the current node The sample set, Representing candidate features, Representing candidate features The number of possible values, Representing candidate features Values a subset of samples Represents a subset of samples The Gini index, Represents a subset of samples The Middle The proportion of class samples, Representing candidate features The weighted Gini index, Represents the sample set of the current node The information gain of a candidate feature is considered; the larger the value, the more helpful the feature is for classification. Represents the sample set Information entropy Represents a subset of samples Information entropy This represents the logarithmic function with base 2.

[0103] In this embodiment, ,and ,because Indicates the first The proportion of each class of samples in the current node; in the sample set of a node, the proportion of all classes (total) The sum of the proportions of samples from each class must equal all samples from that node (because these samples must belong to and only belong to this class). (a certain category within a category), so from the perspective of the definition of proportion and the completeness of the set, the sum of the proportions of all categories is 1. Furthermore, the current node... The formula for calculating the Gini index is as follows: Therefore, the sample subset The formula for calculating the Gini index is: Then, the proportion of the sample size in the subset is used as the weight. Candidate features can be obtained by weighted summation of the Gini indices of all child nodes. The Gini index, also known as the weighted Gini index, is: This design allows the Gini index of child nodes with a large sample size to be more prominent in candidate features. It accounts for a larger proportion of the overall Gini index and can more accurately measure the reduction in impurity of the entire dataset after splitting the parent node with candidate features, thus providing a more reasonable basis for the subsequent decision tree to select the optimal splitting feature (selecting the optimal candidate feature as the splitting feature).

[0104] S622. Using the analytic hierarchy process (AHP), weights are assigned to the weighted Gini index and information gain, and the comprehensive value of each candidate feature is calculated, i.e.:

[0105]

[0106] in, This represents the combined value of each candidate feature. , These represent the weighted Gini index and information gain for a candidate feature, respectively. , These represent the weights of the weighted Gini index and the information gain, respectively.

[0107] In this embodiment, the weights of the weighted Gini index and information gain are determined using the analytic hierarchy process (AHP), specifically by constructing a decision matrix, calculating the largest eigenvalue and eigenvector, and performing a one-time test. Furthermore, in practical applications, cross-validation can also be used to select the weights of the weighted Gini index and information gain, thereby optimizing the performance of a single decision tree.

[0108] S623. Select the candidate feature with the best comprehensive value as the splitting feature to improve a single decision tree. Repeat S621-S623 to improve multiple decision trees, and finally generate multiple improved decision trees.

[0109] In the embodiment, in the construction process of the decision tree, a split criterion based on the combination of Gini index and information gain is used to improve the classification performance of the decision tree. For each node, the weighted Gini index and information gain of each feature are calculated, then the two are combined according to a certain weight, and the feature with the optimal comprehensive value is selected as the split feature, and finally the improvement of the single decision tree is realized.

[0110] S63, the improved multiple decision trees are combined into a random forest model to generate an improved random forest model.

[0111] In the embodiment, the optimized spectral feature set is divided into a training set and a test set according to a ratio of 7:3. The training set is used to train the improved random forest model, and by adjusting the model parameters, i.e. the number of decision trees and the maximum depth, the model achieves good performance on the training set. Then the test set is input into the trained model, and the prediction result of the forest resources can be obtained. Therefore, after obtaining the trained improved random forest model, the spectral features of any target forest area can be input into the model, and the prediction result of the forest resources can be generated.

[0112] S7, the spectral features of the target forest area are reacquired and input into the improved random forest model to generate the prediction result of the forest resources.

[0113] The forest resource prediction method based on remote sensing image analysis provided by the application can filter out invalid data by screening multi-source remote sensing images, and ensure the reliability of data input. Meanwhile, the multi-source remote sensing images screened are subjected to data processing, the sensor, atmosphere and geometric deformation errors are eliminated through radiation and geometric correction, the multi-source remote sensing image spatial coordinates are unified through image registration, and the data deviation problem is solved. Then, based on the registered multi-source remote sensing images, wavelet decomposition is carried out, the low-frequency components are weighted and fused to retain the overall information of the ground objects, the high-frequency components are fused according to the regional energy to strengthen the forest detail features, and the information-rich fused remote sensing images are generated through inverse transformation, so as to provide high-quality data support for feature extraction. Then, the spectral features related to forest prediction resources are extracted from the fused remote sensing images to construct a matrix, the features with strong correlation with the actual observation values of the forest are selected based on the mutual information values, and thus an improved random forest model is constructed. That is, by optimizing the decision tree construction rule, the improved random forest model is improved in the ability to capture the changes of forest resources, and the accuracy of forest resource prediction is finally improved. Therefore, the application gradually reduces the error and strengthens the effectiveness from the data, feature and model levels, and finally significantly improves the forest resource prediction accuracy.

[0114] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used for helping to understand the method of the present application and its core idea; meanwhile, for the ordinary skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges can be changed, and the above description should not be understood as the limitation of the present application.

[0115] Those skilled in the art will appreciate that the examples described herein are presented for purposes of aiding the reader in understanding the principles of the present application and are not intended to limit the scope of the present application to such specifically recited examples and features. Various modifications and alterations of the present application are possible and within the scope of the present application, which is not limited to the above-described embodiments and examples.

Claims

1. A forest resource prediction method based on remote sensing image analysis, characterized in that, Includes the following steps: S1. Acquire optical and radar remote sensing images of the target forest area, perform data filtering, and obtain the filtered multi-source remote sensing images; S2. After performing radiometric and geometric corrections on the selected multi-source remote sensing images, perform image registration to generate registered multi-source remote sensing images. S3. Based on the registered multi-source remote sensing images, wavelet decomposition is performed on each remote sensing image to extract low-frequency and high-frequency components. The low-frequency components are weighted and fused, while the high-frequency components are fused based on regional energy. After generating the fused low-frequency and high-frequency components, inverse wavelet transform is performed to generate the fused remote sensing image. S4. Based on the fused remote sensing images, extract the spectral features of forest resources and generate a spectral feature matrix; S5. Using the spectral feature matrix as the feature variable and the actual observed value of forest resources as the prediction target variable, calculate the mutual information value between each feature variable and the prediction target variable, filter the feature variables with mutual information values ​​greater than the mutual information threshold, and generate the optimized spectral features. S6. Using the optimized spectral features as input, and by calculating the Gini index and information gain, construct improved decision trees to generate an improved random forest model. S7. Reacquire the spectral characteristics of the target forest area and input them into the improved random forest model to generate forest resource prediction results.

2. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Step S1 specifically includes: S11. Determine the latitude and longitude range of the target forest area and the time span for data acquisition; S12. Acquire optical and radar remote sensing images of the target forest area; S13. Perform data screening on optical and radar remote sensing images, remove remote sensing images with large areas of point cloud coverage and data corruption, and generate the screened multi-source remote sensing images.

3. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Step S2 specifically includes: S21. Perform radiometric and geometric corrections on the selected multi-source remote sensing images to generate corrected multi-source remote sensing images. S22. Perform image registration on the corrected multi-source remote sensing images to generate registered multi-source remote sensing images, specifically as follows: S221. From the corrected multi-source remote sensing images, randomly select one remote sensing image with high clarity and resolution, use it as the reference image, and use the remaining images as the images to be registered. S222. After converting the reference image and the image to be registered to grayscale, the feature points of the reference image and the image to be registered are extracted using the scale-invariant feature transformation method. S223. Based on the feature points of the reference image and the image to be registered, a fast nearest neighbor search method is used to perform approximate nearest neighbor matching between the feature points of the reference image and the image to be registered, to obtain the feature points in the image to be registered that are most similar to the feature points of the reference image, and to generate a preliminary matching pair of feature points. S224. Based on the preliminary matching feature point pairs, the random sample consistency method is used for detection. By constructing an affine transformation model, incorrectly matched feature point pairs are eliminated, and optimized feature point pairs are generated. S225. Based on the optimized feature point pairs, calculate the affine transformation matrix from the image to be registered to the reference image; S226. Use the affine transformation matrix to perform spatial transformation on each pixel of the image to be registered, obtain the new coordinates of the image to be registered in the coordinate system of the reference image, and finally generate the registered multi-source remote sensing image.

4. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Step S3 specifically includes: S31. Set the wavelet function and the number of decomposition levels, and perform wavelet decomposition on each registered remote sensing image to obtain the low-frequency and high-frequency components of each remote sensing image. S32. Based on the low-frequency components, calculate the sharpness index of the low-frequency components of each remote sensing image, determine the weight of the low-frequency components, and generate the fused low-frequency components through weighted fusion, specifically as follows: S321. Calculate the average gradient of the low-frequency components of each remote sensing image and use it as an indicator of sharpness, i.e.: in, Indicates the first The average gradient of the low-frequency components of the remote sensing image. , These represent the number of pixels along the horizontal and vertical directions, respectively, for the low-frequency components. Indicates the first Low-frequency components of a remote sensing image at pixel points grayscale value at that location Indicates the first Low-frequency components of a remote sensing image at pixel points grayscale value at that location Indicates the first Low-frequency components of a remote sensing image at pixel points The grayscale value at that location; S322. Sort the average gradients of the low-frequency components of each remote sensing image from largest to smallest. Set the weight of the remote sensing image with the larger average gradient, and the sum of the weights of each remote sensing image is 1. This gives the weights corresponding to the average gradients of the low-frequency components of each remote sensing image. S323. The average gradient of each low-frequency component in the remote sensing image is weighted and averaged with its corresponding weight to obtain the fused low-frequency component, i.e.: in, This represents the low-frequency component after fusion. Indicates the number of remote sensing images. Indicates the first Weights of remote sensing images; S33. Perform region-based energy-based fusion on the high-frequency components to generate fused high-frequency components, specifically as follows: S331. Based on high-frequency components, calculate the energy value of the n×n region surrounding each pixel in each remote sensing image, i.e.: in, Indicates the first High-frequency components of a remote sensing image at pixel points The energy value of the surrounding n×n region. Indicates the first High-frequency components of a remote sensing image at pixel points High-frequency coefficients at the location, Indicates the size of the area surrounding a pixel. This indicates a round-down operation; S332. For each pixel, compare the energy values ​​of all remote sensing images at that pixel, select the remote sensing image with the highest energy value, and use the high-frequency coefficients of that remote sensing image as the optimal high-frequency coefficients for that pixel. Finally, stitch together the optimal high-frequency coefficients of all pixels to generate the fused high-frequency component, i.e.: in, This indicates the high-frequency components after fusion at the pixel level. The coefficient at the location, Indicates at pixel point The high-frequency coefficients of the remote sensing image at the point of maximum energy. Indicates at pixel point Remote sensing image at the point of maximum energy value. This represents the independent variable obtained when the function reaches its maximum value, i.e., the remote sensing image. ; S34. Perform inverse wavelet transform on the fused low-frequency and high-frequency components to generate the fused remote sensing image.

5. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Step S4 specifically includes: S41. Based on the fused remote sensing image, extract the reflectance values ​​of the near-infrared and red bands, calculate the normalized vegetation index value of each pixel, and use it as a spectral feature, i.e.: in, This represents the normalized vegetation index value for each pixel. Represents the reflectivity in the near-infrared band. Indicates the reflectivity in the red light band; S42. The normalized vegetation index values ​​of each pixel are concatenated to generate a spectral feature matrix.

6. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Actual observed values ​​for forest resources include forest cover.

7. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, The formula for calculating the mutual information value between each feature variable and the target variable in step S5 is as follows: in, Representing characteristic variables With the target variable for prediction Mutual information value between them Representing characteristic variables, Indicates the target variable to be predicted. Representing characteristic variables With the target variable for prediction The joint probability distribution, Represents the logarithmic function. Representing characteristic variables Marginal probability distribution, Represents the target variable for prediction The marginal probability distribution.

8. The forest resource prediction method based on remote sensing image analysis according to claim 1, characterized in that, Step S6 specifically includes: S61. Randomly sample the optimized spectral features to construct multiple decision trees; S62. Perform feature evaluation on each decision tree and construct multiple improved decision trees, specifically as follows: S621. For each node of each decision tree, calculate the weighted Gini index and information gain for each candidate feature; S622. The weighted Gini index and information gain are assigned weights using the analytic hierarchy process (AHP) and the comprehensive value of each candidate feature is calculated. S623. Select the candidate feature with the best comprehensive value as the splitting feature to improve a single decision tree. Repeat S621-S623 to improve multiple decision trees, and finally generate multiple improved decision trees. S63. Combine the improved decision trees into a random forest model to generate an improved random forest model.

9. The forest resource prediction method based on remote sensing image analysis according to claim 8, characterized in that, The formula for calculating the weighted Gini index and information gain for each candidate feature is as follows: in, Indicates the current node The Gini index, This indicates the total number of categories of samples in the current node. Indicates the first The proportion of class samples in the current node. Indicates the current node The sample set, Representing candidate features, Representing candidate features The number of possible values, Representing candidate features Values a subset of samples Represents a subset of samples The Gini index, Represents a subset of samples The Middle The proportion of class samples, Representing candidate features The weighted Gini index, Represents the sample set of the current node Information gain of candidate features Represents the sample set Information entropy Represents a subset of samples Information entropy This represents the logarithmic function with base 2.

10. The forest resource prediction method based on remote sensing image analysis according to claim 8, characterized in that, The formula for calculating the comprehensive value of each candidate feature is: in, This represents the combined value of each candidate feature. , These represent the weighted Gini index and information gain for a candidate feature, respectively. , These represent the weights of the weighted Gini index and the information gain, respectively.

Images