Biomass monitoring method based on warming grassland
By collecting 3D point cloud data of warming grasslands using drones and combining iterative optimization with local density zoning and triangular meshes, along with canopy structure and diurnal warming characteristics, a dual-branch neural network model was used to solve the problems of low accuracy and efficiency in grassland biomass monitoring, thus achieving efficient and accurate biomass monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF MOUNTAIN HAZARDS & ENVIRONMENT CHINESE ACADEMY OF SCI
- Filing Date
- 2026-04-16
- Publication Date
- 2026-05-15
AI Technical Summary
Existing grassland biomass monitoring methods suffer from low accuracy, are time-consuming and labor-intensive, and cannot achieve large-scale and long-term monitoring. In particular, the estimation accuracy is low in high biomass areas, and remote sensing monitoring does not fully incorporate warming microenvironment parameters, resulting in insufficient ecological interpretability.
Three-dimensional point cloud data of warming grassland plots were collected by UAVs. Ground points were extracted by local density zoning and triangular network iterative optimization methods. Combined with canopy structure parameters and diurnal warming characteristics, a two-branch neural network model was used to predict biomass, thereby achieving high-precision monitoring of warming grassland.
It improves the accuracy and efficiency of grassland biomass monitoring, is applicable to different warming ranges and grassland types, and enables long-term continuous biomass monitoring, overcoming the shortcomings of existing technologies.
Smart Images
Figure CN122042906A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ecological monitoring technology, specifically to a biomass monitoring method based on warming grasslands. Background Technology
[0002] Grassland ecosystems are an important component of terrestrial ecosystems, and biomass monitoring of warming grasslands is a core aspect of global climate change research. Currently, grassland biomass monitoring is mainly divided into two categories: manual measurement methods and remote sensing monitoring methods. Manual measurement methods obtain quadrat biomass through mowing, which is highly accurate but time-consuming, labor-intensive, and destructive, making it unsuitable for large-scale and long-term monitoring. Remote sensing monitoring methods have the advantages of wide coverage and high timeliness, but they often rely on a single vegetation index (such as NDVI) to construct linear models, making them susceptible to vegetation index saturation issues and resulting in lower estimation accuracy in high biomass areas. Furthermore, most remote sensing monitoring does not fully incorporate warming microenvironment parameters, failing to effectively isolate warming-driven biomass change signals, leading to insufficient ecological interpretability of the monitoring results. Summary of the Invention
[0003] To address the above problems, this invention proposes a biomass monitoring method based on warming grasslands.
[0004] The technical solution of this invention is: a biomass monitoring method based on warming grasslands, comprising the following steps: S1. Use the sensors carried by the drone to collect the original three-dimensional point cloud data of the warming grassland sample plot; S2. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data. Filter the standard 3D point cloud data to obtain the final ground point set and the final vegetation point set. S3. Subtract the digital elevation model corresponding to the final ground point set from the digital surface model corresponding to the standard three-dimensional point cloud data to obtain the canopy relative height model of the warming grassland sample plot. S4. Extract the basic canopy structure parameters of the canopy relative height model; S5. Extract warming characteristic parameters from the temperature data collected by temperature sensors deployed in the warming grassland sample plots; S6. Using the basic canopy structure parameters and warming characteristic parameters as the model input set, the final results are determined using the biomass monitoring model.
[0005] In S3, for the ground point set, an inverse distance weighted interpolation method is used to generate a digital elevation model with a resolution of 0.1m×0.1m, which reflects the surface elevation distribution of the warming sample plot.
[0006] Digital surface model: Using all preprocessed point cloud data (ground points and vegetation points), an I interpolation method is used to generate a digital surface model with a resolution of 0.1m×0.1m, which reflects the highest elevation distribution of the grassland vegetation canopy.
[0007] The elevation value (in meters) of the (i,j)th raster in the digital surface model is subtracted from the elevation value (in meters) of the (i,j)th raster in the digital elevation model. Rasteres with relative height ≤ 0 in the canopy relative height model (representing bare land without vegetation cover) are removed, and valid rasteres with relative height > 0 are retained for subsequent canopy structure parameter extraction.
[0008] Furthermore, S2 includes the following sub-steps; S21. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data; S22. Based on the local density of each point in the standard three-dimensional point cloud data, the sample plot is divided into sparse density area, medium density area and dense density area. S23. Filter the point clouds of each density zone to generate several initial ground points for the warming grassland sample plot, forming an initial ground point set; S24. Connect several initial ground points to form several triangles, and adjust the triangles so that there are no other initial ground points inside the circumcircle of the triangle, thus generating a triangulation network; S25. Extract the remaining point clouds from the standard 3D point cloud data except for the initial ground points, and calculate the nearest vertical distance from the remaining point clouds to the triangulation and the local density of the remaining point clouds. S26. Add the point cloud that satisfies the preset conditions for both the nearest vertical distance and local density to the initial ground point set, and update the triangular mesh until all other point clouds have been traversed to obtain the final ground point set. S27. Take the point cloud data other than the final ground point set from the standard three-dimensional point cloud data as the final vegetation point set.
[0009] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, existing ground and vegetation extraction algorithms generally use static threshold extraction. When directly used for filtering point clouds of warming grasslands, the threshold requirements are different for regions with different warming gradients (such as sparse vegetation and dense vegetation). Fixed thresholds will cause vegetation points in high warming areas to be misjudged as ground points, and ground points in low warming areas to be misjudged as vegetation points.
[0010] Furthermore, existing algorithms only determine the vertical distance from a point to the triangulation network, neglecting the local density characteristics of warming grassland point clouds (dense clusters are mostly vegetation, sparse clusters are mostly ground), which can lead to misclassification. Moreover, even with uniform warming, the vegetation growth status (density and height) will differ in different areas within the sample plot, affecting the distinction between ground and vegetation in the point cloud. Therefore, this invention calculates the local density of each point (e.g., the number of points within a 0.2m neighborhood); and divides the sample plot into three zones based on density, for example: sparse zone: density ≤ 8 points / m². 2 (Sparse vegetation, high ground exposure); Medium zone: 8 points / m 2 <Density ≤ 16 points / m 2 (Medium vegetation); Dense areas: density > 16 points / m² 2 (Dense vegetation, little exposed ground). Of course, the density threshold can be adjusted according to the actual situation to ensure that the number of point clouds in each zone is sufficient.
[0011] After filtering the point cloud of each region, the spatial location of each point is identified; the points are connected in pairs to form triangles; the connection method of the triangles is automatically adjusted to ensure that they conform to the rule that there are no other points on the circumcircle; finally, a network of triangles is obtained, which is called a triangular network.
[0012] This invention does not use random initial ground points, but rather combines warming gradient and point cloud elevation statistics to pinpoint real ground points; preset conditions are dynamically adjusted with distance to adapt to vegetation characteristics in different warming zones. A density criterion is added to distinguish between densely vegetated points and sparsely vegetated ground points.
[0013] Furthermore, S23 includes the following sub-steps: S231. Extract the elevation coordinates of all point clouds in the density region to form an elevation sequence, and extract the lower quartile and median of the elevation sequence. S232. Based on the lower quartile and median of the high-order sequence, construct screening criteria for the density region, specifically as follows: ,in, This represents the lower quartile of the high-order sequence. This represents the median of the high-order sequence. Represents the elevation coordinates of the point cloud; S233. Select the point clouds in the density region that meet the screening conditions as candidate point clouds to form a candidate set; S234. Divide each density region into grids, and randomly select a candidate point cloud from each grid to form the initial ground point set.
[0014] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, instead of mixing candidate points from all three zones together, each zone (sparse, medium, and dense) selects its own points. This ensures that initial ground points exist in all three zones, preventing them from being concentrated in only one zone. Elevation thresholds based on quartiles and the median reduce the inclusion of low-lying vegetation points. Raster sampling ensures that initial ground points are evenly distributed within the plot, preventing local fitting bias caused by point clustering during triangulation construction.
[0015] It's important to note that in S234, dense vegetation in high-density areas results in very few candidate points on the ground. Some grid cells may even have no points at all, indicating that the grids are too small. The grid size needs to be increased. The purpose of grid division is to distribute the selected points, not to use a fixed-size grid. If a small grid cell has no points, increase the grid size or merge grids. As long as the final selected points cover different areas of the partition, it's considered a uniform selection.
[0016] Additionally, in S233, the local density of each point in the candidate set can be calculated, retaining points with a local density ≤ 5 points / m. 2 Find the candidate points and update the candidate set.
[0017] Furthermore, in S4, the average canopy height, standard deviation of canopy height, canopy vegetation volume, and canopy density of the canopy relative height model are extracted as basic canopy structure parameters.
[0018] The arithmetic mean of the relative heights of all effective grids within a quadrat reflects the overall height of the vegetation canopy. The standard deviation of the relative heights of all effective grids within a quadrat reflects the uniformity of the vegetation canopy height; differences in canopy height will change significantly with larger temperature increases. The sum of the relative heights of all effective grids within a quadrat multiplied by the area of a single grid reflects the total volume of the vegetation canopy, which is significantly positively correlated with aboveground biomass. The proportion of effective grids to the total number of grids within a quadrat reflects the vegetation cover.
[0019] Furthermore, S5 includes the following sub-steps: S51. Temperature data collected using temperature sensors deployed in warming grassland sample plots; S52. During the warming cycle, extract the average warming amplitude of the warming grassland plots for each day and the average warming amplitude for each night. S53. Take the average of the average temperature increase for all daytime days as the first temperature increase, and take the average of the average temperature increase for all nighttime days as the second temperature increase. S54. Based on the proportion of daytime and nighttime, the first and second temperature increase amplitudes are weighted and fused to obtain the first temperature increase coefficient. S55. Subtract the first heating amplitude from the second heating amplitude, and take the absolute value of the difference to obtain the second heating coefficient; S56. The ratio between the first heating coefficient and the second heating coefficient is used as the heating characteristic parameter.
[0020] The beneficial effect of the above-mentioned further solution is that, in this invention, daytime hours can be 08:00-20:00, and nighttime hours can be 20:00-08:00. Because the lengths of day and night may differ—for example, days are longer in summer and nights are longer in winter—the time selection can be adjusted adaptively.
[0021] Grassland biomass response to nighttime warming is typically stronger than that during the day, and there is a significant difference between daytime and nighttime warming rates during the warming process. Existing technologies often ignore this diurnal difference. This invention quantifies the differential impact of diurnal warming asymmetry on grassland canopy and biomass, aligning with the physiological response patterns of grassland ecosystems, namely, that nighttime respiration is more significantly affected by temperature.
[0022] The daily average temperature increase is obtained by weighting the average daytime / nighttime temperature increase based on the actual duration of day / nighttime. This gives a greater weight to the longer periods of daytime / nighttime in the daily average temperature increase calculation, making the calculation more consistent with the actual daily average temperature changes.
[0023] Furthermore, the biomass monitoring model includes an input layer, a dual-branch module, a multiplier U1, an adder A1, a feature fusion layer, a fully connected layer, and an output layer; the dual-branch module includes a canopy feature branch layer and a warming feature branch layer; The input terminal of the input layer serves as the input terminal of the biomass monitoring model. Its first and second output terminals are respectively connected to the input terminals of the canopy feature branch layer and the warming feature branch layer. The first output terminal of the canopy feature branch layer and the output terminal of the warming feature branch layer are both input to multiplier U1. The second output terminal of the canopy feature branch layer and the first output terminal of multiplier U1 are both input to adder A1. The second output terminal of multiplier U1 and the output terminal of adder A1 are both input to the feature fusion layer. The output terminal of the feature fusion layer, the fully connected layer, and the output layer are connected sequentially.
[0024] The beneficial effect of the above-mentioned further scheme is that, in this invention, the input layer normalizes the basic canopy structure parameters and warming characteristic parameters, and splices them into a canopy branch vector, thereby converting the warming characteristic parameters into a warming branch vector.
[0025] The input to the canopy branch is four original parameters, four-dimensional in total. A 4x8 weight matrix is assigned to it, and through matrix multiplication and an activation function, the four-dimensional input is mapped to an eight-dimensional output. Multiplying these two results in a 1x8 vector. Adding a bias term (eight-dimensional), and then applying ReLU activation, the final output is an eight-dimensional canopy feature.
[0026] The input to the warming branch is a single original parameter, which is 1-dimensional. Similarly, it is assigned a weight matrix of 1 row × 8 columns, and then mapped to an 8-dimensional output through matrix multiplication. With the addition of a bias term (8-dimensional), the output after activation is an 8-dimensional warming feature.
[0027] The canopy feature branch layer uses the ReLU function corresponding to the convolution kernel to extract features from the canopy branch vector, while the warming feature branch layer uses the ReLU function corresponding to the convolution kernel and a lightweight LSTM to extract features, making the canopy branch features and the warming feature branch have the same dimension.
[0028] The fully connected layer transforms the 16-dimensional features obtained by the feature fusion layer into more abstract 64-dimensional hidden features through a learnable weight matrix, while introducing non-linearity with ReLU activation.
[0029] The output layer maps the features back to biomass, transforming the 64-dimensional hidden features into 1-dimensional aboveground biomass (regression task). Linear activation is used because biomass is a continuous value.
[0030] Multiplier U1 is used to perform nonlinear modulation of the canopy by heating, outputting 8-dimensional coupled features. Adder is used to compensate for the annihilation of weak features that may be caused by multiplicative coupling, preserving the basic features of the canopy. Feature fusion layers are then stitched together to output 16-dimensional fused features.
[0031] Furthermore, the expression for adder A1 is: ; ; in, This represents the output of adder A1. This represents the output of the canopy feature branch layer. This represents the output of the heating feature branch layer. This represents the output of multiplier U1. This represents the compensation coefficient.
[0032] The beneficial effect of the above-mentioned further scheme is that, in this invention, the effect of warming on the canopy is not a linear superposition (for example, an increase of 1°C in warming does not necessarily result in a fixed increase in canopy height), but rather a non-linear correlation. Therefore, if only warming features and canopy features are concatenated (for example, combining 4-dimensional canopy features with 1-dimensional warming features to form 5-dimensional features), the model can only learn the independent effects of the two, and cannot learn the interaction between warming and canopy modulation; while multiplicative coupling can directly quantify this non-linear correlation. The operation is based on Hadamard product (element-by-element multiplication), and the output dimensions of both branches are unified to 8 dimensions.
[0033] Furthermore, if the warming rate in a certain area is extremely low, the result after multiplicative coupling will approach zero, at which point the model will lose the core information of the canopy features. The additive residual can recover the original canopy features, ensuring that the model can still predict biomass based on the canopy features. The compensation coefficient controls the weight of the increment to avoid overshadowing the basic contribution, and can be set manually or dynamically adjusted.
[0034] The beneficial effects of this invention are as follows: This invention proposes an iterative optimization method based on a local density partitioning triangular network to improve the extraction accuracy of ground points and solve the problem of insufficient ground points in dense areas. It also sets adaptive thresholds for different density partitions. Furthermore, this invention constructs feature sets of multiple canopy structure parameters and diurnal warming asymmetry features to extract warming and canopy features. Finally, this invention designs a neural network based on a fusion of bi-branch, multiplicative, and additive methods to complete the nonlinear interaction between warming and canopy, improving biomass prediction accuracy, which differs from conventional linear single-branch models. This invention addresses the spatial heterogeneity of artificially warmed grasslands and the diurnal heating differences of warming equipment, making it suitable for biomass monitoring of different warming amplitudes and grassland types, enabling long-term and continuous biomass monitoring. Attached Figure Description
[0035] Figure 1 A flowchart of a biomass monitoring method based on warming grasslands; Figure 2 This is a schematic diagram of the biomass monitoring model. Detailed Implementation
[0036] The embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0037] like Figure 1 As shown, this invention provides a biomass monitoring method based on warming grasslands, comprising the following steps: S1. Use the sensors carried by the drone to collect the original three-dimensional point cloud data of the warming grassland sample plot; S2. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data. Filter the standard 3D point cloud data to obtain the final ground point set and the final vegetation point set. S3. Subtract the digital elevation model corresponding to the final ground point set from the digital surface model corresponding to the standard three-dimensional point cloud data to obtain the canopy relative height model of the warming grassland sample plot. S4. Extract the basic canopy structure parameters of the canopy relative height model; S5. Extract warming characteristic parameters from the temperature data collected by temperature sensors deployed in the warming grassland sample plots; S6. Using the basic canopy structure parameters and warming characteristic parameters as the model input set, the final results are determined using the biomass monitoring model.
[0038] In S3, for the ground point set, an inverse distance weighted interpolation method is used to generate a digital elevation model with a resolution of 0.1m×0.1m, which reflects the surface elevation distribution of the warming sample plot.
[0039] Digital surface model: Using all preprocessed point cloud data (ground points and vegetation points), an I interpolation method is used to generate a digital surface model with a resolution of 0.1m×0.1m, which reflects the highest elevation distribution of the grassland vegetation canopy.
[0040] The elevation value (in meters) of the (i,j)th raster in the digital surface model is subtracted from the elevation value (in meters) of the (i,j)th raster in the digital elevation model. Rasteres with relative height ≤ 0 in the canopy relative height model (representing bare land without vegetation cover) are removed, and valid rasteres with relative height > 0 are retained for subsequent canopy structure parameter extraction.
[0041] In this embodiment of the invention, S2 includes the following sub-steps; S21. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data; S22. Based on the local density of each point in the standard three-dimensional point cloud data, the sample plot is divided into sparse density area, medium density area and dense density area. S23. Filter the point clouds of each density zone to generate several initial ground points for the warming grassland sample plot, forming an initial ground point set; S24. Connect several initial ground points to form several triangles, and adjust the triangles so that there are no other initial ground points inside the circumcircle of the triangle, thus generating a triangulation network; S25. Extract the remaining point clouds from the standard 3D point cloud data except for the initial ground points, and calculate the nearest vertical distance from the remaining point clouds to the triangulation and the local density of the remaining point clouds. S26. Add the point cloud that satisfies the preset conditions for both the nearest vertical distance and local density to the initial ground point set, and update the triangular mesh until all other point clouds have been traversed to obtain the final ground point set. S27. Take the point cloud data other than the final ground point set from the standard three-dimensional point cloud data as the final vegetation point set.
[0042] In this invention, existing ground and vegetation extraction algorithms generally use static threshold extraction. When directly used for filtering point clouds of warming grasslands, the threshold requirements differ for regions with different warming gradients (such as sparse vegetation and dense vegetation). A fixed threshold will cause vegetation points in high warming areas to be misidentified as ground points, and ground points in low warming areas to be misidentified as vegetation points.
[0043] Furthermore, existing algorithms only determine the vertical distance from a point to the triangulation network, neglecting the local density characteristics of warming grassland point clouds (dense clusters are mostly vegetation, sparse clusters are mostly ground), which can lead to misclassification. Moreover, even with uniform warming, the vegetation growth status (density and height) will differ in different areas within the sample plot, affecting the distinction between ground and vegetation in the point cloud. Therefore, this invention calculates the local density of each point (e.g., the number of points within a 0.2m neighborhood); and divides the sample plot into three zones based on density, for example: sparse zone: density ≤ 8 points / m². 2 (Sparse vegetation, high ground exposure); Medium zone: 8 points / m 2 <Density ≤ 16 points / m 2 (Medium vegetation); Dense areas: density > 16 points / m² 2 (Dense vegetation, little exposed ground). Of course, the density threshold can be adjusted according to the actual situation to ensure that the number of point clouds in each zone is sufficient.
[0044] After filtering the point cloud of each region, the spatial location of each point is identified; the points are connected in pairs to form triangles; the connection method of the triangles is automatically adjusted to ensure that they conform to the rule that there are no other points on the circumcircle; finally, a network of triangles is obtained, which is called a triangular network.
[0045] This invention does not use random initial ground points, but rather combines warming gradient and point cloud elevation statistics to pinpoint real ground points; preset conditions are dynamically adjusted with distance to adapt to vegetation characteristics in different warming zones. A density criterion is added to distinguish between densely vegetated points and sparsely vegetated ground points.
[0046] In this embodiment of the invention, S23 includes the following sub-steps: S231. Extract the elevation coordinates of all point clouds in the density region to form an elevation sequence, and extract the lower quartile and median of the elevation sequence. S232. Based on the lower quartile and median of the high-order sequence, construct screening criteria for the density region, specifically as follows: ,in, This represents the lower quartile of the high-order sequence. This represents the median of the high-order sequence. Represents the elevation coordinates of the point cloud; S233. Select the point clouds in the density region that meet the screening conditions as candidate point clouds to form a candidate set; S234. Divide each density region into grids, and randomly select a candidate point cloud from each grid to form the initial ground point set.
[0047] In this invention, instead of mixing candidate points from all three zones together, each zone (sparse, medium, and dense) selects its own points. This ensures that initial ground points exist in all three zones, preventing them from concentrating in only one zone. Elevation thresholds based on quartiles and the median are used to reduce the inclusion of low-lying vegetation points. Raster sampling ensures that initial ground points are evenly distributed within the plot, preventing local fitting bias caused by point clustering during triangulation.
[0048] It's important to note that in S234, dense vegetation in high-density areas results in very few candidate points on the ground. Some grid cells may even have no points at all, indicating that the grids are too small. The grid size needs to be increased. The purpose of grid division is to distribute the selected points, not to use a fixed-size grid. If a small grid cell has no points, increase the grid size or merge grids. As long as the final selected points cover different areas of the partition, it's considered a uniform selection.
[0049] Additionally, in S233, the local density of each point in the candidate set can be calculated, retaining points with a local density ≤ 5 points / m. 2 Find the candidate points and update the candidate set.
[0050] In this embodiment of the invention, in S4, the average canopy height, standard deviation of canopy height, canopy vegetation volume, and canopy density of the canopy relative height model are extracted as basic canopy structure parameters.
[0051] The arithmetic mean of the relative heights of all effective grids within a quadrat reflects the overall height of the vegetation canopy. The standard deviation of the relative heights of all effective grids within a quadrat reflects the uniformity of the vegetation canopy height; differences in canopy height will change significantly with larger temperature increases. The sum of the relative heights of all effective grids within a quadrat multiplied by the area of a single grid reflects the total volume of the vegetation canopy, which is significantly positively correlated with aboveground biomass. The proportion of effective grids to the total number of grids within a quadrat reflects the vegetation cover.
[0052] In this embodiment of the invention, S5 includes the following sub-steps: S51. Temperature data collected using temperature sensors deployed in warming grassland sample plots; S52. During the warming cycle, extract the average warming amplitude of the warming grassland plots for each day and the average warming amplitude for each night. S53. Take the average of the average temperature increase for all daytime days as the first temperature increase, and take the average of the average temperature increase for all nighttime days as the second temperature increase. S54. Based on the proportion of daytime and nighttime, the first and second temperature increase amplitudes are weighted and fused to obtain the first temperature increase coefficient. S55. Subtract the first heating amplitude from the second heating amplitude, and take the absolute value of the difference to obtain the second heating coefficient; S56. The ratio between the first heating coefficient and the second heating coefficient is used as the heating characteristic parameter.
[0053] In this invention, daytime hours can be 08:00-20:00, and nighttime hours can be 20:00-08:00. Because the lengths of day and night may differ—for example, days are longer in summer and nights are longer in winter—the time selection can be adjusted accordingly.
[0054] Grassland biomass response to nighttime warming is typically stronger than that during the day, and there is a significant difference between daytime and nighttime warming rates during the warming process. Existing technologies often ignore this diurnal difference. This invention quantifies the differential impact of diurnal warming asymmetry on grassland canopy and biomass, aligning with the physiological response patterns of grassland ecosystems, namely, that nighttime respiration is more significantly affected by temperature.
[0055] The daily average temperature increase is obtained by weighting the average daytime / nighttime temperature increase based on the actual duration of day / nighttime. This gives a greater weight to the longer periods of daytime / nighttime in the daily average temperature increase calculation, making the calculation more consistent with the actual daily average temperature changes.
[0056] In S42, the temperature increase is the difference between the actual temperature of the temperature-increasing zone and the natural temperature of the control zone. For example, all raw temperature data collected by the sensor in the temperature-increasing zone during the daytime period (08:00-20:00) and all raw temperature data collected by the sensor in the control zone during the same daytime period are used to obtain two sets. Outliers in both sets are removed, and the arithmetic mean of the filtered temperature-increasing zone data is taken to obtain the average daytime temperature of the temperature-increasing zone on day m. Similarly, the average daytime temperature of the control zone on day m is obtained. Subtracting the average daytime temperature of the control zone from the average daytime temperature of the temperature-increasing zone gives the average daytime temperature increase on day m.
[0057] In embodiments of the present invention, such as Figure 2 As shown, the biomass monitoring model includes an input layer, a dual-branch module, a multiplier U1, an adder A1, a feature fusion layer, a fully connected layer, and an output layer; the dual-branch module includes a canopy feature branch layer and a warming feature branch layer. The input terminal of the input layer serves as the input terminal of the biomass monitoring model. Its first and second output terminals are respectively connected to the input terminals of the canopy feature branch layer and the warming feature branch layer. The first output terminal of the canopy feature branch layer and the output terminal of the warming feature branch layer are both input to multiplier U1. The second output terminal of the canopy feature branch layer and the first output terminal of multiplier U1 are both input to adder A1. The second output terminal of multiplier U1 and the output terminal of adder A1 are both input to the feature fusion layer. The output terminal of the feature fusion layer, the fully connected layer, and the output layer are connected sequentially.
[0058] In this invention, the input layer normalizes the basic canopy structure parameters and warming characteristic parameters, and concatenates them into a canopy branch vector, thus converting the warming characteristic parameters into a warming branch vector.
[0059] The input to the canopy branch is four original parameters, four-dimensional in total. A 4x8 weight matrix is assigned to it, and through matrix multiplication and an activation function, the four-dimensional input is mapped to an eight-dimensional output. Multiplying these two results in a 1x8 vector. Adding a bias term (eight-dimensional), and then applying ReLU activation, the final output is an eight-dimensional canopy feature.
[0060] The input to the warming branch is a single original parameter, which is 1-dimensional. Similarly, it is assigned a weight matrix of 1 row × 8 columns, and then mapped to an 8-dimensional output through matrix multiplication. With the addition of a bias term (8-dimensional), the output after activation is an 8-dimensional warming feature.
[0061] The canopy feature branch layer uses the ReLU function corresponding to the convolution kernel to extract features from the canopy branch vector, while the warming feature branch layer uses the ReLU function corresponding to the convolution kernel and a lightweight LSTM to extract features, making the canopy branch features and the warming feature branch have the same dimension.
[0062] The fully connected layer transforms the 16-dimensional features obtained by the feature fusion layer into more abstract 64-dimensional hidden features through a learnable weight matrix, while introducing non-linearity with ReLU activation.
[0063] The output layer maps the features back to biomass, transforming the 64-dimensional hidden features into 1-dimensional aboveground biomass (regression task). Linear activation is used because biomass is a continuous value.
[0064] Multiplier U1 is used to perform nonlinear modulation of the canopy by heating, outputting 8-dimensional coupled features. Adder is used to compensate for the annihilation of weak features that may be caused by multiplicative coupling, preserving the basic features of the canopy. Feature fusion layers are then stitched together to output 16-dimensional fused features.
[0065] In this embodiment of the invention, the expression for adder A1 is: ; ; in, This represents the output of adder A1. This represents the output of the canopy feature branch layer. This represents the output of the heating feature branch layer. This represents the output of multiplier U1. This represents the compensation coefficient.
[0066] In this invention, the effect of warming on the canopy is not a linear superposition (e.g., a 1°C increase in temperature does not necessarily result in a fixed increase in canopy height), but rather a non-linear correlation. Therefore, if only warming features and canopy features are concatenated (e.g., combining 4-dimensional canopy features with 1-dimensional warming features to form 5-dimensional features), the model can only learn the independent effects of the two, and cannot learn the interaction between warming and canopy modulation; however, multiplicative coupling can directly quantify this non-linear correlation. The calculation is based on Hadamard product (element-by-element multiplication), and the output dimensions of both branches are unified to 8 dimensions.
[0067] Furthermore, if the warming rate in a certain area is extremely low, the result after multiplicative coupling will approach zero, at which point the model will lose the core information of the canopy features. The additive residual can recover the original canopy features, ensuring that the model can still predict biomass based on the canopy features. The compensation coefficient controls the weight of the increment to avoid overshadowing the basic contribution, and can be set manually or dynamically adjusted.
[0068] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A biomass monitoring method based on warming grasslands, characterized in that, Includes the following steps: S1. Use the sensors carried by the drone to collect the original three-dimensional point cloud data of the warming grassland sample plot; S2. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data. Filter the standard 3D point cloud data to obtain the final ground point set and the final vegetation point set. S3. Subtract the digital elevation model corresponding to the final ground point set from the digital surface model corresponding to the standard three-dimensional point cloud data to obtain the canopy relative height model of the warming grassland sample plot. S4. Extract the basic canopy structure parameters of the canopy relative height model; S5. Extract warming characteristic parameters from the temperature data collected by temperature sensors deployed in the warming grassland sample plots; S6. Using the basic canopy structure parameters and warming characteristic parameters as the model input set, the final results are determined using the biomass monitoring model.
2. The biomass monitoring method based on warming grassland according to claim 1, characterized in that, S2 includes the following sub-steps; S21. Preprocess the original 3D point cloud data to obtain standard 3D point cloud data; S22. Based on the local density of each point in the standard three-dimensional point cloud data, the sample plot is divided into sparse density area, medium density area and dense density area. S23. Filter the point clouds of each density zone to generate several initial ground points for the warming grassland sample plot, forming an initial ground point set; S24. Connect several initial ground points to form several triangles, and adjust the triangles so that there are no other initial ground points inside the circumcircle of the triangle, thus generating a triangulation network; S25. Extract the remaining point clouds from the standard 3D point cloud data except for the initial ground points, and calculate the nearest vertical distance from the remaining point clouds to the triangulation and the local density of the remaining point clouds. S26. Add the point cloud that satisfies the preset conditions for both the nearest vertical distance and local density to the initial ground point set, and update the triangular mesh until all other point clouds have been traversed to obtain the final ground point set. S27. Take the point cloud data other than the final ground point set from the standard three-dimensional point cloud data as the final vegetation point set.
3. The biomass monitoring method based on warming grassland according to claim 2, characterized in that, S23 includes the following sub-steps: S231. Extract the elevation coordinates of all point clouds in the density region to form an elevation sequence, and extract the lower quartile and median of the elevation sequence. S232. Based on the lower quartile and median of the high-order sequence, construct screening criteria for the density region, specifically as follows: ,in, This represents the lower quartile of the high-order sequence. This represents the median of the high-order sequence. Represents the elevation coordinates of the point cloud; S233. Select the point clouds in the density region that meet the screening conditions as candidate point clouds to form a candidate set; S234. Divide each density region into grids, and randomly select a candidate point cloud from each grid to form the initial ground point set.
4. The biomass monitoring method based on warming grassland according to claim 1, characterized in that, In S4, the average canopy height, standard deviation of canopy height, canopy vegetation volume, and canopy density of the canopy relative height model are extracted as basic canopy structure parameters.
5. The biomass monitoring method based on warming grassland according to claim 1, characterized in that, S5 includes the following sub-steps: S51. Temperature data collected using temperature sensors deployed in warming grassland sample plots; S52. During the warming cycle, extract the average warming amplitude of the warming grassland plots for each day and the average warming amplitude for each night. S53. Take the average of the average temperature increase for all daytime days as the first temperature increase, and take the average of the average temperature increase for all nighttime days as the second temperature increase. S54. Based on the proportion of daytime and nighttime, the first and second temperature increase amplitudes are weighted and fused to obtain the first temperature increase coefficient. S55. Subtract the first heating amplitude from the second heating amplitude, and take the absolute value of the difference to obtain the second heating coefficient; S56. The ratio between the first heating coefficient and the second heating coefficient is used as the heating characteristic parameter.
6. The biomass monitoring method based on warming grassland according to claim 1, characterized in that, The biomass monitoring model includes an input layer, a dual-branch module, a multiplier U1, an adder A1, a feature fusion layer, a fully connected layer, and an output layer; the dual-branch module includes a canopy feature branch layer and a warming feature branch layer; The input terminal of the input layer serves as the input terminal of the biomass monitoring model, and its first and second output terminals are respectively connected to the input terminals of the canopy feature branch layer and the warming feature branch layer. The first output terminal of the canopy feature branch layer and the output terminal of the warming feature branch layer are both input to multiplier U1. The second output terminal of the canopy feature branch layer and the first output terminal of multiplier U1 are both input to adder A1. The second output terminal of multiplier U1 and the output terminal of adder A1 are both input to the feature fusion layer. The output of the feature fusion layer, the fully connected layer, and the output layer are connected in sequence.
7. The biomass monitoring method based on warming grassland according to claim 6, characterized in that, The expression for adder A1 is: ； ； in, This represents the output of adder A1. This represents the output of the canopy feature branch layer. This represents the output of the heating feature branch layer. This represents the output of multiplier U1. This represents the compensation coefficient.