Multi-spectral image and airborne laser radar point cloud combined individual tree height remote sensing estimation method

Abstract

The invention discloses a multi-spectral image and airborne laser radar point cloud combined single-tree height remote sensing estimation method, which comprises the following steps: acquiring a multi-spectral image and an airborne laser radar point cloud, and taking the multi-spectral image and the airborne laser radar point cloud as data sources; a method based on combination of object-oriented multi-scale tree crown segmentation and CHM constructed by airborne laser radar is adopted, and tree height extraction is realized by means of programming. The method comprises the following steps: firstly, extracting a multi-spectral image crown boundary by adopting object-oriented multi-scale segmentation to obtain an individual tree crown contour SHP layer; carrying out denoising, point cloud classification, interpolation and other processing on airborne laser radar point cloud data, and further constructing a canopy height model CHM; a single tree crown SHP and a canopy height model CHM are superposed, a programming means is applied, the tree crown SHP is used as a window, and the maximum value of the tree crown CHM in the SHP is searched to realize remote sensing estimation of the tree height, so that the estimation precision of the remote sensing tree height is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing estimation, and in particular to a method for remote sensing estimation of the height of a single tree by combining multispectral images and airborne laser radar point clouds. Background Art

[0002] Tree height is an important indicator of forest resource management status. It is also an important parameter for forest resource surveys and one of the key parameters for measuring forest biomass and stock. Traditional tree height information collection relies on manual field measurements, using height measuring instruments to obtain tree height information. This method not only has an impact on forest habitats, but also consumes a lot of human and financial resources, which is not conducive to conducting surveys on large-scale forest resources. In recent years, the application of low-altitude remote sensing technology of unmanned aerial vehicles in the forestry field has provided technical support for the rapid and efficient acquisition of tree height. It has the characteristics of a wide range of data collection, collection time is not restricted by weather, flexible, convenient and efficient use. Compared with traditional field survey methods, it has more outstanding advantages and is currently one of the main methods for remote sensing acquisition of forest parameters.

[0003] UAV low-altitude remote sensing technology is a geographic information acquisition technology that has gradually developed based on remote sensing technology. It mainly uses UAV platforms equipped with sensors of different performance to collect and monitor data in the target range below low altitude (1000m), and uses algorithms to analyze and interpret the ground objects. In the field of forestry, based on the different sensors such as visible light, multi-spectral, lidar, and hyperspectral carried by UAVs, high-resolution images, fine three-dimensional point clouds and other data are obtained, and the data are analyzed and interpreted to achieve remote sensing forestry parameter estimation of target tree species.

[0004] At present, remote sensing tree height estimation is mainly based on a single remote sensing data source, and the estimated tree height is obtained based on multispectral images or airborne LiDAR point cloud data. Based on multispectral images, a digital elevation model (DEM) is constructed by artificially arranging ground points, and a canopy height model is constructed together with a digital surface model (DSM) generated by a jigsaw puzzle software, and the estimated tree height is identified and determined with the help of a local maximum algorithm; while airborne LiDAR uses segmentation software such as LiDAR360 to obtain tree height; or the local maximum method is used to obtain the estimated tree height based on the conversion of point cloud data sets to raster. However, using only a single image data, it is difficult to artificially arrange ground points in forest stands of different age groups, which in turn affects the accuracy of tree height estimation; and using a single LiDAR point cloud data, the segmentation accuracy is low in forest stands with high forest density, and the omission is serious, and the estimation of single tree height is significantly underestimated. Therefore, a remote sensing estimation method for single tree height that combines multispectral images and airborne LiDAR point clouds is needed. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a single tree height remote sensing estimation method combining multispectral imagery and airborne lidar point cloud, using programming means, with the crown SHP as a window, searching for the maximum value of the crown CHM in the SHP to achieve tree height remote sensing estimation, which greatly improves the estimation accuracy of remote sensing tree height.

[0006] To achieve the above object, the present invention adopts the following technical solution: a method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud, comprising the following steps: Step 1: Based on the acquired multispectral photos and airborne lidar data, multispectral images and lidar point cloud data are obtained respectively with the help of jigsaw software and airborne lidar data production software; Step 2: Using multispectral images as the data source, an object-oriented multi-scale segmentation algorithm is used to segment the tree crown and obtain the crown outline of a single tree; Step 3: Use the cloth filtering method to separate the ground points and non-ground points from the pre-processed airborne lidar point cloud, generate the digital surface model DSM and the digital elevation model DEM respectively, and obtain the canopy height model CHM by taking the difference between the two; Step 4: Use the single tree crown profile SHP obtained by the object-oriented multi-scale segmentation method as the search window, superimpose it on the CHM constructed using the airborne lidar point cloud, and use programming to search for the maximum value in the CHM area corresponding to the crown SHP of each tree one by one, so as to obtain the remote sensing estimated tree height result of the single tree.

[0007] In a preferred embodiment, the method for stitching the multispectral images includes feature point matching, camera self-checking, and dense point cloud reconstruction to obtain a stitched multispectral image.

[0008] In a preferred embodiment, the step 2 comprises the following steps: Step 21: Select the best segmentation scale, apply the segmentation scale, and segment the tree crown; Step 22: According to the best crown segmentation result, a classification system is established, representative training samples are selected, and five features, including mean value, brightness value, shape index, correlation of co-occurrence gray matrix, and texture mean, are selected for feature screening to select the best feature; Step 23: Select the best features and use the random forest classification method for object-oriented classification to obtain the crown classification results.

[0009] In a preferred embodiment, the method for preprocessing the airborne laser radar point cloud data includes flight strip stitching, point cloud denoising, point cloud filtering, and projection transformation.

[0010] In a preferred embodiment, the method for constructing the canopy height model CHM comprises the following steps: Step 41: Use the cloth filter algorithm in the point cloud processing software CloudCompare to separate ground points from non-ground points to obtain a ground point cloud, and use the inverse distance weighted method to interpolate and derive a 0.1 m digital elevation model DEM; Step 42: Separate the ground points and non-ground points by cloth filtering to obtain non-ground points, convert the non-ground point dataset into a raster, and derive a digital surface model DSM with a resolution of 0.1 m; Step 43: The canopy height model CHM is obtained by subtracting the digital surface model DSM from the digital elevation model DEM, and the resolution of CHM is 0.1m.

[0011] In a preferred embodiment, the single tree crown profile SHP obtained by object-oriented multi-scale segmentation is used as a search window, and the maximum value of CHM in the corresponding SHP window is searched by programming, including the following steps: First, use the geopandas library to read vector data and the rasterio library to read raster data; then, initialize an empty list max_values ​​to store the local maximum value and the coordinates of the maximum value point of each face; Secondly, use a for loop to traverse each face and extract the local maximum value with the face's geometric information as the neighborhood; specifically, use the geometry_mask function to generate a mask, convert the face's geometric information into a mask in the grid coordinate system, then use the np.where function to obtain the pixel value within the mask, and use the np.max function to obtain the local maximum value; Again, use the np.unravel_index function to get the coordinates of the local maximum point and add the result to the list max_values; here use the raster_data.xy function to convert the coordinates in the raster coordinate system to the coordinates in the geographic coordinate system; Finally, we create a GeoDataFrame of the maximum value points using the gpd.GeoDataFrame function and export the result to a point vector file using the to_file function.

[0012] In a preferred embodiment, the object-oriented multi-scale segmentation method based on multispectral images is used to obtain the single tree crown profile SHP, the CHM is constructed using the airborne lidar point cloud, and the single tree crown profile SHP obtained by object-oriented multi-scale segmentation is used as the search window. Programming is used to search for the maximum value of the corresponding CHM tree by tree, that is, the remote sensing single tree height estimation result.

[0013] Compared with the prior art, the present invention has the following beneficial effects: the present invention combines multispectral images and lidar point clouds, uses the tree crown SHP boundary obtained by object-oriented multi-scale segmentation of multispectral images as a basis, and superimposes it on the canopy height model CHM constructed by point clouds, and uses programming means to obtain the maximum value of CHM estimated tree height in the tree crown SHP area, thereby realizing tree height estimation by combining multispectral images and airborne lidar point clouds, avoiding the shortcomings of estimating tree height with single remote sensing data, and improving the accuracy of tree height estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a technical flow chart of a preferred embodiment of the present invention; Figure 2 This is a tree crown segmentation diagram for all age groups according to a preferred embodiment of the present invention; Figure 3 A CHM flow chart for constructing a laser radar point cloud for a preferred embodiment of the present invention; Figure 4 It is a schematic diagram of the tree crown boundary superimposed on the CHM programming process in the preferred embodiment of the present invention; Figure 5 A schematic diagram of a tree height result superimposed on an image in a preferred embodiment of the present invention; Figure 6 This is a fitting diagram of the measured tree height and the estimated tree height according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0016] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0017] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0018] like Figure 1-6 As shown, this implementation example includes the following steps: A. obtain a multispectral image and an airborne laser radar point cloud, preprocess the multispectral image to obtain a multispectral image, and preprocess the airborne laser radar point cloud to obtain a laser radar point cloud; The drone data was acquired using the Phanttom4 Multispectral drone, which was equipped with an integrated multispectral imaging system that integrated visible light and five multispectral cameras. During flight, six different image data, including visible light (RGB), red, green, blue, red edge, and near infrared (NIR), were acquired. The present invention used ArcGIS 10.8 software to merge the five single-band images to obtain multispectral data.

[0019] Pegasus LiDAR acquires LiDAR point cloud data and is equipped with an integrated system LiDAR scanning module D-LiDAR500. The D-LiDAR500 laser scanning module is mainly composed of a system controller, a laser controller, a camera controller, a laser scanner, and an operation and navigation terminal. The D-LiDAR500 has an inertial sampling frequency of 300Hz, a laser class of Class 1, a measurement accuracy of 2cm, and a laser field of view of 40.3°. The data was collected in August 2022, with clear and windless weather and flying at noon.

[0020] The present invention selects a typical southern tree species, a Chinese fir plantation, sets up 10 standard plots, and surveys a total of 300 Chinese fir trees. The forests are divided into 2 age groups, including 2 each of young forests, middle-aged forests, near-mature forests, mature forests, and over-mature forests. RTK positioning is performed on the 300 randomly selected Chinese fir trees and the boundaries of the standard plots.

[0021] In this implementation example, multispectral images are preprocessed using DJI Zhitu software, which integrates puzzle, calibration, coordinate and other functions to obtain multispectral images with a resolution of 0.1m, and ArcGIS 10.8 software is used to crop standard ground images. The lidar data uses CloudCompare software to perform point cloud denoising, filtering, classification, etc.

[0022] B. Segmenting the tree crown in the multispectral image by using an object-oriented multi-scale segmentation algorithm, including: multi-scale segmentation, feature optimization, and random forest classification.

[0023] (1) Select the best segmentation scale, apply the segmentation scale, and segment the tree crown; (2) Based on the best crown segmentation results, a classification system is established, representative training samples are selected, and five features, including mean, brightness value, shape index, correlation of co-occurrence grayscale matrix, and texture mean, are used for feature screening to select the best feature; (3) Select the best features and use the random forest classification method for object-oriented classification to obtain the crown classification results.

[0024] The digital surface model DSM and the digital elevation model DEM constructed by the pre-processed airborne laser radar point cloud are subtracted to construct a canopy height model CHM; The lidar point cloud data is classified using the point cloud classification software CloundCompare, and the cloth filter algorithm is used to separate ground points and non-ground points to obtain the ground point cloud. The inverse distance weighted method is used to interpolate and derive a 0.1m digital elevation model DEM. The non-ground point dataset is converted to a raster to derive a digital surface model DSM with a resolution of 0.1m. The canopy height model CHM is obtained by subtracting the digital surface model DSM from the digital elevation model DEM, and the CHM resolution is 0.1m.

[0025] C uses programming methods, takes the single tree crown SHP obtained by object-oriented multi-scale segmentation as the search window, and uses programming to search for the maximum tree height in each crown CHM to achieve tree height extraction.

[0026] Table 1 shows the accuracy evaluation of multi-scale segmentation of single trees in different age groups. Standard Encoding Age Group Measured number of plants Accurate segmentation Missing segmentation Accuracy 1 Young forest 151 147 4 97.4% 2 Young forest 162 153 9 94.4% 3 Middle-aged forest 84 84 0 100.0% 4 Middle-aged forest 74 74 0 100.0% 5 Nearly mature forest 75 75 0 100.0% 6 Nearly mature forest 62 62 0 100.0% 7 Mature forest 92 78 14 84.8% 8 Mature forest 102 76 26 74.5% 9 Overmature forest 47 47 0 100.0% 10 Overmature forest 40 40 0 100.0% The evaluation results of segmentation accuracy in Table 1 show that the overall segmentation accuracy of young forest, middle-aged forest, nearly mature forest, mature forest and over-mature forest is as high as 100% and as low as 74.5%. The main reason for the high segmentation accuracy is that the plant density is low and the crowns are not adhered, so the multi-scale segmentation effect is better. However, when the plant density is high, the crown segmentation accuracy will be affected even if it is a young forest.

[0027] E uses the crown SHP obtained by segmentation as a search window and superimposes it on the CHM constructed by the airborne lidar point cloud. The maximum tree height in each crown CHM is searched by programming to output the estimated tree height result. The evaluation results of the tree height estimation accuracy of each age group are shown in Table 2. According to the fitting results with the measured tree height, the young forest R 2 is 0.86, and RMSE is 0.3m; the middle-aged forest R 2 is 0.89, and RMSE is 0.7m; 2 is 0.92, and RMSE is 0.8m; R 2 is 0.82, RMSE is 1.4m; overmature forest R 2 The main reason for this phenomenon is the combined effect of the difference in canopy density within the standard plot and the age of the tree species.

[0028] Table 2. Tree height estimation accuracy for each age group Age Group <![CDATA[R 2 ]]> RMSE Young forest 0.86 0.3 Middle-aged forest 0.89 0.7 Nearly mature forest 0.92 0.8 Mature forest 0.82 1.4 Overmature forest 0.95 1.1 The measured tree heights of the 300 trees in this survey were matched with the tree heights estimated by remote sensing ( Figure 1 ), the overall estimate R 2 is 0.98 and the RMSE is 1m.

[0029] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud, characterized in that: The following steps are involved: Step 1: Based on the acquired multispectral photos and airborne lidar data, multispectral images and lidar point cloud data are obtained respectively with the help of jigsaw software and airborne lidar data production software; Step 2: Using multispectral images as the data source, an object-oriented multi-scale segmentation algorithm is used to segment the tree crown and obtain the crown outline of a single tree; Step 3: Use the cloth filtering method to separate the ground points and non-ground points from the pre-processed airborne lidar point cloud, generate the digital surface model DSM and the digital elevation model DEM respectively, and obtain the canopy height model CHM by taking the difference between the two; Step 4: Use the single tree crown profile SHP obtained by the object-oriented multi-scale segmentation method as the search window, superimpose it on the CHM constructed using the airborne lidar point cloud, and use programming to search for the maximum value in the CHM area corresponding to the crown SHP of each tree one by one, so as to obtain the remote sensing estimated tree height result of the single tree.

2. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The method for stitching the multispectral images includes feature point matching, camera self-checking, and dense point cloud reconstruction to obtain a stitched multispectral image.

3. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The step 2 comprises the following steps: Step 21: Select the best segmentation scale, apply the segmentation scale, and segment the tree crown; Step 22: According to the best crown segmentation result, a classification system is established, representative training samples are selected, and five features, including mean value, brightness value, shape index, correlation of co-occurrence gray matrix, and texture mean, are selected for feature screening to select the best feature; Step 23: Select the best features and use the random forest classification method for object-oriented classification to obtain the crown classification results.

4. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The method for preprocessing the airborne laser radar point cloud data includes flight strip stitching, point cloud denoising, point cloud filtering, and projection transformation.

5. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The method for constructing the canopy height model CHM comprises the following steps: Step 41: Use the cloth filter algorithm in the point cloud processing software CloudCompare to separate ground points from non-ground points to obtain a ground point cloud, and use the inverse distance weighted method to interpolate and derive a 0.1 m digital elevation model DEM; Step 42: Separate the ground points and non-ground points by cloth filtering to obtain non-ground points, convert the non-ground point dataset into a raster, and derive a digital surface model DSM with a resolution of 0.1 m; Step 43: The canopy height model CHM is obtained by subtracting the digital surface model DSM from the digital elevation model DEM. The resolution of CHM is 0.1 m.

6. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The single tree crown profile SHP obtained by object-oriented multi-scale segmentation is used as the search window, and the maximum value of CHM in the corresponding SHP window is searched by programming, including the following steps: First, use the geopandas library to read vector data and the rasterio library to read raster data; then, initialize an empty list max_values ​​to store the local maximum value and the coordinates of the maximum value point of each face; Secondly, use a for loop to traverse each face and extract the local maximum value with the face's geometric information as the neighborhood; specifically, use the geometry_mask function to generate a mask, convert the face's geometric information into a mask in the grid coordinate system, then use the np.where function to obtain the pixel value within the mask, and use the np.max function to obtain the local maximum value; Again, use the np.unravel_index function to get the coordinates of the local maximum point and add the result to the list max_values; here use the raster_data.xy function to convert the coordinates in the raster coordinate system to the coordinates in the geographic coordinate system; Finally, we create a GeoDataFrame of the maximum value points using the gpd.GeoDataFrame function and export the result to a point vector file using the to_file function.

7. The method for remote sensing estimation of single tree height by combining multispectral imagery and airborne laser radar point cloud according to claim 1, characterized in that: The method adopts an object-oriented multi-scale segmentation method based on multispectral images to obtain the single tree crown profile SHP, uses the CHM constructed by airborne laser radar point cloud, and uses programming to search for the maximum value of the corresponding CHM tree by tree, that is, the remote sensing single tree height estimation result.

Images