Forest resource detection method based on satellite image data

Abstract

The invention relates to the technical field of forest resource detection, and provides a forest resource detection method based on satellite image data, comprising the following steps: acquiring initial satellite image data of a target forest area, and preprocessing the initial satellite image data to obtain first satellite image data; performing radiation atmospheric correction on the first satellite image data through a radiation atmospheric correction model to obtain second satellite image data; performing super-resolution reconstruction on the second satellite image data according to the reconstruction enhancement model to obtain third satellite image data; identifying the third satellite image data based on a forest ground feature extraction model to obtain forest ground feature information; and identifying the third satellite image data and the forest ground feature information through a forest resource detection model to obtain a forest resource distribution map of the target forest area. According to the method, the precision and accuracy of forest resource detection are improved, the resolution of satellite image data is enhanced, and accurate recognition of forest ground feature information is realized through the forest ground feature extraction model.

Description

Technical Field

[0001] The present invention relates to the technical field of forest resource detection, and in particular to a forest resource detection method based on satellite image data. Background Art

[0002] Forests are the main body of terrestrial ecosystems, playing a leading and supporting role. Protecting forest resources is to create a foundation for ecological civilization. Studying the dynamic changes of forest resources helps to understand the quantity, quality and structural change mechanism of forest resources. Through scientific and technological measures and reasonable regulatory policies, its structure can be more reasonable. As the largest ecosystem, forests can regulate climate, conserve water, prevent soil erosion, prevent wind and sand, prevent drought, prevent waterlogging, reduce noise, beautify the environment, and play a vital role in maintaining the biosphere ecosystem.

[0003] Traditional forest resource survey methods usually rely on ground surveys, which have disadvantages such as long time, high cost, and limited coverage, and are difficult to meet the rapidly changing forest resource monitoring needs. With the advancement of remote sensing technology, the use of satellite images for forest resource monitoring has become an effective means. Remote sensing technology can quickly obtain surface information over a large area and has the advantages of high efficiency and accuracy. Although there are some forest resource detection methods based on remote sensing, there are still problems such as insufficient processing of atmospheric influences and insufficient image resolution, which affects the accuracy and reliability of the detection results.

[0004] Existing remote sensing-based forest resource detection methods ignore the impact of water vapor and aerosols in the atmosphere, causing satellite images to appear foggy and blurred, reducing the contrast and clarity of the images. Summary of the invention

[0005] In view of this, the present invention proposes a forest resource detection method based on satellite image data. By adopting data preprocessing, radiation atmosphere correction, super-resolution reconstruction, forest object information extraction and forest resource detection, the radiation influence and atmospheric influence of satellite image data are eliminated, the resolution of satellite image data is enhanced, the resources of the target forest area are effectively identified and a forest resource distribution map is generated, the accurate identification of forest object information is achieved, the precision and accuracy of forest resource detection are improved, and the problem of foggy and blurred satellite images is solved.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides a forest resource detection method based on satellite image data, comprising the following steps:

[0007] S1, obtaining initial satellite image data of a target forest area, and preprocessing the initial satellite image data to obtain first satellite image data;

[0008] S2, performing radiation atmospheric correction on the first satellite image data by using a radiation atmospheric correction model to obtain second satellite image data;

[0009] S3, performing super-resolution reconstruction on the second satellite image data according to the reconstruction enhancement model to obtain third satellite image data;

[0010] S4, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information;

[0011] S5, identifying the third satellite image data and forest feature information through a forest resource detection model to obtain a forest resource distribution map of the target forest area.

[0012] Based on the above technical solution, preferably, step S2 includes:

[0013] The radiation atmosphere correction model includes radiation correction and atmospheric correction;

[0014] Perform radiometric correction on the first satellite image data:

[0015]

[0016] in, is the first satellite image data after radiometric correction, is the average solar intensity, g min is the minimum solar intensity, g max is the highest sunlight intensity, is the maximum solar radiation value, is the minimum solar radiation value;

[0017] Perform atmospheric correction on the first satellite image data after radiometric correction:

[0018]

[0019] in, is the second satellite image data, is the first satellite image data after radiometric correction, d dr is the distance between the Earth and the Sun, is the solar radiation flux, α s is the solar zenith angle, is the atmospheric adjustment factor, τ sq is the water vapor simulation factor, τ qr is the aerosol simulation factor, A dx is the terrain correction factor.

[0020] On the basis of the above technical solution, preferably, step S3 includes:

[0021] Decomposing the second satellite image data into a plurality of low-resolution frames, and extracting a feature vector of each low-resolution frame;

[0022] Align the feature vectors of low-resolution frames at different times through deformable convolution;

[0023] Extracting scale-adaptive feature vectors of low-resolution frames based on multiple scale factors;

[0024] The low-resolution frame is super-reconstructed according to a plurality of scale factors and scale-adaptive feature vectors to obtain third satellite image data.

[0025] On the basis of the above technical solution, preferably, the calculation formula of the third satellite image data is:

[0026]

[0027] in, is the low-resolution frame at time t, for The corresponding eigenvector, f FE (·) is the feature extraction function, for The feature vector obtained after time alignment is is the feature vector corresponding to the low-resolution frame at time i, w i is the weight at time i, p is the index position coordinate, Δp i is the index position coordinate offset at time i, N is the number of supporting frames on both sides of time t, is the scale-adaptive feature vector at scale n, f SFEM (·) is the scale-adaptive extraction function, r1 n and r2 n are the scaling factors for length and width at scale n, for The corresponding low-resolution frame, is the third satellite image data, Conv 1×1 and Conv 3×3 are 1×1 convolution operation and 3×3 convolution operation respectively, ReLU(·) is the ReLU activation function, and f GFCUM (·) is a global implicit function.

[0028] Based on the above technical solution, preferably, step S4 includes:

[0029] S41, obtaining historical forest feature information data, and dividing the historical forest feature information data into a training set and a test set;

[0030] S42, constructing an initial forest feature extraction model based on the U-NET model, iteratively training the initial forest feature extraction model through a training set, configuring a loss function and hyperparameters, until an iteration stop condition is reached, and obtaining a trained initial forest feature extraction model, wherein the iteration stop condition includes reaching a preset number of iterations and the loss function converges;

[0031] S43, evaluating the trained initial forest feature extraction model through the test set, and evaluating the model based on the mAP value. When the model evaluation meets the requirements, the forest feature extraction model is obtained. If the model evaluation does not meet the requirements, the hyperparameters are adjusted and the model is iteratively trained again.

[0032] S44, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information of the target forest area.

[0033] On the basis of the above technical solution, preferably, the calculation formula of the forest feature extraction model is:

[0034]

[0035] in, is the forest feature information of the target forest area, is the extraction score of the kth type of forest features, K is the number of types of forest features, W k is the weight of the kth type of forest features, b k is the bias term of the kth type of forest objects.

[0036] On the basis of the above technical solution, preferably, the calculation formula of the loss function is:

[0037]

[0038] Among them, Z is the model input, E(Z) is the loss function value of Z, Ω is the pixel set of Z, Y is a pixel in Ω, w(Y) is the weight of Y, and p(Y) is the predicted probability distribution of the model input.

[0039] Based on the above technical solution, preferably, step S5 includes:

[0040] The forest resource detection model includes a first regression model and a second regression model;

[0041] The first regression model is used to identify the third satellite image data, divide the target forest area into a plurality of target forest sub-areas of the same size, and obtain forest resource classification results of the plurality of target forest sub-areas;

[0042] The second regression model is used to identify forest feature information and obtain the regional content of forest resource types in multiple target forest sub-areas.

[0043] On the basis of the above technical solution, preferably, the calculation formula of the first regression model is:

[0044]

[0045] The calculation formula of the second regression model is:

[0046]

[0047] in, is the predicted value of forest resource types in the target forest sub-region, m is the number of decision trees in the first regression model, is the predicted value of the forest resource type of the target forest sub-region by the j-th decision tree in the first regression model, P is the classification result of the forest resource type of the target forest sub-region, P1, P2, P v are the target forest sub-areas of forest resource type 1, forest resource type 2, and forest resource type v, respectively. v-1 , Y v are the 1st, 2nd, v-1th, and vth forest resource category thresholds, respectively, f(X u ) is the predicted value of the u-th type of forest resources in the target forest sub-area, α d is the weight of the dth decision tree of the second regression model, h d (X u ) is the predicted value of the u-th type of forest resources in the target forest sub-region by the d-th decision tree of the second regression model, D is the number of decision trees in the second regression model, L(θ) is the loss function of the second regression model, and y u is the actual value of the u-th type of forest resources in the target forest sub-area, Q is the number of forest resource types in the target forest sub-area, and Ω(θ) is the regularization term.

[0048] On the basis of the above technical solution, preferably, step S5 further includes:

[0049] A proportional scaling map of the target forest area is drawn, and the proportional scaling map is annotated based on the forest resource classification results of multiple target forest sub-areas in the target forest area and the predicted values ​​of various types of forest resources to obtain a forest resource distribution map of the target forest area.

[0050] The forest resource detection method based on satellite image data of the present invention has the following beneficial effects compared with the prior art:

[0051] (1) By adopting data preprocessing, radiation atmosphere correction, super-resolution reconstruction, forest feature information extraction and forest resource detection, the radiation and atmospheric effects of satellite image data are eliminated, the resolution of satellite image data is enhanced, the resources of the target forest area are effectively identified and a forest resource distribution map is generated, the accurate identification of forest feature information is achieved, and the precision and accuracy of forest resource detection are improved; (2) Through deformable convolution and scale-adaptive feature vector extraction, the characteristics of forest features at different scales can be effectively captured, thereby achieving high-quality super-resolution reconstruction, ensuring that the reconstructed satellite image data is improved in detail and accuracy, and providing a more reliable data basis for forest resource detection; (3) By combining the first regression model and the second regression model, forest resources are accurately classified and regional content ratios are evaluated. The decision tree-based regression model is used to deal with the complex relationship between forest resources and satellite image data, which improves the classification accuracy and the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 A flow chart of a forest resource detection method based on satellite image data of the present invention;

[0054] Figure 2 It is a structural diagram of the reconstruction enhancement model of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See also Figure 1 This embodiment provides a forest resource detection method based on satellite image data, comprising the following steps:

[0057] S1, obtaining initial satellite image data of a target forest area, and preprocessing the initial satellite image data to obtain first satellite image data;

[0058] S2, performing radiation atmospheric correction on the first satellite image data by using a radiation atmospheric correction model to obtain second satellite image data;

[0059] S3, performing super-resolution reconstruction on the second satellite image data according to the reconstruction enhancement model to obtain third satellite image data;

[0060] S4, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information;

[0061] S5, identifying the third satellite image data and forest feature information through a forest resource detection model to obtain a forest resource distribution map of the target forest area.

[0062] Specifically, the forest resource detection method based on satellite image data of this embodiment eliminates the radiation influence and atmospheric influence of satellite image data, enhances the resolution of satellite image data, effectively identifies the resources of the target forest area and generates a forest resource distribution map, realizes the accurate identification of forest object information, and improves the precision and accuracy of forest resource detection by adopting data preprocessing, radiation atmosphere correction, super-resolution reconstruction, forest object information extraction and forest resource detection.

[0063] The preprocessing includes standardizing, denoising and filtering the initial satellite image data;

[0064] The standardized calculation formula is:

[0065]

[0066] Among them, z is the standardized initial satellite image data, Take the value of the initial satellite image data, is the mean of the initial satellite image data, and σ is the standard deviation of all data values ​​corresponding to the data source of the initial satellite image data.

[0067] Specifically, the standardization processing of this embodiment can eliminate the dimensional differences and numerical distribution differences in the initial satellite image data, making the data from different data sources or channels comparable, which is conducive to subsequent data fusion and analysis processing. This preprocessing method can improve the accuracy and stability of the radiation atmosphere correction and super-resolution reconstruction steps, and lay a good data foundation for forest object identification and resource detection.

[0068] Step S2 includes:

[0069] The radiation atmosphere correction model includes radiation correction and atmospheric correction;

[0070] Perform radiometric correction on the first satellite image data:

[0071]

[0072] in, is the first satellite image data after radiometric correction, is the average solar intensity, g min is the minimum solar intensity, g max is the highest sunlight intensity, is the maximum solar radiation value, is the minimum solar radiation value;

[0073] Perform atmospheric correction on the first satellite image data after radiometric correction:

[0074]

[0075] in, is the second satellite image data, is the first satellite image data after radiometric correction, d dr is the distance between the Earth and the Sun, is the solar radiation flux, α s is the solar zenith angle, is the atmospheric adjustment factor, τ sq is the water vapor simulation factor, τ qr is the aerosol simulation factor, A dx is the terrain correction factor.

[0076] Specifically, this embodiment eliminates the image data error caused by the change of sunlight intensity through radiation correction, so that the image data after radiation correction more truly reflects the actual situation of the target forest area.

[0077] Based on the atmospheric correction step, the interference of atmospheric components (such as water vapor and aerosol) on satellite image data is removed, ensuring the clarity and accuracy of the image data.

[0078] After radiation and atmospheric correction, the image data are more comparable at different times and conditions, facilitating time series analysis and change detection, and supporting dynamic monitoring of forest resources.

[0079] Through radiation correction and atmospheric correction, the second satellite image data generated provides high-quality basic data for super-resolution reconstruction and forest object extraction, and improves the accuracy of forest resource detection.

[0080] Step S3 includes:

[0081] Decomposing the second satellite image data into a plurality of low-resolution frames, and extracting a feature vector of each low-resolution frame;

[0082] Align the feature vectors of low-resolution frames at different times through deformable convolution;

[0083] Extracting scale-adaptive feature vectors of low-resolution frames based on multiple scale factors;

[0084] The low-resolution frame is super-reconstructed according to a plurality of scale factors and scale-adaptive feature vectors to obtain third satellite image data.

[0085] Specifically, this embodiment decomposes the second satellite image data into multiple low-resolution frames and extracts the feature vector of each frame, thereby making full use of the time series information and improving the time consistency of the reconstruction result.

[0086] By aligning feature vectors at different times through deformable convolution, the geometric distortion caused by shooting angle, time and other factors between low-resolution frames is overcome, ensuring the accurate alignment of feature information.

[0087] Extracting scale-adaptive feature vectors of low-resolution frames based on multiple scale factors can better capture detail information at different scales and improve the detail preservation ability of the reconstruction results.

[0088] The comprehensive use of multi-scale factors and scale-adaptive feature vectors for super-resolution reconstruction fully utilizes the advantages of time series information, geometric alignment and multi-scale features, and ultimately obtains high-quality third-party satellite image data to enhance the effect of forest object recognition and resource detection.

[0089] The calculation formula of the third satellite image data is:

[0090]

[0091] in, is the low-resolution frame at time t, for The corresponding eigenvector, f FE (·) is the feature extraction function, for The feature vector obtained after time alignment is is the feature vector corresponding to the low-resolution frame at time i, w i is the weight at time i, p is the index position coordinate, Δp i is the index position coordinate offset at time i, N is the number of supporting frames on both sides of time t, is the scale-adaptive feature vector at scale n, f SFEM (·) is the scale-adaptive extraction function, r1 n and r2 n are the scaling factors for length and width at scale n, for The corresponding low-resolution frame, is the third satellite image data, Conv 1×1 and Conv 3×3 are 1×1 convolution operation and 3×3 convolution operation respectively, ReLU(·) is the ReLU activation function, and f GFCUM (·) is a global implicit function.

[0092] Specifically, Figure 2 , which is a structural diagram of the reconstruction enhancement model of this embodiment. This embodiment utilizes time series information, extracts feature vectors of each low-resolution frame, and performs time alignment on them, thereby making full use of the correlation between multiple frames of data.

[0093] By introducing scale-adaptive feature vectors, we can better capture detail information at different scales, improve the detail preservation ability of the reconstruction results, and enhance the reconstruction effect of high-frequency information.

[0094] The super-resolution reconstruction layer uses 1×1 and 3×3 convolution operations, combined with ReLU activation function and global implicit function, which can make full use of multi-scale feature information. This embodiment fully considers time series information, scale adaptive features and multi-scale feature fusion, realizes effective super-resolution reconstruction of low-resolution frames, and obtains high-quality third satellite image data.

[0095] Step S4 includes:

[0096] S41, obtaining historical forest feature information data, and dividing the historical forest feature information data into a training set and a test set;

[0097] S42, constructing an initial forest feature extraction model based on the U-NET model, iteratively training the initial forest feature extraction model through a training set, configuring a loss function and hyperparameters, until an iteration stop condition is reached, and obtaining a trained initial forest feature extraction model, wherein the iteration stop condition includes reaching a preset number of iterations and the loss function converges;

[0098] S43, evaluating the trained initial forest feature extraction model through the test set, and evaluating the model based on the mAP value. When the model evaluation meets the requirements, the forest feature extraction model is obtained. If the model evaluation does not meet the requirements, the hyperparameters are adjusted and the model is iteratively trained again.

[0099] S44, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information of the target forest area.

[0100] Specifically, this embodiment utilizes historical forest feature information data and divides it into a training set and a test set, thereby providing a good data basis for model training and evaluation and ensuring the generalization ability of the model.

[0101] This embodiment uses the U-NET model as the initial forest feature extraction model, which has good semantic segmentation capabilities and can effectively identify various types of forest feature information in the third satellite image data. Through iterative training, configuring appropriate loss functions and hyperparameters, a forest feature extraction model with excellent performance is finally obtained.

[0102] The trained model is evaluated through the test set, and the mAP value is used as the evaluation indicator to objectively measure the recognition performance of the model. When the model evaluation meets the requirements, the final forest feature extraction model can be obtained. If the evaluation does not meet the requirements, the hyperparameters can be adjusted in a targeted manner to further optimize the model.

[0103] The mAP value is calculated as follows:

[0104]

[0105] Where b is the number of forest feature categories, R(c) is the recall rate of the current position, R(c+1) is the recall rate of the next position, and precision(c, c+1) is the accuracy of the current position and the next position;

[0106] Specify the following parameters,

[0107] TP: The number of samples correctly classified as positive samples; they are actually positive samples and are also classified as positive samples by the model;

[0108] FP: The number of samples that are misclassified as positive samples; they are actually negative samples, but are classified as positive samples by the model;

[0109] TN: The number of samples correctly classified as negative samples; actually negative samples, but also classified as negative samples by the model;

[0110] FN: The number of samples that are misclassified as negative samples; they are actually positive samples, but are classified as negative samples by the model;

[0111] Precision = TP / (TP+FP), recall = TP / (TP+FN);

[0112] Using the Precision and Recall of each type as the horizontal and vertical axes, finding the AP value is actually equivalent to finding the area enclosed by Precision and Recall. Since the curve area is very jittery, interpolation is used here to smooth the curve area. If the current recall value is R(c), the interpolation value is the maximum Precision value between the next position recall R(c+1). Where c represents the current position. The mAP value is the average of all types of AP values, where b represents the number of forest feature categories.

[0113] Based on the optimized forest feature extraction model, the third satellite image data can be accurately identified to obtain information on various types of forest features in the target forest area, providing data support for forest resource detection.

[0114] The calculation formula of the forest feature extraction model is:

[0115]

[0116] in, is the forest feature information of the target forest area, is the extraction score of the kth type of forest features, K is the number of types of forest features, W k is the weight of the kth type of forest features, b k is the bias term of the kth type of forest objects.

[0117] The calculation formula of the loss function is:

[0118]

[0119] Among them, Z is the model input, E(Z) is the loss function value of Z, Ω is the pixel set of Z, Y is a pixel in Ω, w(Y) is the weight of Y, and p(Y) is the predicted probability distribution of the model input.

[0120] Specifically, the forest feature extraction model of this embodiment adopts a score-based approach to feature recognition, where the score of each category is determined by a weight and a bias term, which can effectively capture the feature differences of different categories of forest features and improve recognition accuracy.

[0121] The design of the loss function takes into account the pixel set of the model input Z, as well as the weight and predicted probability distribution of each pixel Y. This pixel-level loss function can better optimize the model's ability to recognize detailed features.

[0122] The forest feature extraction model obtained through training and optimization can accurately identify various types of forest features in the third-party satellite image data and provide reliable feature information data for forest resource detection.

[0123] The forest feature extraction model fully considers the characteristic differences of forest features, adopts a score-based recognition method, and combines pixel-level loss function optimization to improve the accuracy of forest feature recognition through an improved machine learning model.

[0124] Step S5 includes:

[0125] The forest resource detection model includes a first regression model and a second regression model;

[0126] The first regression model is used to identify the third satellite image data, divide the target forest area into a plurality of target forest sub-areas of the same size, and obtain forest resource classification results of the plurality of target forest sub-areas;

[0127] The calculation formula of the first regression model is:

[0128]

[0129] The second regression model is used to identify forest feature information and obtain the regional content of forest resource types in multiple target forest sub-areas;

[0130] The calculation formula of the second regression model is:

[0131]

[0132] in, is the predicted value of forest resource types in the target forest sub-region, m is the number of decision trees in the first regression model, is the predicted value of the forest resource type of the target forest sub-region by the j-th decision tree in the first regression model, P is the classification result of the forest resource type of the target forest sub-region, P1, P2, P v are the target forest sub-areas of forest resource type 1, forest resource type 2, and forest resource type v, respectively. v-1 , Y v are the 1st, 2nd, v-1th, and vth forest resource category thresholds, respectively, f(X u ) is the predicted value of the u-th type of forest resources in the target forest sub-area, α d is the weight of the dth decision tree of the second regression model, h d (X u ) is the predicted value of the u-th type of forest resources in the target forest sub-region by the d-th decision tree of the second regression model, D is the number of decision trees in the second regression model, L(θ) is the loss function of the second regression model, and y u is the actual value of the u-th type of forest resources in the target forest sub-area, Q is the number of forest resource types in the target forest sub-area, and Ω(θ) is the regularization term.

[0133] Specifically, the first regression model of this embodiment is used to identify the third satellite image data, and the target forest area is divided into multiple target forest sub-areas of the same size, and the forest resource category classification of each sub-area is obtained. This sub-area-based analysis method can better reflect the spatial distribution characteristics of forest resources.

[0134] The second regression model is used to identify forest features and obtain the proportion of each type of forest resource in each target forest sub-area. This analysis based on feature information can more accurately assess the distribution of different forest resources.

[0135] The calculation formulas of the two regression models both adopt the form of decision trees, which can effectively capture the complex nonlinear relationship between forest resources and satellite image data and ground feature information, thereby improving the accuracy of the detection results.

[0136] The second regression model also introduced a loss function and a regularization term, which can optimize the model parameters during the training process, improve the generalization ability of the model, and enhance the predictability of the actual distribution of forest resources.

[0137] Step S5 also includes:

[0138] A proportional scaling map of the target forest area is drawn, and the proportional scaling map is annotated based on the forest resource classification results of multiple target forest sub-areas in the target forest area and the predicted values ​​of various types of forest resources to obtain a forest resource distribution map of the target forest area.

[0139] Specifically, in a specific embodiment, firstly, a scaled image of the target forest area is drawn, the geographical boundary information of the target forest area is obtained, and then a scaled image is drawn according to the actual area of ​​the target forest area;

[0140] Secondly, the forest resource classification results of the target forest sub-regions in the target forest region are marked, and the forest resource classification results of each target forest sub-region obtained by the first regression model in step S5 are marked on the proportional scale map, and different colors or legends are used to identify different types of forest resource types;

[0141] Third, the scaled map is annotated based on the predicted values ​​of each type of forest resources. The predicted content proportions of each type of forest resources in each target forest sub-area obtained by the second regression model in step S5 are annotated on the scaled map in the form of area ratios, and different colors or legends are used to identify the distribution of different types of forest resources.

[0142] Finally, the forest resource distribution map of the target forest area is obtained, and the annotation results of the above two steps are integrated on the proportional scaling map to form a forest resource distribution map that intuitively reflects the distribution of various types of forest resources in the target forest area.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A forest resource detection method based on satellite image data, characterized in that: The following steps are involved: S1, obtaining initial satellite image data of a target forest area, and preprocessing the initial satellite image data to obtain first satellite image data; S2, performing radiation atmospheric correction on the first satellite image data by using a radiation atmospheric correction model to obtain second satellite image data; S3, performing super-resolution reconstruction on the second satellite image data according to the reconstruction enhancement model to obtain third satellite image data; S4, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information; S5, identifying the third satellite image data and forest feature information through a forest resource detection model to obtain a forest resource distribution map of the target forest area.

2. The forest resource detection method based on satellite image data according to claim 1, characterized in that: Step S2 includes: The radiation atmosphere correction model includes radiation correction and atmospheric correction; Perform radiometric correction on the first satellite image data: in, is the first satellite image data after radiometric correction, is the average solar intensity, g min is the minimum solar intensity, g max is the highest sunlight intensity, is the maximum solar radiation value, is the minimum solar radiation value; Perform atmospheric correction on the first satellite image data after radiometric correction: in, is the second satellite image data, is the first satellite image data after radiometric correction, d dr is the distance between the Earth and the Sun, is the solar radiation flux, α s is the solar zenith angle, is the atmospheric adjustment factor, τ sq is the water vapor simulation factor, τ qr is the aerosol simulation factor, A dx is the terrain correction factor.

3. The forest resource detection method based on satellite image data as claimed in claim 2, characterized in that: Step S3 includes: Decomposing the second satellite image data into a plurality of low-resolution frames, and extracting a feature vector of each low-resolution frame; Align the feature vectors of low-resolution frames at different times through deformable convolution; Extracting scale-adaptive feature vectors of low-resolution frames based on multiple scale factors; The low-resolution frame is super-reconstructed according to a plurality of scale factors and scale-adaptive feature vectors to obtain third satellite image data.

4. The forest resource detection method based on satellite image data as claimed in claim 3, characterized in that: The calculation formula of the third satellite image data is: in, is the low-resolution frame at time t, for The corresponding eigenvector, f FE (·) is the feature extraction function, for The feature vector obtained after time alignment is is the feature vector corresponding to the low-resolution frame at time i, w i is the weight at time i, p is the index position coordinate, Δp i is the index position coordinate offset at time i, N is the number of supporting frames on both sides of time t, is the scale-adaptive feature vector at scale n, f SFEM (·) is the scale-adaptive extraction function, r1 n and are the scaling factors for length and width at scale n, for The corresponding low-resolution frame, is the third satellite image data, Conv 1×1 and Conv 3×3 are 1×1 convolution operation and 3×3 convolution operation respectively, ReLU(·) is the ReLU activation function, and f GFCUM (·) is a global implicit function.

5. The forest resource detection method based on satellite image data as claimed in claim 4, characterized in that: Step S4 includes: S41, obtaining historical forest feature information data, and dividing the historical forest feature information data into a training set and a test set; S42, constructing an initial forest feature extraction model based on the U-NET model, iteratively training the initial forest feature extraction model through a training set, configuring a loss function and hyperparameters, until an iteration stop condition is reached, and obtaining a trained initial forest feature extraction model, wherein the iteration stop condition includes reaching a preset number of iterations and the loss function converges; S43, evaluating the trained initial forest feature extraction model through the test set, and evaluating the model based on the mAP value. When the model evaluation meets the requirements, the forest feature extraction model is obtained. If the model evaluation does not meet the requirements, the hyperparameters are adjusted and the model is iteratively trained again. S44, identifying the third satellite image data based on the forest feature extraction model to obtain forest feature information of the target forest area.

6. The forest resource detection method based on satellite image data as claimed in claim 5, characterized in that: The calculation formula of the forest feature extraction model is: in, is the forest feature information of the target forest area, is the extraction score of the kth type of forest features, K is the number of types of forest features, W k is the weight of the kth forest feature, b k is the bias term of the kth type of forest objects.

7. The forest resource detection method based on satellite image data according to claim 6, characterized in that: The calculation formula of the loss function is: Among them, Z is the model input, E(Z) is the loss function value of Z, Ω is the pixel set of Z, Y is a pixel in Ω, w(Y) is the weight of Y, and p(Y) is the predicted probability distribution of the model input.

8. The forest resource detection method based on satellite image data as claimed in claim 7, characterized in that: Step S5 includes: The forest resource detection model includes a first regression model and a second regression model; The first regression model is used to identify the third satellite image data, divide the target forest area into a plurality of target forest sub-areas of the same size, and obtain forest resource classification results of the plurality of target forest sub-areas; The second regression model is used to identify forest feature information and obtain the regional content of forest resource types in multiple target forest sub-areas.

9. The forest resource detection method based on satellite image data as claimed in claim 8, characterized in that: The calculation formula of the first regression model is: The calculation formula of the second regression model is: in, is the predicted value of forest resource types in the target forest sub-region, m is the number of decision trees in the first regression model, is the predicted value of the forest resource type of the target forest sub-region by the j-th decision tree in the first regression model, P is the classification result of the forest resource type of the target forest sub-region, P1, P2, P v are the target forest sub-areas of forest resource type 1, forest resource type 2, and forest resource type v, respectively. v-1 , Y v are the 1st, 2nd, v-1th, and vth forest resource category thresholds, respectively, f(X u ) is the predicted value of the u-th type of forest resources in the target forest sub-area, α d is the weight of the dth decision tree of the second regression model, h d (X u ) is the predicted value of the u-th type of forest resources in the target forest sub-region by the d-th decision tree of the second regression model, D is the number of decision trees in the second regression model, L(θ) is the loss function of the second regression model, and y u is the actual value of the u-th type of forest resources in the target forest sub-area, Q is the number of forest resource types in the target forest sub-area, and Ω(θ) is the regularization term.

10. The forest resource detection method based on satellite image data according to claim 9, characterized in that: Step S5 also includes: A proportional scaling map of the target forest area is drawn, and the proportional scaling map is annotated based on the forest resource classification results of multiple target forest sub-areas in the target forest area and the predicted values ​​of various types of forest resources to obtain a forest resource distribution map of the target forest area.

Images