Method for establishing farmland protection forest belt decline degree division model based on remote sensing data

Abstract

The invention provides a method for establishing a farmland protection forest belt decline degree division model based on remote sensing data, which is characterized by comprising the following steps of: 1) scanning a farmland protection forest belt by using a laser radar, a hyperspectral camera and an infrared thermal imaging camera to obtain farmland protection forest belt point cloud, a hyperspectral image and infrared thermal imaging data; investigating to obtain the decline degree of the farmland protection forest belt; 2) processing the obtained farmland protection forest belt point cloud, the hyperspectral image and the thermal imaging data; 3) extracting model training parameters by using the data processed in the step 2); obtaining a model training data set; and 4) training based on the model training data set in the step 3) to obtain a model for dividing the decline degree of the farmland protection forest belt.

Description

Technical Field

[0001] The invention belongs to the technical field of forestry, and more specifically, relates to a method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data. Background Art

[0002] Forest decline is a major problem of global ecological environment degradation. Farmland shelterbelts are ecological barriers for oases in arid areas, and they play an important role in preventing and controlling sandstorm disasters, improving regional microclimate, and ensuring people's production and life. However, due to climate change, immature initial forestry technology, and harmful organisms, a large number of trees have died and grown poorly in farmland shelterbelts, and the stability of the forest stand structure has decreased, leading to the degradation or loss of the protective function. Therefore, it is urgent to classify the degree of decline of farmland shelterbelts in order to provide necessary technical support for subsequent renewal and transformation.

[0003] At present, the assessment of the decline of farmland shelterbelts mainly relies on manual methods, which requires forestry workers to conduct in-depth field investigations. However, this method is limited by geography and time, which is not only time-consuming and labor-intensive, but also has a limited scope of investigation. Compared with traditional manual monitoring, remote sensing technology is simple, real-time and economical. As emerging forestry survey methods, unmanned aerial vehicle-borne laser radar, hyperspectral and infrared thermal imaging technologies have shown great potential. Laser radar can provide three-dimensional spatial information of the target, accurately measure and analyze the structure and morphology of farmland shelterbelts; hyperspectral can obtain rich spectral information to reveal the health status and physiological characteristics of vegetation; infrared thermal imaging can understand the physiological status of trees, such as water status, photosynthesis efficiency, etc. Healthy trees usually have a more uniform temperature distribution, while trees that are dehydrated, malnourished or under other stress may experience abnormal local temperature increases or decreases, thereby assessing the overall health of the forest.

[0004] In the field of forestry monitoring, a single remote sensing data source is currently used. However, multi-source data fusion has advantages such as complementarity in time and space, and can more comprehensively reflect the characteristics of the research object. Machine learning models are known for their strong adaptability, efficiency and flexibility. Fusion of LiDAR, hyperspectral and infrared thermal imaging data and classification using machine learning models is an accurate and efficient technical approach, which lays a solid technical foundation for the classification of the degree of decline of farmland shelterbelts. Summary of the invention

[0005] The present invention aims to solve the problem that the division of the decline degree of farmland shelterbelts relies on manpower and is inefficient, and provides a method for establishing a farmland shelterbelt decline degree division model based on remote sensing data. The present invention can ensure high detection accuracy and speed. It can conduct forestry monitoring intelligently, quickly and accurately. The technical solution is as follows:

[0006] A method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data comprises the following steps:

[0007] Step 1) Scanning the farmland shelterbelt using laser radar, hyperspectral and infrared thermal imaging cameras to obtain point cloud, hyperspectral image and infrared thermal imaging data of the farmland shelterbelt; and investigating and obtaining the degree of decline of the farmland shelterbelt;

[0008] Step 2) Processing the acquired farmland shelterbelt point cloud, hyperspectral image and thermal imaging data;

[0009] Step 3) extracting model training parameters using the farmland shelterbelt point cloud, hyperspectral image and thermal imaging data processed in step 2) to obtain a model training data set;

[0010] Step 4) Based on the model training data set in step 3), a model for classifying the degree of decline of farmland shelterbelts is trained.

[0011] Optionally, in step 1), an unmanned aerial vehicle equipped with a laser radar, a hyperspectral camera and an infrared thermal imaging camera are used to scan the farmland shelterbelt to obtain point cloud, hyperspectral camera and infrared thermal imaging data of the farmland shelterbelt.

[0012] Optionally, in step 2), processing the farmland shelterbelt point cloud includes: sequentially performing calculation, denoising, ground point classification, normalization, and single tree segmentation on the farmland shelterbelt point cloud;

[0013] Optionally, processing of the farmland shelterbelt point cloud includes:

[0014] Use LiAcquire software to screen the range of the target object and perform point cloud data solution; perform point cloud data denoising in LiDAR360 software, delete point clouds that do not belong to the target object, and remove undeleted noise points;

[0015] The denoised farmland shelterbelt point cloud data is separated from the ground points using an improved progressive triangulated network encryption filtering algorithm;

[0016] Then, the DEM and DSM are generated using irregular triangulation and inverse distance weighted interpolation respectively, and the CHM is obtained by subtracting the DEM and DSM. The point cloud is normalized based on the elevation value of the point cloud data minus the pixel value of the corresponding DEM.

[0017] Generate seed points based on CHM, and then perform single tree segmentation on the normalized point cloud data based on the generated seed points.

[0018] Optionally, in step 2), processing the hyperspectral image includes: sequentially performing geometric correction, reflectivity correction, denoising, mosaicking, and cropping on the hyperspectral image;

[0019] Optionally, processing of hyperspectral images includes:

[0020] ENVI 5.3 was used for geometric correction. The hyperspectral image was registered based on the three-dimensional coordinates of the ground points obtained by the total station, and the coordinates of the hyperspectral image were converted into the WGS-84 coordinate system.

[0021] Use standard white plate data from hyperspectral images to perform reflectance correction;

[0022] The hyperspectral image is smoothed using a polynomial smoothing algorithm to remove noise; the denoised hyperspectral data is mosaicked and then the area to be studied is cropped.

[0023] Optionally, processing the infrared thermal imaging data includes: sequentially performing geometric correction, splicing, and cropping on the infrared thermal imaging data;

[0024] Optionally, the infrared thermal imaging data processing method in step 2) is: in ENVI 5.3, using the mapping relationship between geographic control points and actual geographic coordinates, geometrically correcting the data and changing the coordinate system to the WGS-84 coordinate system;

[0025] The infrared thermal imaging data are then stitched together, and finally the area of ​​the research object is cropped out.

[0026] Optionally, in step 3), the method for extracting lidar parameters is to use the airborne forestry module of the LiDAR360 software to extract the height variables, density variables, intensity variables and structural variables of the farmland shelterbelt.

[0027] Optionally, in the step 3), the method for extracting hyperspectral data is: using ENVI 5.3 software to extract the spectral reflectance of the farmland shelterbelt, and using the Band Math function to calculate the vegetation index; derivatizing the spectral reflectance curve to obtain the spectral derivative of the farmland shelterbelt; and extracting the texture features of the hyperspectral image based on the gray level co-occurrence matrix method.

[0028] Optionally, the method for extracting infrared thermal imaging data in step 3) is to use Image J analysis software to select the area where the research object is located, extract the grayscale value, and then calculate the temperature value of the selected object based on the mapping relationship between the grayscale value and the temperature value.

[0029] Optionally, in step 4), a data set is constructed using parameters extracted from lidar, hyperspectral and infrared thermal imaging data combined with the degree of decline of farmland shelterbelts surveyed on the spot, and a machine learning model is established in Python. The model is repeatedly trained, verified and adjusted based on the parameters of the data set to obtain the optimal model.

[0030] Optionally, the decline degree of the farmland shelterbelt includes:

[0031] Not declining: Dead and dying trees account for less than 10%;

[0032] Mild decline: dead and dying wood accounts for between 10% and 20%;

[0033] Moderate decline: dead and dying wood accounts for between 20% and 40%;

[0034] Severe decline: Dead and dying wood accounts for more than 40%.

[0035] The present invention has the following beneficial effects:

[0036] The method for establishing a model for classifying the degree of decline of shelterbelts proposed in the present invention combines laser radar, hyperspectral and infrared thermal imaging, and uses multi-level data to establish a model training data set. The trained model for classifying the degree of decline of shelterbelts can accurately judge the decline of shelterbelts. Using this model to evaluate the decline of shelterbelts can conduct forestry monitoring intelligently, quickly and accurately, solving the time-consuming and labor-intensive shortcomings of manual investigation in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of the present invention.

[0038] Figure 2 Farmland shelterbelt after point cloud data processing.

[0039] Figure 3 Farmland shelterbelt after hyperspectral data processing.

[0040] Figure 4 Farmland shelterbelt after infrared thermal imaging processing. DETAILED DESCRIPTION

[0041] The following examples are provided for a better understanding of the present invention, but are not limited to the best embodiment, and do not limit the content and protection scope of the present invention. Any product identical or similar to the present invention obtained by anyone under the enlightenment of the present invention or by combining the features of the present invention with other prior arts shall fall within the protection scope of the present invention. If no specific experimental steps or conditions are indicated in the examples, the operations or conditions of the conventional experimental steps described in the literature in the field can be carried out.

[0042] A method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data, the method comprising the following steps:

[0043] (1) Data Acquisition

[0044] This example uses a part of a forest belt in a forest area as the research object, and uses a DJI m600pro drone equipped with a laser radar, hyperspectral and infrared thermal imaging system to collect data on the research object. Before scanning, plan the drone flight route and adjust the gimbal posture in advance, set equipment parameters, reduce data redundancy, and improve flight efficiency and data quality. The main parameters of the measurement equipment are as follows:

[0045] Table 1 Main parameters of the equipment

[0046]

[0047] (2) Data processing

[0048] 1) Point cloud data processing:

[0049] ① Solution: Import the point cloud, base station and trajectory data obtained by the lidar into the LiAcquire software, use the Pos selection tool to cut the flight trajectory, select the WGS-84 coordinate system, perform data solution, and export the solved point cloud data into .las format;

[0050] ② Denoising: Use the LiDAR360 software to remove outlier noise points using the Statistical Outlier Removal (SOR) algorithm. The parameters that need to be set before using the algorithm are the number of neighboring points N and the standard deviation multiple T. In this example, N is set to 10 and T is set to 5. Visually inspect the denoised point cloud data and manually delete the noise points that cannot be removed by the algorithm.

[0051] ③ Ground point classification: In LiDAR360, the Improved Progressive TIN Densification (IPTD) algorithm is used to classify ground points;

[0052] ④ Normalization: In LiDAR360, the ground points and non-ground points are interpolated by triangulated irregular network (TIN) and inverse distance weighting (IDW) respectively to generate a digital elevation model (DEM) and digital surface model (DSM) with a resolution of 0.6m; the canopy height model (CHM) with the same resolution can be obtained by subtracting the DEM from the DSM; the point cloud data is normalized by subtracting the pixel value of the corresponding DEM from the elevation value (Z) of the point cloud data;

[0053] ⑤ Single tree segmentation: Import CHM into LiDAR360, click "Airborne Forestry" > "Single Tree Segmentation" > "CHM Generate Seed Points" in sequence, set the maximum tree height to 40 meters, the minimum to 2 meters, and the buffer pixel to 50, perform single tree segmentation based on the generated seed points, and use the editing tool to check the single tree segmentation results. At the same time, you can add, delete, and other manual interactive editing of seed points, and segment the point cloud again based on the edited seed points to improve the accuracy of single tree segmentation.

[0054] 2) Hyperspectral data processing:

[0055] Geometric correction: Import the forest belt hyperspectral image to be registered in ENVI 5.3, and use the three-dimensional coordinates of the ground points obtained by the total station to perform geometric correction on the forest belt hyperspectral image; Use the geometric correction tool to record the coordinates of the control points obtained by the total station, and accurately input these coordinates into the coordinates of the ground control points selected in the hyperspectral image to be registered. At least 9 control points should be selected for precise registration, convert the UTM coordinate system to the WGS-84 coordinate system, and export the hyperspectral data after geometric correction;

[0056] Reflectance correction: In ENVI 5.3, use the region of interest tool (ROI) to select the unsaturated standard white plate area. Use the spectrum curve tool to extract the reflectance of the standard white plate, and then calculate the correction coefficient of each band based on the actual reflectance of the standard white plate and the measured reflectance. Then, use the band operation to apply the correction coefficient to each band of the geometrically corrected hyperspectral data to achieve reflectance correction;

[0057] Denoising: In order to effectively remove the noise caused by various factors, the polynomial smoothing algorithm was used to filter the spectral curve after reflectance correction using Matlab 2018. During the processing, the derivative order was set to 0, the number of smoothing points was set to 5, and the polynomial degree was set to 3, which can better remove the noise;

[0058] Mosaic cropping: In ENVI 5.3 software, the denoised hyperspectral data is first mosaicked, and then the coordinate points of the farmland shelterbelt that needs to be cropped are recorded. The ROI tool is used to input the corresponding control point coordinates for cropping, and finally the cropped image data is exported.

[0059] 3) Infrared thermal imaging data processing:

[0060] In ENVI 5.3, the mapping relationship between geographic control points and actual geographic coordinates obtained by the total station is used to perform geometric correction on the infrared thermal imaging data, and the coordinate system is changed to the WGS-84 coordinate system. The geometrically corrected infrared thermal imaging data is then spliced, the area of ​​the research object is cropped out, and finally the scope of the research object is selected to read the grayscale value of the research object in the infrared thermal imaging data.

[0061] (3) Parameter extraction

[0062] 1) Point cloud data parameter extraction:

[0063] In the LiDAR360 software, statistical analysis was performed based on the elevation value (Z) of the normalized point cloud data to obtain relevant statistics of the height variable. Click "Airborne Forestry" - "Forest Parameters" - "Height Variable" in sequence, set the x size / y size to 15 meters, and the height threshold to 2 meters to obtain the distribution characteristics of the point cloud data in the height dimension. Select the output format as a CSV file, with a total of 46 parameters related to height.

[0064] By clicking the "Export LiDAR Data" command on the LiDAR360 toolbar, open the Export Wizard dialog box, select the "Surface" type, then select "Density" in "Surface Properties", and then you can select the "Points per Cell" or "Density" option to write the corresponding density information into the cell value of the output image, thereby realizing the calculation and export of density variables, a total of 10 different slices of point cloud parameters;

[0065] The intensity variable is similar to the height variable, except that the intensity value of the point is used to calculate the intensity variable instead of the height value. Click "Airborne Forestry" - "Forest Parameters" - "Intensity Variables" in sequence, set the x size / y size to 15 meters, select the output format as CSV file, a total of 42 parameters related to intensity;

[0066] In LiDAR360, click "Airborne Forestry" - "Forest Parameters" - "Coverage", set x size / y size to 15 meters, and the height threshold to 2 meters to obtain coverage parameters;

[0067] In LiDAR360, click "Airborne Forestry" - "Forest Parameters" - "Leaf Area Index" in sequence, set x size / y size to 15 meters, height threshold to 2 meters, and leaf inclination angle distribution to 0.5 to obtain the leaf area index;

[0068] In LiDAR360, click "Airborne Forestry" - "Forest Parameters" - "Gap Rate" in sequence, set x size / y size to 15 meters, and the height threshold to 2 meters to obtain the gap rate;

[0069] Based on the point cloud data after individual tree segmentation, the structural variables of each individual tree are extracted, including tree height, breast diameter, crown diameter, crown area and crown volume, and the structural variables of the forest belt are obtained by taking the average value.

[0070] 2) Hyperspectral data parameter extraction:

[0071] Vegetation Index: 28 vegetation indices are calculated using the Band Math function in ENVI 5.3;

[0072] Spectral derivative: In order to better distinguish the spectral curves of different research objects, the spectral reflectance of the research object was derived using Matlab 2018 to obtain the first-order and second-order reflectance data;

[0073] Texture features: In ENVI 5.3, the texture features of hyperspectral images are extracted based on the Gray-level Co-occurrence Matrix (GLCM) method, and the calculation of indicators is based on pixel values.

[0074] 3) Infrared thermal imaging data parameter extraction:

[0075] In ENVI 5.3 software, the corresponding relationship between the grayscale value and the temperature value set by the device is used: temperature = (grayscale value) * 0.04-273.15, the temperature is calculated, and the temperature data is exported.

[0076] (4) Construction of recession degree model:

[0077] The measured decline degree of farmland shelterbelts was divided into four decline degrees according to the proportion of dead and dying trees in the farmland shelterbelts, namely no decline (dead and dying trees accounted for less than 10%), mild decline (dead and dying trees accounted for between 10% and 20%), moderate decline (dead and dying trees accounted for between 20% and 40%), and severe decline (dead and dying trees accounted for more than 40%). All the parameters extracted in (3) were combined with the decline degree of the forest belts surveyed in the sample plots to establish a data set. The data set was standardized using the Z-Score standardization method to ensure that different features had similar scales. The standardized data set was divided into 70% for the training set and 30% for the validation set. The random forest model was constructed using Python 3.8 and the model was validated using the validation set data to obtain the farmland shelterbelt decline degree identification model.

[0078] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data, characterized in that: The following steps are involved: Step 1) Scanning the farmland shelterbelt using laser radar, hyperspectral and infrared thermal imaging cameras to obtain point cloud, hyperspectral image and infrared thermal imaging data of the farmland shelterbelt; and investigating and obtaining the degree of decline of the farmland shelterbelt; Step 2) Processing the acquired farmland shelterbelt point cloud, hyperspectral image and thermal imaging data; Step 3) extracting model training parameters using the farmland shelterbelt point cloud, hyperspectral image and thermal imaging data processed in step 2) to obtain a model training data set; Step 4) Based on the model training data set in step 3), a model for classifying the degree of decline of farmland shelterbelts is trained.

2. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1 is characterized in that: In the step 1), a drone is used to carry a laser radar, a hyperspectral camera and an infrared thermal imaging camera to scan the farmland shelterbelt to obtain point cloud, hyperspectral camera and infrared thermal imaging data of the farmland shelterbelt.

3. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1 is characterized in that: In the step 2), processing the point cloud of the farmland shelterbelt includes: solving, denoising, ground point classification, normalization, and single tree segmentation of the point cloud of the farmland shelterbelt in sequence; Preferably, processing the farmland shelterbelt point cloud includes: Use LiAcquire software to screen the range of the target object and perform point cloud data solution; perform point cloud data denoising in LiDAR360 software, delete point clouds that do not belong to the target object, and remove undeleted noise points; The denoised farmland shelterbelt point cloud data is separated from the ground points using an improved progressive triangulated network encryption filtering algorithm; Then, the DEM and DSM are generated using irregular triangulation and inverse distance weighted interpolation respectively, and the CHM is obtained by subtracting the DEM and DSM. The point cloud is normalized based on the elevation value of the point cloud data minus the pixel value of the corresponding DEM. Generate seed points based on CHM, and then perform single tree segmentation on the normalized point cloud data based on the generated seed points.

4. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1 is characterized in that: In the step 2), processing the hyperspectral image includes: sequentially performing geometric correction, reflectivity correction, denoising, mosaicking, and cropping on the hyperspectral image; Preferably, processing the hyperspectral image includes: ENVI 5.3 was used for geometric correction. The hyperspectral image was registered based on the three-dimensional coordinates of the ground points obtained by the total station, and the coordinates of the hyperspectral image were converted into the WGS-84 coordinate system. Use standard white plate data from hyperspectral images to perform reflectance correction; The hyperspectral image is smoothed using a polynomial smoothing algorithm to remove noise; the denoised hyperspectral data is mosaicked and then the area to be studied is cropped.

5. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: Processing the infrared thermal imaging data includes: geometric correction, splicing and cropping of the infrared thermal imaging data in sequence; Preferably, the infrared thermal imaging data processing method in step 2) is: in ENVI 5.3, using the mapping relationship between geographic control points and actual geographic coordinates, geometrically correcting the data and changing the coordinate system to the WGS-84 coordinate system; The infrared thermal imaging data are then stitched together, and finally the area of ​​the research object is cropped out.

6. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: In the step 3), the method for extracting laser radar parameters is to use the airborne forestry module of the LiDAR360 software to extract the height variables, density variables, intensity variables and structure variables of the farmland shelterbelt.

7. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: In the step 3), the method for extracting hyperspectral data is to use ENVI 5.3 software to extract the spectral reflectance of the farmland shelterbelt, and use the Band Math function to calculate the vegetation index; derive the spectral reflectance curve to obtain the spectral derivative of the farmland shelterbelt; and extract the texture features of the hyperspectral image based on the gray level co-occurrence matrix method.

8. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: The method for extracting infrared thermal imaging data in step 3) is to use Image J analysis software to select the area where the research object is located, extract the grayscale value, and then calculate the temperature value of the selected object based on the mapping relationship between the grayscale value and the temperature value.

9. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: In the step 4), a data set is constructed using parameters extracted from lidar, hyperspectral and infrared thermal imaging data combined with the degree of decline of farmland shelterbelts investigated on the spot, and a machine learning model is established in Python. The model is repeatedly trained, verified and adjusted based on the parameters of the data set to obtain the optimal model.

10. The method for establishing a farmland shelterbelt decline degree classification model based on remote sensing data according to claim 1, characterized in that: The decline degree of the farmland shelterbelt includes: Not declining: Dead and dying trees account for less than 10%; Mild decline: dead and dying wood accounts for between 10% and 20%; Moderate decline: dead and dying wood accounts for between 20% and 40%; Severe decline: Dead and dying wood accounts for more than 40%.

Images