A remote sensing calculation method, system and medium for regional 30-meter forest canopy height

Abstract

The present invention discloses a remote sensing calculation method, system and medium for regional 30-meter forest canopy height, and relates to the technical field of forestry canopy height estimation. The method comprises: obtaining forest canopy height data and multi-source remote sensing factors of a target area; using a local noise algorithm to screen the forest canopy height data, and determining the prediction variables corresponding to the screened forest canopy height data according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data; using a random forest algorithm to model the screened forest canopy height data and the corresponding prediction variables to obtain a canopy height inversion model; using the canopy height inversion model to estimate the regional 30-meter forest canopy height, and drawing according to the estimated prediction data. The present invention can improve the drawing accuracy of the regional 30-meter forest canopy height.

Description

Technical Field

[0001] The invention relates to the technical field of forestry canopy height estimation, and in particular to a remote sensing calculation method, system and medium for regional 30-meter forest canopy height. Background Art

[0002] As the main body of terrestrial ecosystems, forests play an important role in the global carbon cycle and climate change. Among them, forest vegetation biomass accounts for 80% of the total terrestrial vegetation biomass. Therefore, accurately calculating the spatial distribution pattern of forest carbon storage and its dynamic changes is an important basis for understanding the carbon cycle mechanism and calculating forest carbon sinks. Estimating forest vegetation carbon storage is mainly through calculating the aboveground biomass of the forest. The aboveground biomass mainly depends on forest stand parameters such as tree species, tree height, coverage and breast diameter. Among them, tree height is a key indicator of forest aboveground biomass. Accurately mapping a large range of forest heights is an important data basis for regional forest carbon stock estimation.

[0003] The forest canopy height refers to the distance between the top of the tree canopy and the ground. At present, the estimation methods of forest canopy height include: field inventory method, model simulation method and remote sensing calculation method. First, the traditional plot measurement method is time-consuming and labor-intensive, and it will also damage the forest vegetation, and it does not have the ability to monitor large areas over a long period of time. Secondly, the model simulation method is based on field measurements and is modeled based on the measured data of canopy height, thereby reducing the field work of measuring canopy height. Once the model is established, the canopy height of the entire forest can be estimated for the same forest stand, and the estimation results have a certain degree of accuracy. However, due to the different modeling systems and applicable scopes of various types of forest stand canopy height models, it is impossible to realize the application of the model in large areas and multiple scenarios.

[0004] Remote sensing technology has achieved continuous and large-scale ground observation, with rich data sources, making up for the shortcomings of traditional field measurement methods, and has been recognized as a powerful means of mapping forest height in large areas. Among them, passive optical remote sensing data can better reflect the two-dimensional information of the horizontal direction of the forest canopy, but it cannot directly obtain the vertical structure information of the canopy and the signal is easy to saturate, especially the image is easily contaminated by clouds, and its estimation accuracy is generally low. Compared with passive remote sensing, microwave remote sensing has a certain canopy penetration ability and has certain advantages in extracting forest vertical structure parameters. Synthetic aperture radar technology has certain advantages in detecting forest height, but microwave signals are greatly affected by terrain undulations, especially in the face of high biomass problems, the signal may be seriously saturated. Compared with the above remote sensing technologies, lidar technology realizes direct measurement of forest vertical structure including canopy height through laser ranging, and has broad application prospects. Lidar mainly emits laser beams to densely sample the surface of the ground object, is not affected by clouds and fog, directly penetrates the forest canopy, and accurately obtains the three-dimensional structure information of the forest canopy. Depending on the platform they are mounted on, airborne lidar and ground-based laser scanners can provide small-scale, high-precision forest laser data, but they are expensive and have limited spatial coverage.

[0005] At present, there are the following methods to filter out unreliable photon fragments in laser data: (i) select nighttime; (ii) select strong light; (iii) exclude high cloud confidence; (iv) exclude high canopy uncertainty; (v) exclude sparsely distributed photon fragments; (vi) set a reasonable canopy height range; (vii) determine whether it is an abnormal photon based on the difference between the terrain height of the photon and the reference height. However, the photons retained by the above traditional screening methods still have large uncertainties, which is relatively limited in improving the estimation accuracy of forest canopy height. Summary of the invention

[0006] The purpose of the present invention is to provide a remote sensing calculation method, system and medium for regional 30-meter forest canopy height, which can improve the accuracy of mapping the height of regional 30-meter forest canopy.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A remote sensing calculation method for regional 30-meter forest canopy height, comprising:

[0009] Obtain forest canopy height data and multi-source remote sensing factors in the target area;

[0010] The forest canopy height data are screened using a local noise algorithm, and prediction variables corresponding to the screened forest canopy height data are determined according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data;

[0011] Using a random forest algorithm, modeling the screened forest canopy height data and corresponding prediction variables to obtain a canopy height inversion model;

[0012] The canopy height inversion model is used to estimate the 30-meter forest canopy height in the region, and the height is drawn based on the estimated prediction data.

[0013] Optionally, the forest canopy height data is screened using a local noise algorithm, and the prediction variables corresponding to the screened forest canopy height data are determined according to the multi-source remote sensing factors, specifically including:

[0014] Expand the sampling point footprints of the forest canopy height data in a two-dimensional plane, search for the nearest point for each sampling point in turn, and judge each sampling point by calculating the neighborhood average height difference, filter out abnormal points, and obtain the filtered forest canopy height data;

[0015] The cloud-covered pixels in the multi-source remote sensing factors are removed, the images are synthesized into a median image according to the image time series, and the images are unified to a spatial resolution of 30 meters, and the prediction variables corresponding to the forest canopy height data screened at the unified resolution are extracted.

[0016] Optionally, a random forest algorithm is used to model the screened forest canopy height data and corresponding prediction variables to obtain a canopy height inversion model, specifically including:

[0017] spatially matching the screened forest canopy height data and corresponding prediction variables to combine them into modeling data;

[0018] A decision tree set is constructed according to the modeling data by using a bagging algorithm or a bootstrap algorithm, a random forest model is trained by randomly selecting samples from the decision tree set, and the trained random forest model is determined as a canopy height inversion model.

[0019] Optionally, a bagging algorithm or a bootstrap algorithm is used to construct a decision tree set according to the modeling data, a random forest model is trained by randomly selecting samples from the decision tree set, and the trained random forest model is determined as the canopy height inversion model, specifically including:

[0020] Using a bagging algorithm or a bootstrap algorithm, construct N training subsets of a decision tree set according to the modeling data, and for a test set of n, sample n times with replacement to obtain a training set of n;

[0021] Randomly select features, randomly select m features from the M features of the sample, where m << M is satisfied, and then select 1 as the splitting feature of the node from these m features by calculating the Gini index;

[0022] Construct N decision trees. During the formation of each decision tree, split according to the splitting feature of the node until it cannot be split. Finally, establish N decision trees to form a random forest model;

[0023] Input the sample data not drawn in the decision tree set into the random forest model for training, adjust the number of leaf nodes of the model, and determine the trained random forest model as the canopy height inversion model.

[0024] Optionally, use the canopy height inversion model to estimate the forest canopy height of 30 meters in the area, and draw according to the estimated prediction data, specifically including:

[0025] Perform standard spatial grid processing on the modeling data, use the canopy height inversion model to estimate the canopy height of each pixel point in the target area to obtain prediction data, and draw according to the prediction data.

[0026] The present invention provides a remote sensing measurement and calculation system for the forest canopy height of 30 meters in the area, including:

[0027] An acquisition module for obtaining the forest canopy height data and multi-source remote sensing factors of the target area;

[0028] A screening module for screening the forest canopy height data by using the local noise algorithm, and determining the prediction variables corresponding to the screened forest canopy height data according to the multi-source remote sensing factors; the prediction variables include topographic factors, surface reflectance, vegetation indices, vegetation coverage, and climate data;

[0029] A model construction module for modeling the screened forest canopy height data and the corresponding prediction variables by using the random forest algorithm to obtain a canopy height inversion model;

[0030] A height drawing module for using the canopy height inversion model to estimate the forest canopy height of 30 meters in the area, and drawing according to the estimated prediction data.

[0031] The present invention provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the remote sensing measurement and calculation method for the forest canopy height of 30 meters in the area as described above.

[0032] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0033] The present invention discloses a remote sensing calculation method, system and medium for regional 30-meter forest canopy height, the method comprising obtaining forest canopy height data and multi-source remote sensing factors of a target area; using a local noise algorithm to screen the forest canopy height data, and determining the prediction variables corresponding to the screened forest canopy height data according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data; using a random forest algorithm to model the screened forest canopy height data and the corresponding prediction variables to obtain a canopy height inversion model; using the canopy height inversion model to estimate the regional 30-meter forest canopy height, and drawing according to the estimated prediction data. The present invention can improve the drawing accuracy of the height of the regional 30-meter forest canopy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] 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 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 labor.

[0035] Figure 1 It is a flow chart of the remote sensing calculation method of the regional 30-meter forest canopy height of the present invention;

[0036] Figure 2 This is a schematic diagram of neighborhood search of the local noise removal algorithm in this embodiment;

[0037] Figure 3 The comparison diagram of ICESat-2 / ATL08 canopy parameters before and after denoising in this embodiment; wherein, part (a) is a schematic diagram before denoising; part (b) is a schematic diagram after denoising;

[0038] Figure 4 is a graph showing the error variation of different random forest model parameters in this embodiment;

[0039] Figure 5 : is a scatter plot between the canopy height estimated by the two methods in this embodiment and the ICESat-2 canopy height; wherein, part (a) is a schematic diagram of the result of direct verification by the traditional method; part (b) is a schematic diagram of the result of direct verification by the method of this embodiment; part (c) is a schematic diagram of the result of 10-fold cross-validation by the traditional method; part (d) is a schematic diagram of the result of 10-fold cross-validation by the method of this embodiment;

[0040] Figure 6 The spatial distribution map of the 30-meter forest canopy height in Hunan Province predicted by the model in this embodiment;

[0041] Figure 7 It is a distribution diagram of the pixel level difference between the product in this embodiment and two publicly available products; among them, part (a) is a distribution diagram of the pixel level difference between this product and the first publicly available tree height product; part (b) is a distribution diagram of the pixel level difference between this product and the second publicly available tree height product. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in 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.

[0043] The purpose of the present invention is to provide a remote sensing calculation method, system and medium for regional 30-meter forest canopy height, which can improve the accuracy of mapping the height of regional 30-meter forest canopy.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, the present invention provides a method for remote sensing calculation of regional 30-meter forest canopy height, comprising:

[0046] Step 100: Acquire forest canopy height data and multi-source remote sensing factors of the target area;

[0047] Step 200: Filter the forest canopy height data using a local noise algorithm, and determine the prediction variables corresponding to the filtered forest canopy height data according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data; specifically include:

[0048] The sampling point footprints of the forest canopy height data are expanded in a two-dimensional plane, and the nearest point search is performed on each sampling point in turn, and each sampling point is judged by calculating the average height difference of the neighborhood, and abnormal points are screened out to obtain the screened forest canopy height data; the cloud-covered pixels in the multi-source remote sensing factor are removed, and the images are synthesized into a median image according to the image time series, and the images are unified to a spatial resolution of 30 meters, and the prediction variables corresponding to the screened forest canopy height data at the unified resolution are extracted.

[0049] Step 300: Using a random forest algorithm, modeling the screened forest canopy height data and corresponding prediction variables to obtain a canopy height inversion model; specifically including:

[0050] Spatially match the filtered forest canopy height data and the corresponding predictor variables, and combine them into modeling data; use the bagging algorithm or the bootstrap algorithm to construct a set of decision trees based on the modeling data, train a random forest model by randomly sampling samples from the set of decision trees, and determine the trained random forest model as the canopy height inversion model.

[0051] Among them, the training process of the canopy height inversion model is as follows:

[0052] Use the bagging algorithm or the bootstrap algorithm to construct N training subsets of the set of decision trees according to the modeling data. For a test set with a quantity of n, sample with replacement n times to obtain a training set with a quantity of n. The probability of each sample being drawn is calculated using the following formula:

[0053]

[0054] Randomly select features, randomly select m features from the M features of the sample, satisfying the condition m << M, and then select 1 as the splitting feature of the node from these m features by calculating the Gini index;

[0055] Construct N decision trees, and perform splitting according to the splitting features of the nodes during the formation of each decision tree until it cannot be split. Finally, establish N decision trees to form a random forest model;

[0056] Input the sample data not drawn in the set of decision trees into the random forest model for training, adjust the number of leaf nodes of the model, and determine the trained random forest model as the canopy height inversion model.

[0057] Step 400: Estimate the forest canopy height of 30 meters in the area using the canopy height inversion model, and draw according to the estimated prediction data; specifically include:

[0058] Perform standard spatial grid processing on the modeling data, use the canopy height inversion model to estimate the canopy height of each pixel point in the target area to obtain prediction data, and draw according to the prediction data.

[0059] Based on the above technical solutions, the following specific embodiments are proposed.

[0060] Step 1: Download and process spaceborne lidar canopy height data and multi-source remote sensing factors.

[0061] (1) Download the vegetation height product ATL08 of the ICESat-2 satellite laser radar located in the example area, and use a local noise algorithm to screen the quality of the canopy parameter h_canopy.

[0062] By visiting the National Snow and Ice Data Center (NSIDC), we downloaded the fifth version of ATL08 products between April 2019 and September 2021. h_canopy is used as the height inversion indicator, which represents the 98% relative height in the energy return process. In addition, in order to ensure the accuracy of the data, we also extracted a series of indicators, including longitude, latitude, best fit, terrain uncertainty, canopy uncertainty, land cover category, day and night identification, cloud confidence, and flight direction. According to the field survey results, outliers with canopy heights less than 0 meters and greater than 60 meters were removed, and footprints outside the embodiment area were excluded.

[0063] All ICESat-2 height footprints are expanded in two dimensions, and the nearest point search is performed for each point in turn ( Figure 2 ), by calculating the average height difference of the neighborhood, each point is judged to be an outlier. The specific algorithm flow is as follows:

[0064] ① Taking the current point o as the center, within the given maximum range d max Search for its k nearest points in the two-dimensional plane and calculate the average of the absolute differences in height between all the nearest points and the center point;

[0065]

[0066] ② When the mean height difference is greater than a given threshold (here set as the neighborhood standard deviation), the point is considered a noise point and is removed, otherwise it is retained as a valid point.

[0067]

[0068] Since there is a certain positive correlation between the slope and the error of canopy height extraction, the number of nearest points k is determined by the slope of the center point. This method takes 5 as the initial number of points, and divides the number of points positively according to the size of its own slope, and finally k does not exceed 25. In addition, the threshold here is set as the height standard deviation of all points in the neighborhood. When the threshold becomes smaller and smaller, the number of points retained by this method becomes smaller. Figure 3 A comparison of the original ATL08 data before and after denoising is given.

[0069] In order to verify the feasibility of the local noise removal method, quality screening was performed based on the ATL08 product attributes, which is referred to as the traditional method, including: selecting night light data; removing points with cloud content >= 2; removing points with canopy uncertainty greater than 20; removing points with an absolute difference between the photon terrain height and the reference DEM greater than 50. All points obtained by the two methods were involved in the subsequent machine learning modeling to facilitate direct comparison of accuracy.

[0070] (2) Analyze and collect predictive variables of forest canopy height, including topographic factors, surface reflectivity, vegetation index, vegetation coverage, and climate data, synthesize the data and unify them to a spatial resolution of 30 meters.

[0071] The Landsat-8 and Sentinel-2 surface reflectance products for the growing season (June to November) from 2020 to 2021 were downloaded from the Earthdata website with a spatial resolution of 30 meters. The cloud in the image was removed using the quality assessment product, and the image time series was synthesized into a median image. For Landsat-8 data, the prediction variables include surface reflectance in five spectral bands: blue, green, red, near infrared (NIR), and medium-wave infrared (SWIR); and six vegetation indices: ratio vegetation index (RVI), greenness index (GLI), normalized red-green index (RGVI), atmosphere-tolerant vegetation index (ARVI), normalized difference infrared index (NDII2), and optimized soil conditioning index (MSAVI). For Sentinel-2 data, the prediction variables include red edge vegetation index (NDVI58A, NDVI56, and NDVI47), red edge chlorophyll index (RECI), and plant sensitive reflectance index (PSRI). All images are reassembled and stored in geographic coordinates to facilitate subsequent data processing.

[0072] The elevation data of the example area were obtained using the globally rectified SRTM digital terrestrial elevation product, with a spatial resolution of approximately 0.00023 radians, and the slope of the area was calculated based on 33 grid cells.

[0073] Two vegetation products from the MODIS sensor are used: one is the vegetation continuous field (VCF) product, with a spatial resolution of about 250 meters. Cloud-covered pixels are excluded based on the provided data quality index (value > 2); the other is the leaf area index (LAI) product, with a spatial resolution of about 500 meters. Cloud-covered pixels are also removed for the same reason.

[0074] The annual mean temperature (AMT), temperature seasonality (TQ), precipitation seasonality (PQ), precipitation in the hottest quarter (PHQ), and precipitation in the coldest quarter (PCQ) were selected as climate prediction variables using the WorldClim global climate database, with a spatial resolution of approximately 1 km.

[0075] The GLC_FCS30 global land cover product was used to obtain information on nearly 30 fine-scale cover classifications for the global land area, with a spatial resolution of approximately 30 m, including different forest types.

[0076] All the ICESat-2 / ATL08 sub-satellite point data, terrain data, surface reflectance, vegetation indices, and climate data were projected onto the Universal Transverse Mercator (UTM) system and re-gridded into a grid of 0.00027 radians × 0.00027 radians (30 m resolution) for subsequent modeling processing.

[0077] Step 2: Spatially match the forest canopy height with the prediction variables to form a modeling dataset, and use the random forest algorithm to model the matched dataset, and evaluate the model accuracy through 10-fold cross-validation.

[0078] Take the forest height as the response variable, divide all the data into a training set and a validation set, and use the random forest algorithm to train the samples and construct a regression model. The process of constructing the random forest model is as follows:

[0079] Random forest is an ensemble learning algorithm with decision trees as the base learners. The bagging or bootstrap algorithm is used to construct a set of N decision trees. By randomly sampling the samples, N independent decision models are trained, and finally the voting results of the model average are output. The specific process is as follows:

[0080] (1) Construct N training subsets. For a test set with a size of n, sample n times with replacement to obtain a training set with a size of n. The probability of each sample being drawn is calculated using the following formula:

[0081]

[0082] For each subtree, when the number of samples is large enough, the probability of each element being selected is approximately 0.6. The data that is not selected as the training set is called out-of-bag data.

[0083] (2) Randomly select features. Randomly select m features from the M features of the samples, satisfying the condition m << M, and then select 1 of these m features as the splitting feature of this node by calculating the Gini index.

[0084] As for how to select the best feature to be split during the splitting process, first, traverse all the eigenvalues ​​of the corresponding feature of the data set, and calculate the error when the data set is divided into two parts according to the eigenvalue; then, select the feature and its corresponding value that minimizes the error of the two subsets after splitting as the best splitting feature value; the subset error is represented by the Gini index, which represents the probability of a randomly selected sample being misclassified. The larger the Gini index, the greater the probability of misclassification. The calculation formula is as follows:

[0085]

[0086] Among them, p k Indicates the probability that the selected sample belongs to category k;

[0087] Split the data set into left and right subtrees according to the best split feature. If the feature value is greater than the best split feature value, it belongs to the left subtree, otherwise it belongs to the right subtree. Get the left and right child nodes. If it cannot be split further, it will be used as a leaf node.

[0088] (3) Construct N decision trees. In the process of forming each decision tree, each node is split strictly according to step (2) until it cannot be split anymore. Finally, N decision trees are established to form a random forest.

[0089] (4) Based on the prediction of random forest, the out-of-bag samples are input into the regression model to obtain their eigenvalues ​​and calculate the out-of-bag error, which can be used as the accuracy of the model prediction.

[0090] During the model building process, two model hyperparameters need to be determined: one is the number of trees (base learners); the other is the number of leaf nodes in the tree containing predicted values. We used the original data for testing, set the leaf nodes from 1 to 50, and analyzed the changes in the cumulative mean square error of the model when the number of trees increases successively. In order to prevent the results from being interfered with, the test process was repeated 100 times. Figure 4 The final debugging results are shown. When the number of growing trees is greater than 100, the model generalization error tends to be stable. Here, the optimal number of leaf nodes is selected as 5.

[0091] In addition to the direct fitting results after the statistical model training is completed, the model is verified by a 10-fold cross-validation method based on the sample sequence, which can avoid potential overfitting problems in the model. The training set is randomly divided into 10 parts, one of which is selected as the validation set, and the remaining 9 are used as the training set. The cycle is repeated 10 times to traverse all the data and calculate the average performance index. We use the determination coefficient R2, root mean square error RMSE and mean absolute error MAE to evaluate the final model accuracy. The following are the formulas for the three statistical indicators.

[0092] The coefficient of determination R2 indicates the degree to which the predicted value explains the true value and can be used to measure the predictive ability of the model.

[0093]

[0094] The root mean square error (RMSE) reflects the degree of fit between the predicted value and the true value and is particularly sensitive to outliers.

[0095]

[0096] The mean absolute error (MAE) is used to indicate the gap between the predicted value and the true value, and better reflects the actual situation of the predicted value error.

[0097]

[0098] The mean absolute error (MAE) is used to indicate the gap between the predicted value and the true value, and better reflects the actual situation of the predicted value error.

[0099]

[0100] Among them, h i is the height observation value, is a highly predicted value, is the average value of height observations and n is the number of samples.

[0101] The accuracy of the model is verified using the above accuracy indicators, and the results are as follows Figure 5 As shown. Direct fitting results: R 2 =0.94, RMSE = 1.31m, MAE = 1.04m, 10-fold cross validation results based on sample sequence: R 2 =0.55, RMSE = 3.09 meters, MAE = 2.44 meters. From the verification results, it can be seen that compared with the traditional model, the model in this embodiment has better unbiasedness in estimating the original data, R 2 The accuracy of the proposed method is about doubled, with the RMSE decreasing from 6.28 meters to 3.09 meters, and the MAE decreasing from 4.63 meters to 2.44 meters. Therefore, the proposed method can effectively avoid the interference of unreliable photons, improve the model estimation capability, and can be used to invert the forest canopy height in the area.

[0102] Step 3: Use the established canopy height inversion model to make an estimate, so as to achieve accurate mapping of the 30-meter forest canopy height in the region.

[0103] Spaceborne LiDAR has discrete sampling characteristics ( Figure 3), the canopy height products provided are all distributed in strips, and continuous observation of the canopy height domain cannot be achieved. Fusion of spatially continuous remote sensing images (including multispectral data, terrain data, and other data), combined with satellite-borne LiDAR forest height products, and the use of machine learning methods to establish prediction models have become the main method for large-scale continuous mapping of vegetation heights. This embodiment integrates multi-source remote sensing data and discrete satellite-borne LiDAR canopy data, and uses the established random forest regression model to estimate the canopy height of each pixel in the forest area, and ultimately successfully achieves accurate mapping of forest canopy height at a resolution of 30 meters in Hunan Province ( Figure 6 ). In general, the forest canopy height in Hunan Province ranged from 2.53 to 22.79 meters, with an average of 15.74 meters and a standard deviation of 1.36 meters.

[0104] In order to analyze the forest canopy height dataset of this embodiment, the regional accuracy is compared with two existing canopy height data, and the pixel-level differences between the obtained dataset and other height products are analyzed. Figure 7 It can be seen that the average difference between the product of this embodiment and the two reference height products at the pixel level is 3.31 meters and 2.40 meters, respectively, the average absolute error is about 4.17 meters and 4.11 meters, respectively, and the standard deviation of the error is 3.73 meters and 4.43 meters. In addition, the pixel-level correlation with the two products is 0.59 and 0.35, respectively, which shows that this product has a relatively good match with the reference data set. In short, the accuracy of the Hunan Province forest height product generated using a combined local noise removal and random forest method is relatively reliable and has a fine resolution. Therefore, the present embodiment can achieve accurate mapping of high-resolution, spatially continuous regional forest canopy heights.

[0105] In addition, the present invention provides a regional 30-meter forest canopy height remote sensing calculation system, comprising:

[0106] The acquisition module is used to obtain forest canopy height data and multi-source remote sensing factors in the target area;

[0107] A screening module, for screening the forest canopy height data using a local noise algorithm, and determining prediction variables corresponding to the screened forest canopy height data according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data;

[0108] A model building module is used to model the screened forest canopy height data and corresponding prediction variables using a random forest algorithm to obtain a canopy height inversion model;

[0109] The height drawing module is used to estimate the 30-meter forest canopy height in the region using the canopy height inversion model, and draw according to the estimated prediction data.

[0110] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for remote sensing calculation of regional 30-meter forest canopy height.

[0111] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0112] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A remote sensing method for calculating the height of a 30-meter forest canopy in a region, characterized in that: include: Obtain forest canopy height data and multi-source remote sensing factors in the target area; The forest canopy height data are screened using a local noise algorithm, and prediction variables corresponding to the screened forest canopy height data are determined according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data; Using a random forest algorithm, modeling the screened forest canopy height data and corresponding prediction variables to obtain a canopy height inversion model; Using the canopy height inversion model to estimate the 30-meter forest canopy height in the region, and drawing based on the estimated prediction data; The local noise algorithm is used to screen the forest canopy height data, and the prediction variables corresponding to the screened forest canopy height data are determined according to the multi-source remote sensing factors, specifically including: The sampling point footprint of the forest canopy height data is expanded in a two-dimensional plane, and the nearest point search is performed for each sampling point in turn. Each sampling point is judged by calculating the neighborhood average height difference, and abnormal points are screened out to obtain the screened forest canopy height data. The specific algorithm flow is as follows: Taking the current point o as the center, within the given maximum range d max Search for k nearest points in the two-dimensional plane, and calculate the mean of the absolute height differences between all the nearest points and the center point; when the mean height difference is greater than a given threshold, it is regarded as a noise point and removed; when the mean height difference is less than or equal to a given threshold, it is retained as a valid point; the given threshold is set as the neighborhood standard deviation; The cloud-covered pixels in the multi-source remote sensing factors are removed, the images are synthesized into a median image according to the image time series, and the images are unified to a spatial resolution of 30 meters, and the prediction variables corresponding to the forest canopy height data screened at the unified resolution are extracted.

2. The remote sensing calculation method for regional 30-meter forest canopy height according to claim 1 is characterized in that: The random forest algorithm is used to model the screened forest canopy height data and the corresponding prediction variables to obtain a canopy height inversion model, which specifically includes: spatially matching the screened forest canopy height data and corresponding prediction variables to combine them into modeling data; A decision tree set is constructed according to the modeling data by using a bagging algorithm or a bootstrap algorithm, a random forest model is trained by randomly selecting samples from the decision tree set, and the trained random forest model is determined as a canopy height inversion model.

3. The remote sensing calculation method for regional 30-meter forest canopy height according to claim 2 is characterized in that: Using a bagging algorithm or a bootstrap algorithm, a decision tree set is constructed according to the modeling data, a random forest model is trained by randomly selecting samples from the decision tree set, and the trained random forest model is determined as a canopy height inversion model, specifically including: Using a bagging algorithm or a bootstrap algorithm, construct N training subsets of a decision tree set according to the modeling data, and for a test set of n, sample n times with replacement to obtain a training set of n; Randomly select features, randomly select m features from the M features of the sample, satisfying the condition m << M, and then select 1 from these m features as the splitting feature of the node by calculating the Gini index; Construct N decision trees, and split them according to the node splitting characteristics during the formation of each decision tree until they cannot be split anymore. Finally, N decision trees are established to form a random forest model. The sample data not extracted from the decision tree set is input into the random forest model for training, the number of leaf nodes of the model is adjusted, and the trained random forest model is determined as the canopy height inversion model.

4. The remote sensing calculation method for regional 30-meter forest canopy height according to claim 2 is characterized in that: The canopy height inversion model is used to estimate the 30-meter forest canopy height in the region, and the estimated prediction data is used for drawing, including: The modeling data is subjected to standard spatial gridding processing, and the canopy height of each pixel in the target area is estimated using the canopy height inversion model to obtain predicted data, and drawing is performed based on the predicted data.

5. A regional 30-meter forest canopy height remote sensing calculation system, applied to the method described in any one of claims 1 to 4, characterized in that: include: The acquisition module is used to obtain forest canopy height data and multi-source remote sensing factors in the target area; A screening module, for screening the forest canopy height data using a local noise algorithm, and determining prediction variables corresponding to the screened forest canopy height data according to the multi-source remote sensing factors; the prediction variables include terrain factors, surface reflectivity, vegetation index, vegetation coverage and climate data; A model building module is used to model the screened forest canopy height data and corresponding prediction variables using a random forest algorithm to obtain a canopy height inversion model; The height drawing module is used to estimate the 30-meter forest canopy height in the region using the canopy height inversion model, and draw according to the estimated prediction data.

6. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, the remote sensing calculation method for the regional 30-meter forest canopy height is implemented as described in any one of claims 1 to 4.

Images