System for estimating diameter at breast height of forest in Yunnan province based on multi-source remote sensing data

Abstract

The invention relates to the technical field of forest resource monitoring, and particularly discloses a Yunnan province forest diameter at breast height estimation system based on multi-source remote sensing data, which comprises a data preprocessing module, a feature extraction module, a model construction module, a model training and verification module, a diameter at breast height estimation module and a result visualization module, the method has remarkable advantages in the aspects of data source fusion, algorithm application, automation degree and result visualization. Through multi-source remote sensing data and an advanced machine learning technology, high-precision estimation of the forest diameter at breast height in the Yunnan province is realized, a solution for dynamic monitoring of forest resources is provided, and the method is suitable for large-scale ecological monitoring and forestry management.

Description

Technical Field

[0001] The invention belongs to the technical field of forest resource monitoring, and in particular relates to a Yunnan forest breast diameter estimation system based on multi-source remote sensing data. Background Art

[0002] Forest diameter at breast height (DBH) is an important indicator for assessing forest structure and biomass, and its measurement accuracy directly affects the assessment of forest ecosystem service functions. However, due to the complex terrain and diverse vegetation types in Yunnan Province, traditional field survey methods are time-consuming and labor-intensive and difficult to fully cover.

[0003] The development of multi-source remote sensing technology has provided a new way to quickly monitor large-scale forest resources. By combining multiple remote sensing data such as optical, radar and terrain, and applying advanced machine learning algorithms, forest diameter at breast height can be effectively estimated, thus providing support for accurate forest resource management. Summary of the invention

[0004] The purpose of the present invention is to provide a system for estimating the diameter at breast height of forests in Yunnan Province based on multi-source remote sensing data, so as to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A system for estimating the diameter at breast height of forests in Yunnan Province based on multi-source remote sensing data, the system comprising the following modules:

[0007] Data preprocessing module: radiometric calibration, atmospheric correction, geometric correction, cloud detection, shadow removal and georeferencing for optical remote sensing, radar remote sensing, terrain data and climate data;

[0008] Feature extraction module: extract vegetation index, radar polarization information, terrain factors and other related features from the pre-processed remote sensing data; the extraction of vegetation index includes NDVI and EVI; the polarization characteristics of the radar remote sensing image include VV and VH polarization information; the extraction of terrain features includes slope, aspect and elevation factors;

[0009] Model building module: Based on the field survey data, a forest DBH estimation model is constructed using a machine learning algorithm. The model includes at least one of the random forest model, support vector machine model, XGBoost model and other machine learning algorithms. The DBH estimation model is established in combination with the measured data.

[0010] Model training and validation module: train the DBH estimation model and perform cross-validation, RMSE, R 2The model training and verification module is based on sample data from different regions and forest types in Yunnan Province to conduct regional model training and accuracy verification to improve the applicability of the model.

[0011] DBH estimation module: Analyze remote sensing images based on the trained model to generate a DBH spatial distribution map within Yunnan Province;

[0012] Result visualization module: supports diversified output of estimation results, including Shapefile, GeoTIFF and other formats, and provides visualization and dynamic monitoring functions of DBH distribution. The result visualization module further supports dynamic monitoring of time series remote sensing data, and can display time series graphs and animation effects of DBH changes.

[0013] Preferably, the radiation calibration is to read the DN (Digital Number) value of the remote sensing image, and use the radiation calibration formula to convert the DN value into a radiation brightness value or a reflectance value. The atmospheric correction is to use an atmospheric correction model, and the atmospheric correction model includes but is not limited to FLAASH and MODTRAN. The image is geometrically precisely corrected to collect ground control point (GCPs) data, identify the corresponding positions of GCPs on the image, use polynomial transformation or affine transformation to calculate the transformation relationship between the image and the geographic coordinate system, and apply the transformation relationship to resample the image to obtain a geometrically precisely corrected image. The cloud detection and shadow removal are to prepare a training data set for clouds and shadows, including remote sensing images with cloud and shadow annotations, train a Mask R-CNN model to enable it to identify cloud and shadow areas in the image, apply the trained Mask R-CNN model to the remote sensing image to be processed, obtain a mask image of clouds and shadows, and use the mask image to fill or remove cloud and shadow areas in the original image.

[0014] Preferably, the smaller the RMSE is, the smaller the difference between the model prediction value and the measured value is. 2 The closer it is to 1, the better the model fits the data.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] Existing forest DBH estimation methods often rely on a single data source, such as optical remote sensing or radar remote sensing data, resulting in limited model accuracy. The present invention enriches the feature information required for DBH estimation by integrating multiple sources such as optical, radar, terrain, and climate data, thereby improving the overall accuracy of the model. By combining the penetration ability of radar data, the problem of missing image information under cloud cover and complex terrain is solved.

[0017] Advantages: The fusion of multi-source data significantly improves the stability and applicability of the model, especially under complex terrain and climatic conditions.

[0018] Compared with traditional empirical models or linear regression models, this paper introduces advanced machine learning algorithms such as random forest, SVM, and XGBoost. These algorithms can automatically capture the nonlinear relationship between remote sensing features and forest diameter at breast height, greatly improving the estimation accuracy. At the same time, the model can automatically adjust according to the characteristics of different regions, enhancing its versatility.

[0019] Advantages: The machine learning algorithm can process high-dimensional complex data and has good generalization ability, which is particularly suitable for the forest ecosystem in Yunnan Province with large heterogeneity.

[0020] Traditional methods usually require a lot of manual participation for data processing and result analysis. The present invention greatly improves work efficiency through fully automated data preprocessing, model training and result estimation processes. At the same time, the system provides intuitive visualization functions, and users can directly view the DBH distribution and its changing trends through the interface, simplifying the interpretation and application of the results.

[0021] Advantages: Automated processing and result visualization not only reduce human errors, but also accelerate forest resource monitoring and management decision-making.

[0022] The present invention is particularly targeted at the complex and diverse forest types and climate conditions in Yunnan Province. Through regional model training and verification, the model can achieve high estimation accuracy in different regions. This feature enables the system to be widely used in DBH estimation in a large range of regions, rather than just limited to small-scale experimental studies.

[0023] Advantages: It has wide regional adaptability and strong potential for promotion and application, and can provide reliable technical support for forest monitoring in different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the distribution map of forest ecosystem types in Yunnan Province;

[0025] Figure 2 A graph showing the estimation of aboveground biomass using the random forest method combined with the breast diameter in Example 1 of the present invention;

[0026] Figure 3 Schematic diagram of breast diameter distribution in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] Embodiment 1:

[0029] Combination Figures 1 to 3 : A forest DBH estimation system based on multi-source remote sensing data, DBH estimation in a forest area in Yunnan Province.

[0030] 1. Data Collection and Processing

[0031] Data Collection:

[0032] Obtain optical remote sensing images of the area with a resolution of 30 meters in 2023 from the Landsat 8 satellite.

[0033] Acquire C-band radar remote sensing images of the same period from the Sentinel-1 satellite.

[0034] Obtain high-precision DEM data with a resolution of 30 meters from SRTM.

[0035] Get the monthly average temperature and precipitation data for 2023 from the local weather station.

[0036] Through field surveys, DBH data were measured in 100 sample plots, 10 trees were measured in each sample plot, and a total of 1,000 DBH data were obtained.

[0037] Data preprocessing:

[0038] Radiometric calibration and atmospheric correction are performed on Landsat 8 optical remote sensing images to eliminate the influence of the atmosphere on the images.

[0039] Perform geometric precision correction on the image to ensure that the spatial position of the image is accurate.

[0040] Use deep learning models such as Mask R-CNN for cloud detection and shadow removal to ensure image quality.

[0041] Optical remote sensing images, radar remote sensing images, DEM data and climate data are georeferenced and unified into the same coordinate system.

[0042] 2. Feature extraction

[0043] Vegetation index extraction:

[0044] Calculate NDVI: NDVI = (NIR-R) / (NIR+R), where NIR is the near-infrared band reflectance and R is the red band reflectance. Extract the NIR and R band data from the Landsat 8 image and calculate the NDVI value for each pixel.

[0045] Radar polarization information extraction:

[0046] VV and VH polarization information is extracted from Sentinel-1 radar remote sensing images. This information reflects the scattering characteristics of the forest and is closely related to the structure and biomass of the forest.

[0047] Terrain factor extraction:

[0048] Extract terrain factors such as slope, aspect and elevation from DEM data. Slope indicates the inclination of the surface, aspect indicates the direction of the surface, and elevation indicates the height of the surface.

[0049] The specific steps are as follows:

[0050] Vegetation index extraction method

[0051] NDVI is one of the most commonly used vegetation indices. The method for calculating NDVI using ENVI software is:

[0052] Open the ENVI software and select "Transform" → "NDVI" in the main menu.

[0053] In the pop-up "NDVI Calculation Input File" dialog box, select the preprocessed remote sensing image file, such as Landsat TM data.

[0054] Click the “OK” button to open the “NDVI Calculation Parameters” dialog box.

[0055] In the "Input File Type" drop-down list, select the remote sensing sensor type, such as Landsat TM, and set the band numbers of the red band (Red) and the near infrared band (Near IR) (for Landsat TM, the red band is 3 and the near infrared band is 4).

[0056] Set the output file path and click the "OK" button to calculate the NDVI image.

[0057] 2. EVI takes into account factors such as atmospheric correction and can more accurately reflect vegetation conditions. Open the ENVI software and select "Transform" → "Band Math" in the main menu.

[0058] In the pop-up "Band Math" dialog box, enter the EVI calculation formula: fix(2.5*(b5-b4) / (b5+6.0*b4-7.5*b1+1)), where b5 represents the near-infrared band, b4 represents the red light band, and b1 represents the blue light band.

[0059] Click the "OK" button to perform band calculation and get the EVI image.

[0060] 2. Radar Polarization Information Extraction Method

[0061] Radar polarization information reflects the scattering characteristics of surface targets. The extraction method is:

[0062] Obtain preprocessed radar polarimetric image data.

[0063] According to the radar system parameters and image metadata, the band number or data layer corresponding to the VV and VH polarization channels is determined.

[0064] Use radar image processing software (such as ENVI, PolSARpro, etc.) to open the image data and extract the data of the VV and VH polarization channels respectively.

[0065] The extracted data are further processed and analyzed, such as polarization decomposition and scattering matrix calculation, to obtain more detailed polarization feature information.

[0066] 3. Terrain Factor Extraction Method

[0067] Terrain factors include slope, aspect and elevation, which play an important role in terrain analysis and object recognition. The following is a DEM extraction method:

[0068] Get the preprocessed DEM data.

[0069] Use GIS software (such as ArcGIS, QGIS, etc.) to open the DEM data.

[0070] Use terrain analysis tools in GIS software, such as “slope analysis”, “aspect analysis” and “elevation extraction”, to calculate the slope, aspect and elevation factors respectively.

[0071] Export the calculation results into the required format (such as raster data, vector data, etc.) for subsequent analysis and application.

[0072] 3. Model construction and training

[0073] Data preparation:

[0074] The extracted vegetation index (NDVI, EVI), radar polarization information (VV, VH), terrain factors (slope, aspect, elevation) and climate data (average monthly temperature, precipitation) were matched with the DBH measured data. Since the resolution of remote sensing data is low (30 meters), the average value of remote sensing data in the sample plot corresponding to each DBH measured data is taken as the characteristic value of the sample plot.

[0075] Construct training and validation data sets, using 70% of the plot data as the training set and the remaining 30% as the validation set.

[0076] Model selection:

[0077] The Stacking model was selected as the DBH estimation model. Random Forest, Support Vector Machine (SVM) and XGBoost were selected for machine learning.

[0078] Model training:

[0079] For the random forest model, the number of decision trees is set to 100 and the maximum depth is set to 10.

[0080] For the support vector machine model, the radial basis function (RBF) is selected as the kernel function, the penalty parameter C is set to 10, and the kernel function parameter gamma is set to 0.1.

[0081] For the XGBoost model, the number of decision trees was set to 100, the learning rate was set to 0.1, and the maximum depth was set to 6.

[0082] The model was trained using a 5-fold cross-validation method combined with geospatial partitioning. In each cross-validation, the training set was further divided into a training subset and a validation subset for adjusting model parameters and evaluating model performance.

[0083] Model Validation:

[0084] By calculating the RMSE (root mean square error) and R 2 (coefficient of determination) is used to evaluate the accuracy and robustness of the model. The smaller the RMSE, the smaller the difference between the model prediction value and the measured value; R 2 The closer it is to 1, the better the model fits the data.

[0085] 4. DBH estimation and result visualization

[0086] Diameter at breast height estimation:

[0087] The pre-processed remote sensing images are analyzed using the trained Stacking model. For each pixel (or pixel), the extracted feature value is input into the model to obtain the DBH prediction value corresponding to the pixel.

[0088] The discrete pixel DBH prediction values ​​are converted into a continuous DBH spatial distribution map through interpolation methods (such as Kriging interpolation).

[0089] Result visualization:

[0090] Develop an interactive WebGIS platform to output the DBH spatial distribution map in Shapefile or GeoTIFF format.

[0091] It provides a visual display function for the DBH distribution, and users can browse the DBH distribution in different areas through the platform.

[0092] It provides dynamic monitoring function, and users can view the trend animation of DBH changes over time. For example, users can compare the DBH spatial distribution maps of different years to analyze the growth changes of forests.

[0093] NDVI calculation formula:

[0094] NDVI = (NIR-R) / (NIR+R)

[0095] Among them, NIR is the reflectance of the near infrared band, and R is the reflectance of the red light band. The data of these two bands can be directly extracted from the processed remote sensing image.

[0096] Stacking model formula:

[0097] For each base learner h_k(x) (k=1,2,...,K) in the Stacking model, its prediction result can be calculated by the following formula:

[0098] h_k(x)=f_k(Φ(x))

[0099] Among them, Φ(x) is the feature vector of the input sample x, and f_k is the prediction function of the kth base learner.

[0100] The final prediction result H(x) of the Stacking model can be calculated by the following formula:

[0101] H(x)=Σ_kα_k h_k(x)+β

[0102] Among them, α_k and β are the parameters of the meta-learner, which can be obtained through training. In practical applications, the meta-learner usually chooses simple models such as linear regression or logistic regression.

[0103] Embodiment 2:

[0104] May 2023 was selected as the time point for data collection to ensure time consistency among all data sources.

[0105] Remote sensing data: Download remote sensing images covering the target forest area from the Landsat 8OLI satellite with a resolution of 30 meters. Download remote sensing images of the same area from the Sentinel 2A satellite with a resolution of 10 meters. Download remote sensing images of the same area from the Gaofen-2 (GF2) satellite with a resolution of 4 meters.

[0106] DEM data: Download DEM data of the area from the Geospatial Data Cloud with a resolution of 30 meters.

[0107] Meteorological data: The average monthly precipitation (MAP) and average monthly temperature (MAT) data for May 2023 were obtained from the local meteorological station.

[0108] Field survey data: 100 representative plots were selected in the target forest area, each with a size of 10 m × 10 m. The coordinates of the centroid of the plots were recorded using RTK-GPS, and the aboveground biomass and underground biomass of each plot were measured.

[0109] Data preprocessing

[0110] Radiometric calibration and atmospheric correction: Radiometric calibration and atmospheric correction are performed on Landsat 8OLI, Sentinel 2A and GF2 remote sensing images to eliminate the influence of the atmosphere on the images.

[0111] Geometric precision correction: Use known control points to perform geometric precision correction on remote sensing images to ensure that the spatial position of the image is accurate.

[0112] Image resampling: Landsat 8OLI and GF2 images were resampled to the same resolution (10 m) as Sentinel 2A images for subsequent analysis.

[0113] Vegetation index calculation: Calculate the NDVI and GNDVI vegetation indices to describe the vegetation coverage and growth conditions of the forest.

[0114] Extraction of terrain factors: Extract terrain factors such as slope and aspect from DEM data to analyze the impact of terrain on forest biomass.

[0115] Specifically, use ENVI software for radiation calibration. Open the remote sensing image in ENVI, select the radiation calibration tool, enter the gain and offset parameters obtained from the metadata file, and perform calibration. Taking Landsat 8OLI as an example, use the FLAASH atmospheric correction tool in ENVI. Enter the imaging time, select the appropriate atmospheric model and aerosol model, perform atmospheric correction, and use ENVI software for geometric precision correction. Select at least 3 control points with known geographic coordinates on the image, select the polynomial model, enter the image coordinates and geographic coordinates of the control points, perform correction, and use ENVI software for resampling. Select bilinear interpolation as the resampling method, set the target resolution to 10 meters, and perform resampling.

[0116] Vegetation index calculation: NDVI (Normalized Difference Vegetation Index)

[0117] Formula: NDVI = (NIR-R) / (NIR+R)

[0118] Among them, NIR is the reflectivity of the near-infrared band, and R is the reflectivity of the red light band.

[0119] Open the remote sensing image in ENVI, select the band operation tool, enter the NDVI calculation formula, and perform the calculation to obtain the NDVI image.

[0120] GNDVI (Green Normalized Difference Vegetation Index)

[0121] Formula: GNDVI = (NIR-G) / (NIR+G)

[0122] Among them, NIR is the reflectivity of the near-infrared band, and G is the reflectivity of the green light band.

[0123] Also in ENVI, use the band operation tool to enter the GNDVI calculation formula and perform the calculation to obtain the GNDVI image.

[0124] Terrain factor extraction

[0125] Slope formula:

[0126] slope=arctan(√((dz / dx) 2 +(dz / dy) 2 ))

[0127] Among them, dz / dx and dz / dy are the elevation gradients of the DEM data in the x and y directions, respectively.

[0128] Slope aspect formula:

[0129] aspect = arctan(dz / dy) / dz / dx)

[0130] Open the DEM data in ENVI, select the terrain factor extraction tool, enter the necessary parameters (such as output resolution, etc.), select to extract slope and aspect factors, and execute the extraction to obtain the corresponding terrain factor image.

[0131] 2. Feature Selection and Variable Construction

[0132] Feature Selection

[0133] Boruta algorithm: The Boruta algorithm is combined with a random forest learner to select feature variables that are highly correlated with forest biomass from vegetation indices, texture features, terrain factors, and environmental factors extracted from remote sensing images.

[0134] Feature variable screening: Based on the output of the Boruta algorithm, feature variables with higher scores are selected as input for subsequent modeling.

[0135] Variable construction

[0136] Sample data set construction: The selected characteristic variables were matched with the measured data of forest biomass obtained from the field survey to construct a sample data set.

[0137] Dataset division: The sample dataset is randomly divided into a training set (70%) and a test set (30%) for subsequent model training and evaluation.

[0138] 3. Model construction and training

[0139] Model selection

[0140] Machine learning algorithms: Four algorithms, namely random forest (RF), gradient boosted decision tree (GBDT), support vector machine (SVM) and k-nearest neighbor (k-NN), were selected for forest biomass estimation.

[0141] Model Training

[0142] Parameter tuning: perform parameter tuning for each algorithm, adjust the number of decision trees and maximum depth of random forest; adjust the learning rate and number of iterations of GBDT; adjust the penalty parameter C and kernel function parameters of SVM; adjust the number of neighbors of k-NN, etc.

[0143] Cross-validation: Use 5-fold cross-validation to determine the best parameter combination for each algorithm, and train the model using the training set data.

[0144] IV. Model Evaluation and Selection

[0145] Model Evaluation

[0146] Evaluation indicators: Use the test set data to evaluate the trained model and calculate the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R2 ) and other indicators.

[0147] Performance comparison: Compare the performance of the four algorithms on the test set and select the algorithm with the smallest RMSE and R 2 The model closest to 1 was taken as the final forest biomass estimation model.

[0148] Model selection

[0149] Optimal model determination: Assume that after evaluation, the GBDT model performs best on the test set (e.g., RMSE = 10Mg / ha, R 2 =0.95), the GBDT model was selected as the final forest biomass estimation model.

[0150] 5. Biomass estimation and result output

[0151] Biomass estimation

[0152] Model application: The selected GBDT model was applied to the preprocessed remote sensing images to estimate the biomass of the target forest area.

[0153] Result output: Output the predicted biomass value of each pixel and generate a spatial distribution map of forest biomass.

[0154] Result Output

[0155] Data format: The biomass spatial distribution map was exported in GeoTIFF format to facilitate subsequent analysis and application.

[0156] Develop visualization tools or platforms to allow users to browse and query biomass estimation results in different regions through WebGIS and other means.

[0157] In summary, the present invention has significant advantages in data source fusion, algorithm application, automation and result visualization. Through multi-source remote sensing data and advanced machine learning technology, the present invention achieves high-precision estimation of forest DBH in Yunnan Province and provides a solution for dynamic monitoring of forest resources, which is suitable for large-scale ecological monitoring and forestry management.

[0158] It will be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but will be a routine task of design, fabrication, and production for those of ordinary skill having the benefit of this disclosure without undue experimentation.

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A forest diameter at breast height estimation system based on multi-source remote sensing data, characterized in that: The system includes the following modules: Data preprocessing module: radiometric calibration, atmospheric correction, geometric correction, cloud detection, shadow removal and georeferencing for optical remote sensing, radar remote sensing, terrain data and climate data; Feature extraction module: extract vegetation index, radar polarization information, terrain factors and other relevant features from pre-processed remote sensing data; The extraction of vegetation index includes NDVI and EVI; the polarization characteristics of radar remote sensing images include VV and VH polarization information; the extraction of terrain features includes slope, aspect and elevation factors; Model building module: Based on the field survey data, a forest DBH estimation model is constructed using a machine learning algorithm. The model includes at least one of the random forest model, support vector machine model, XGBoost model and other machine learning algorithms. The DBH estimation model is established in combination with the measured data. Model training and validation module: train the DBH estimation model and perform cross-validation, RMSE, R 2 The model training and verification module is based on sample data from different regions and forest types in Yunnan Province to conduct regional model training and accuracy verification to improve the applicability of the model. DBH estimation module: Analyze remote sensing images based on the trained model to generate a DBH spatial distribution map within Yunnan Province; Result visualization module: supports diversified output of estimation results, including Shapefile, GeoTIFF and other formats, and provides visualization and dynamic monitoring functions of DBH distribution. The result visualization module further supports dynamic monitoring of time series remote sensing data, and can display time series graphs and animation effects of DBH changes.

2. The Yunnan Province forest diameter at breast height estimation system based on multi-source remote sensing data according to claim 1 is characterized by: The radiation calibration is to read the DN (Digital Number) value of the remote sensing image, and use the radiation calibration formula to convert the DN value into a radiation brightness value or a reflectivity value. The atmospheric correction is to use an atmospheric correction model, and the atmospheric correction model includes but is not limited to FLAASH and MODTRAN. The image is geometrically precisely corrected to collect ground control point (GCPs) data, identify the corresponding positions of GCPs on the image, use polynomial transformation or affine transformation to calculate the transformation relationship between the image and the geographic coordinate system, and apply the transformation relationship to resample the image to obtain a geometrically precisely corrected image. The cloud detection and shadow removal are to prepare a training data set for clouds and shadows, including remote sensing images with cloud and shadow annotations, train a Mask R-CNN model to enable it to identify cloud and shadow areas in the image, apply the trained Mask R-CNN model to the remote sensing image to be processed, obtain a mask image of clouds and shadows, and use the mask image to fill or remove cloud and shadow areas in the original image.

3. The system for estimating the forest diameter at breast height in Yunnan Province based on multi-source remote sensing data according to claim 1 is characterized in that: The smaller the RMSE, the smaller the difference between the model prediction value and the measured value. 2 The closer it is to 1, the better the model fits the data.

Images