Artificial forest remote sensing extraction method based on satellite-borne laser waveform vertical structure complexity

Abstract

The invention discloses a man-made forest remote sensing extraction method based on satellite-borne laser waveform vertical structure complexity. Comprising the following steps: step 1, acquiring L1B geographic positioning waveform, L2A light spot scale global surface height and vegetation height and L2B light spot scale global crown coverage and vertical profile index product data of a satellite-borne full-waveform laser radar GEDI in a forest area; 2, carrying out the preprocessing of the obtained data, and obtaining a high-quality satellite-borne laser radar light spot in the region; 3, carrying out feature extraction, calculating the information entropy of the echo waveform of a single light spot and the waveform similarity of adjacent light spots from L1B geographic positioning waveform data, and extracting the canopy coverage and the leaf height diversity index from L2B light spot scale global crown coverage and vertical profile indexes; constructing a linear model of the height percentile in the L2A light spot scale global surface height and vegetation height data, and calculating the slope of the linear model as a feature; 4, inputting the five features, and performing classified extraction of artificial forests and natural forests by adopting a random forest model; and 5, evaluating the extraction precision and feature importance of the artificial forest. According to the method, a plurality of vertical structure characteristics of the forest are obtained through waveform information of the satellite-borne laser radar, the stand of the artificial forest is extracted in combination with the random forest model, and the artificial forest and the natural forest can be distinguished by fully utilizing the vertical structure characteristics of the satellite-borne laser waveform of the forest area.

Description

Technical Field

[0001] The invention relates to the technical field of artificial forest extraction, and in particular to an artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveforms. Background Art

[0002] Planted forests and natural forests have different functions and values ​​in terms of carbon storage, biodiversity and ecosystem services. Accurately identifying the extent of planted forests in forest ecosystems not only helps to understand their ecological benefits, but also provides a scientific basis for the formulation of management and protection measures. The National Forest Inventory is currently the only source of data on the distribution of planted forests across the country. However, forest inventory data are collected through field surveys and depend to a large extent on the accessibility, remoteness and ownership of forests. In addition, the National Forest Inventory covers a long time span and cannot provide an accurate and up-to-date spatial distribution of China's planted forests. In recent years, advances in remote sensing technology have provided new possibilities for planted forest extraction. By efficiently extracting and analyzing large-area forest data, the distribution of large-scale planted forests can be updated in a timely manner. The current methods for extracting planted forests mainly include extracting spectral and time series features of satellite images for classification, as well as feature recognition combining multi-source remote sensing data.

[0003] The paper "Spracklen, B., & Spracklen, DV (2021). Synergistic use of Sentinel-1 and Sentinel-2 to map natural forest and Acacia plantation and stand ages in North-Central Vietnam. Remote Sensing, 13 (2), 185. https: / / doi.org / 10.3390 / rs13020185", based on the texture features of Sentinel-1 and the spectral features of Sentinel-2, a random forest model was used to distinguish between plantations and natural forests in North-Central Vietnam. In forest areas with a forest age of more than 2 years, the classification accuracy is significantly reduced. The traditional spectral and texture features are greatly affected by the forest age in the process of extracting plantations.

[0004] The paper "Cheng, K., Su, Y., Guan, H., Tao, S., Ren, Y., Hu, T., Ma, K., Tang, Y., & Guo, Q. (2023). Mapping China's planted forests using high resolution imagery and massive amounts of crowdsourced samples. ISPRS Journal of Photogrammetry and Remote Sensing, 196, 356-371. https: / / doi.org / 10.1016 / j.isprsjprs.2023.01.005" constructed a mapping framework based on Google Earth Engine based on the spectral, temporal, structural, textural and topographic characteristics of Landsat and Sentinel-1 time series images, digital elevation models and China's forest canopy height data, as well as crowdsourced data, and produced the first 30-meter resolution distribution map of China's planted and natural forests. Among them, the temporal feature is of great importance in the process of extracting artificial forests, but multi-temporal passive optical remote sensing data are difficult to obtain in cloudy and rainy areas.

[0005] A method for spatial pattern recognition of artificial forests based on temporal classification and spatial analysis disclosed in Chinese patent literature, with announcement number CN112380994A, includes: collecting available Landsat data and PALSAR data; classifying land cover types by random gradient boosting based on PALSAR polarization and texture; extracting the cumulative maximum value of vegetation index NDVI of different land cover categories from multi-temporal Landsat images for phenological feature recognition; determining the threshold and criteria for identifying forest pixels and non-forest pixels; generating forest and non-forest products based on PALSAR, Landsat, thresholds and criteria; and finally generating the spatial distribution of artificial forests by combining knowledge discrimination and spatial analysis methods. In this method, a fixed NDVI threshold of 0.72 is used to identify forest and non-forest pixels. However, factors such as climate conditions, vegetation phenological characteristics, and soil background in different regions will affect the performance of NDVI values. A single threshold may not be able to adapt to diverse environmental characteristics, especially in areas with large climate differences between the north and the south. The threshold of 0.72 may be too high or too low in some places. Moreover, only the polarization information and texture features of PALSAR data are used for classification, and more features in the radar images are not fully explored. The surface penetration ability and vertical structure reflection ability of radar images have not been fully utilized.

[0006] A "remote sensing identification method for evergreen plantations and remote sensing monitoring method for the growth of evergreen plantations" disclosed in Chinese patent documents. Its announcement number is CN111310639A. The method includes: determining the evergreen forest coverage index at each pixel position in the first-type time series remote sensing image based on the first-type time series remote sensing image of the working area in each year within the target time range; then determining the evergreen forest planting area index at each pixel position within the target time range based on the evergreen forest coverage index at each pixel position; finally, determining the historical evergreen forest distribution area of ​​the working area within the target time range based on the evergreen forest planting area index at each pixel position. This method is relatively limited in application and cannot meet the needs of different forest types and large-scale mapping, and its versatility is poor. Summary of the invention

[0007] Purpose of the invention: The present invention provides a remote sensing extraction method for artificial forests based on the vertical structure complexity of satellite-borne laser waveforms. It utilizes the ability of satellite-borne laser full waveform data to reflect vertical structural characteristics, effectively distinguishes artificial forests from natural forests, extracts artificial forest stands, and provides a scientific basis for the formulation of management and protection measures.

[0008] Technical solution: The present invention discloses a method for extracting artificial forests by remote sensing based on the vertical structure complexity of satellite-borne laser waveforms, comprising the following steps:

[0009] Step 1: Obtain the L1B geolocation waveform, L2A spot-scale global surface height and vegetation height, and L2B spot-scale global canopy cover and vertical profile indicator product data of the spaceborne full-waveform lidar GEDI in the forest area;

[0010] Step 2: Crop the acquired data, screen the quality, and remove the ground point preprocessing operations to obtain high-quality satellite-borne laser radar spots in the area;

[0011] Step 3: Perform feature extraction, calculate the information entropy of a single spot echo waveform and the waveform similarity of adjacent spots from the L1B geolocation waveform data, extract the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile indicators, construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature;

[0012] Step 4: Input the above five features and use the random forest model to classify and extract artificial forests and natural forests;

[0013] Step 5: Evaluate the extraction accuracy and feature importance of artificial forests.

[0014] Furthermore, in step 1, the artificial forest and natural forest sample areas were selected by combining high-definition images, existing distribution map products of artificial forests and natural forests, and text records. The L1B geolocation waveform data, L2A spot-scale global surface height and vegetation height data, and L2B spot-scale global canopy cover and vertical profile indicator data of the space-borne lidar GEDI that passed through the forest area were downloaded from the NASA Earth Observation Data Center. The data were published in the .hdf5 format, and the echo waveforms and other information were stored in gediSDS.

[0015] Furthermore, in step 2, the pre-processing operations such as clipping, quality screening, and removing ground points on the acquired data specifically include the following steps:

[0016] Step 21, read latitude_bin0 and longitude_bin0 in gediSDS, and filter the GEDI light spots within the forest range;

[0017] Step 22, read stale_return_flag in gediSDS, degrade, filter out GEDI spots that meet the quality standards, read beamNames, and only keep the GEDI full power intensity tracks of BEAM0101, BEAM0110, BEAM1000, and BEAM1011;

[0018] Step 23: Read rxwaveform in gediSDS, obtain the DN value of the lidar echo, normalize and smooth the DN value, detect the peak value, delete the light spots with only a single peak value, that is, the ground point light spots, and only keep the light spots with trees in the area.

[0019] Furthermore, in step 3, feature extraction is performed, the information entropy of a single spot echo waveform and the waveform similarity of adjacent spots are calculated from the L1B geolocation waveform data, the canopy coverage and leaf height diversity index are extracted from the L2B spot scale global canopy coverage and vertical profile index data, a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data is constructed, and the slope of the linear model is calculated as a feature. The specific steps include the following:

[0020] Step 31, calculate the information entropy of a single spot echo waveform to measure the complexity of the vertical structure of the forest; discretize the DN value into 250 intervals; use a histogram to count the probability of each DN value; perform probability distribution calculation and entropy value calculation, and the information entropy calculation formula is:

[0021]

[0022] Where i is the index variable, H(X) is the information entropy of the DN value, and x iis the discretized DN value, p(x i ) is the DN value taken to x i The probability of

[0023] Step 32, calculate the similarity between the waveform curves of a single light spot and the spatially adjacent light spots, which is used to measure the similarity of the vertical structure between the light spots. Considering the first law of geography, only the similarity of the light spots with a distance of no more than 120 meters is calculated; perform linear interpolation on the waveform curves so that the curves have the same number of DN values; after the curve lengths are consistent, calculate the similarity to quantify the correlation between the two curves in shape. The similarity calculation formula is:

[0024]

[0025] where x i and i are the corresponding point values ​​in the two curves, and is the mean of the two curves, and n is the number of DN values;

[0026] Step 33, extracting the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile index data, where the canopy coverage represents the percentage of the area covered by the vertical projection of the canopy part to the spot area, and the leaf height diversity represents the leaf density and height distribution inside the canopy;

[0027] Step 34: construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature. The slope feature can reflect the density and tree height of the understory vegetation; extract 101 height indicators: rh0-rh100 from the L2A spot scale global surface height and vegetation height data, perform linear regression on the rh value and the square root of the height percentile (0-100), and use the slope of the linear model as a feature. The calculation formula of the linear model is:

[0028]

[0029] Among them, rh x is the height index and x is the height percentile.

[0030] Furthermore, in step 4, 5 features were input, and the random forest model was used to classify and extract artificial forests and natural forests. The data set was divided into training set, validation set and test set, which accounted for 70%, 20% and 10% of the total data respectively. The random forest model was used for classification training, 150 decision trees were set to improve the robustness of classification, and random seeds were set to ensure the repeatability of the results.

[0031] Furthermore, in step 5, evaluating the extraction accuracy and feature importance of artificial forests specifically includes the following steps:

[0032] Step 51: After the model training is completed, the performance of the model is evaluated using the test set, and the confusion matrix, classification accuracy, and classification report are calculated to evaluate the model's extraction accuracy for plantations;

[0033] Step 52: Evaluate feature importance based on Gini Impurity to measure the contribution of each feature in the artificial forest extraction process.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention uses an active satellite-borne lidar remote sensing method, which is not affected by cloudy and rainy weather and supports the extraction of vertical structural features in any forest area covered by a satellite-borne lidar; (2) The present invention only uses the vertical structural features obtained from waveform data to extract artificial forests, paying full attention to the differences in vertical structure between artificial forests and natural forests, and achieving better extraction effects; (3) The present invention supports large-scale remote sensing monitoring of artificial forests and can provide policy support for the construction and management of artificial forests by relevant departments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0036] Figure 2 Violin plots to verify feature availability for the present invention. DETAILED DESCRIPTION

[0037] like Figure 1 As shown, a remote sensing extraction method for artificial forests based on the vertical structure complexity of satellite-borne laser waveforms includes the following steps:

[0038] Step 1: Obtain the L1B geolocation waveform, L2A spot-scale global surface height and vegetation height, and L2B spot-scale global canopy cover and vertical profile indicator product data of the spaceborne full-waveform lidar GEDI in the forest area.

[0039] Combining high-definition images, existing artificial forest and natural forest distribution map products and text records, artificial forest and natural forest sample areas were selected, and L1B geolocation waveform data, L2A spot scale global surface height and vegetation height data, and L2B spot scale global canopy cover and vertical profile indicator data of the space-borne lidar GEDI passing through the forest area were downloaded from the NASA Earth Observation Data Center. The data were published in .hdf5 format, and the echo waveforms and other information were stored in gediSDS.

[0040] The present invention firstly selects four northern forest regions, and the regional information is shown in Table 1.

[0041] Table 1 Regional information

[0042]

[0043]

[0044] Step 2: Perform preprocessing operations such as cropping, quality screening, and removing ground points on the acquired data to obtain high-quality satellite-borne laser radar spots in the area; specifically, the following steps are included:

[0045] Step 2.1, read latitude_bin0 and longitude_bin0 in gediSDS, and filter the GEDI light spots within the forest range;

[0046] Step 2.2, read stale_return_flag in gediSDS, degrade, filter out GEDI spots that meet the quality standards, read beamNames, and only keep the GEDI full power intensity tracks of BEAM0101, BEAM0110, BEAM1000, and BEAM1011;

[0047] Step 2.3, read rxwaveform in gediSDS, obtain the DN value of the lidar echo, normalize and smooth the DN value, detect the peak, delete the light spots with only a single peak, that is, the ground point light spots, and only keep the light spots with trees in the area.

[0048] Step 3: Perform feature extraction, calculate the information entropy of a single spot echo waveform and the waveform similarity of adjacent spots from the L1B geolocation waveform data, extract the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile indicators, construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature; specifically, the following steps are included:

[0049] Step 3.1, calculate the information entropy of a single spot echo waveform to measure the complexity of the vertical structure of the forest. The specific calculation is: discretize the DN value into 250 intervals; use a histogram to count the probability of each DN value; perform probability distribution calculation and entropy value calculation. The information entropy calculation formula is:

[0050]

[0051] Where i is the index variable, H(X) is the information entropy of the DN value, and x i is the discretized DN value, p(x i ) is the DN value taken to x i The probability of

[0052] Step 3.2, calculate the similarity of the waveform curve of a single light spot and the spatially adjacent light spots, which is used to measure the similarity of the vertical structure between the light spots. Considering the first law of geography, only the similarity of the light spots with a distance of no more than 120 meters is calculated. The specific calculation is: linearly interpolate the waveform curves so that the curves have the same number of DN values; after the curve lengths are consistent, calculate the similarity to quantify the correlation between the two curves in shape. The similarity calculation formula is:

[0053]

[0054] where x i and i are the corresponding point values ​​in the two curves, and is the mean of the two curves, and n is the number of DN values;

[0055] Step 3.3, extract the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile index data. The canopy coverage indicates the percentage of the area covered by the vertical projection of the canopy part to the spot area, and the leaf height diversity indicates the leaf density and height distribution inside the canopy.

[0056] Step 3.4, construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature. The slope feature can reflect the density and tree height of the understory vegetation. The specific calculation is: extract 101 height indicators: rh0-rh100 from the L2A spot scale global surface height and vegetation height data, perform linear regression on the rh value and the square root of the height percentile (0-100), and use the slope of the linear model as a feature. The calculation formula of the linear model is:

[0057]

[0058] Among them, rh x is the height index and x is the height percentile.

[0059] In order to verify that the extracted features are different between artificial forests and natural forests and can be used for classification, the present invention uses violin plots to compare the differences of the five features in the two forest stands. The results are as follows: Figure 2 shown.

[0060] Step 4. Input the above five features and use the random forest model to classify and extract artificial forests and natural forests; divide the data set into training set, validation set and test set, which occupy 70%, 20% and 10% of the total data respectively, use the random forest model for classification training, set 150 decision trees to improve the robustness of classification, and set random seeds to ensure the repeatability of the results.

[0061] Step 5: Evaluate the extraction accuracy and feature importance of artificial forests; specifically include the following steps:

[0062] Step 5.1: After the model training is completed, the performance of the model is evaluated using the test set, and the confusion matrix, classification accuracy, and classification report are calculated to evaluate the model's extraction accuracy for plantations.

[0063] Step 5.2: Evaluate feature importance based on Gini impurity to measure the contribution of each feature in the plantation extraction process.

[0064] In order to evaluate the accuracy of plantation extraction, the present invention calculates the accuracy, producer precision, user precision and F1 score through the confusion matrix, and the results are shown in Table 2. The results show that the method proposed in the present invention has achieved good results in plantation extraction.

[0065] Table 2 Accuracy, producer precision, user precision and F1 score

[0066]

[0067] In order to evaluate the contribution of each feature in the random forest to the extraction of artificial forests, the present invention evaluates the importance of features based on Gini impurity, and the results are shown in Table 3.

[0068] Table 3 Evaluation of feature importance based on Gini impurity

[0069]

Claims

1. A remote sensing extraction method for artificial forests based on the vertical structure complexity of satellite-borne laser waveforms, characterized in that: The steps include: Step 1: Obtain the L1B geolocation waveform, L2A spot-scale global surface height and vegetation height, and L2B spot-scale global canopy cover and vertical profile indicator product data of the spaceborne full-waveform lidar GEDI in the forest area; Step 2: preprocess the acquired data to obtain high-quality satellite-borne laser radar spots in the area; Step 3: Perform feature extraction, calculate the information entropy of a single spot echo waveform and the waveform similarity of adjacent spots from the L1B geolocation waveform data, extract the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile indicators, construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature; Step 4: Input the above five features and use the random forest model to classify and extract artificial forests and natural forests; Step 5: Evaluate the extraction accuracy and feature importance of artificial forests.

2. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 1, characterized in that: In step 1, the artificial forest and natural forest sample areas were selected by combining high-definition images, existing artificial forest and natural forest distribution map products and text records. The L1B geolocation waveform data, L2A spot scale global surface height and vegetation height data, and L2B spot scale global canopy cover and vertical profile index data of the space-borne laser radar GEDI passing through the forest area were downloaded from the NASA Earth Observation Data Center. The data was published in .hdf5 format, and the echo waveform and other information were stored in gediSDS.

3. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 1, characterized in that: In step 2, the acquired data is cropped, quality screened, and ground point preprocessing operations are performed.

4. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 3 is characterized in that: The preprocessing operations of clipping, quality screening and removing ground points of the acquired data include the following steps: Step 21, read latitude_bin0 and longitude_bin0 in gediSDS, and filter the GEDI light spots within the forest range; Step 22, read stale_return_flag in gediSDS, degrade, filter out GEDI spots that meet the quality standards, read beamNames, and only keep the GEDI full power intensity tracks of BEAM0101, BEAM0110, BEAM1000, and BEAM1011; Step 23: Read rxwaveform in gediSDS, obtain the DN value of the lidar echo, normalize and smooth the DN value, detect the peak value, delete the light spots with only a single peak value, that is, the ground point light spots, and only keep the light spots with trees in the area.

5. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 1, characterized in that: In step 3, feature extraction is performed, the information entropy of a single spot echo waveform and the waveform similarity of adjacent spots are calculated from the L1B geolocation waveform data, the canopy coverage and leaf height diversity index are extracted from the L2B spot scale global canopy coverage and vertical profile index data, a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data is constructed, and the slope of the linear model is calculated as a feature. The specific steps include the following: Step 31, calculating the information entropy of a single light spot echo waveform to measure the complexity of the forest vertical structure; Step 32, calculating the similarity of the waveform curves of a single light spot and the spatially adjacent light spots, which is used to measure the similarity of the vertical structures between the light spots. Considering the first law of geography, only the similarity of the light spots with a distance of no more than 120 meters is calculated; Step 33, extracting the canopy coverage and leaf height diversity index from the L2B spot scale global canopy coverage and vertical profile index data, where the canopy coverage represents the percentage of the area covered by the vertical projection of the canopy part to the spot area, and the leaf height diversity represents the leaf density and height distribution inside the canopy; Step 34, construct a linear model of the height percentiles in the L2A spot scale global surface height and vegetation height data, and calculate the slope of the linear model as a feature. The slope feature can reflect the density and tree height of the understory vegetation; extract 101 height indicators: rh0-rh100 from the L2A spot scale global surface height and vegetation height data, perform linear regression on the rh value and the square root of the height percentile (0-100), and use the slope of the linear model as a feature.

6. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 5, characterized in that: In step 31, the DN value is discretized into 250 intervals; the probability of occurrence of each DN value is counted using a histogram; Probability distribution calculation and entropy value calculation are performed, and the information entropy calculation formula is: Where i is the index variable, H(X) is the information entropy of the DN value, and x i is the discretized DN value, p(x i ) is the DN value taken to x i probability.

7. The method for extracting artificial forests based on the vertical structure complexity of satellite-borne laser waveforms according to claim 5, characterized in that: In step 32, linear interpolation is performed on the waveform curves so that the curves have the same number of DN values; after the curve lengths are consistent, the similarity is calculated to quantify the correlation between the two curves in shape. The similarity calculation formula is: where x i and i are the corresponding point values ​​in the two curves, and is the mean of the two curves, and n is the number of DN values.

8. The method for extracting artificial forests by remote sensing based on the vertical structure complexity of satellite-borne laser waveforms according to claim 5, characterized in that: In step 34, the calculation formula of the linear model is: Among them, rh x is the height index and x is the height percentile.

9. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 1, characterized in that: In step 4, 5 features were input, and the random forest model was used to classify and extract artificial forests and natural forests. The data set was divided into training set, validation set and test set, which accounted for 70%, 20% and 10% of the total data respectively. The random forest model was used for classification training, 150 decision trees were set to improve the robustness of the classification, and random seeds were set to ensure the repeatability of the results.

10. The artificial forest remote sensing extraction method based on the vertical structure complexity of satellite-borne laser waveform according to claim 1, characterized in that: In step 5, the evaluation of plantation extraction accuracy and feature importance specifically includes the following steps: Step 51: After the model training is completed, the performance of the model is evaluated using the test set, and the confusion matrix, classification accuracy, and classification report are calculated to evaluate the model's extraction accuracy for plantations; Step 52: Evaluate feature importance based on Gini impurity to measure the contribution of each feature in the artificial forest extraction process.

Images