Ecological vegetation coverage rate measuring and calculating method, device and equipment based on unmanned aerial vehicle and medium

Abstract

The invention relates to an ecological vegetation coverage rate measuring and calculating method and device based on an unmanned aerial vehicle, equipment and a medium. The method comprises the following steps: acquiring a multispectral image and three-dimensional point cloud data of a to-be-detected area acquired by an unmanned aerial vehicle; performing space-time registration and processing on the multispectral image and the three-dimensional point cloud data to generate a high-resolution digital orthoimage and a canopy height model; carrying out stacking fusion on the spectral band of the high-resolution digital orthoimage and the height data of the canopy height model at a pixel level to generate a multi-channel input tensor; and segmenting the multi-channel input tensor by using a deep learning semantic segmentation model to generate a reference vegetation coverage rate true value map of the to-be-detected region. By adopting the method, the vertical structure information of the ground features can be introduced into semantic segmentation, the ground features with similar spectrums but different heights are effectively distinguished, and the accuracy and robustness of vegetation extraction are improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology application, and in particular relates to a method, device, equipment and medium for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the increasing demand for high-precision data in the fields of ecological environment monitoring and geographic information systems (GIS), unmanned aerial vehicle (UAV) remote sensing technology has emerged. UAV remote sensing technology, with its advantages of high spatiotemporal resolution, high flexibility, and low cost, has become an important means of acquiring high-precision land surface information, leading to the widely used traditional method of local fine-grained measurement of vegetation cover (FVC) based on UAVs.

[0003] In traditional techniques, many methods rely on visible light (RGB) or multispectral imagery acquired by UAVs and employ conventional image processing techniques for segmentation. For example, digital orthophotos (DOM) are converted to the HSV color space and Otsu's thresholding method is used to segment the H component to extract vegetation. When LiDAR or SfM point cloud data is acquired simultaneously, traditional methods perform "terrain correction" on the remote sensing images or use the extracted 3D parameters as one of many input features for downstream applications.

[0004] However, methods based on RGB or HSV values ​​heavily rely on instantaneous sunlight conditions. Shadows (whether cloud shadows or shadows cast by the vegetation canopy itself) can cause drastic changes in the spectral characteristics of healthy vegetation. Traditional thresholding methods such as Otsu's cannot achieve clear separation in complex environments where vegetation, shadows, and bare soil features highly overlap. This prevents the provision of reliable "ground truth" data for subsequent FVC calculations. Furthermore, current methods fail to effectively integrate 3D structural data (such as canopy height models, CHM) with 2D spectral data during the FVC segmentation stage. The fundamental difficulties of FVC segmentation (such as distinguishing between tall canopies and green bare soil) could have been solved using CHM, but the misuse or functional degradation of existing technologies completely wastes the core advantages of high-value data such as LiDAR. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) to address the aforementioned technical problems.

[0006] Firstly, this application provides a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs), including:

[0007] Acquire multispectral images and 3D point cloud data of the area to be tested, collected by the UAV platform;

[0008] Spatiotemporal registration is performed on the multispectral image to obtain the spatiotemporally registered multispectral image. Orthophoto processing is then performed on the spatiotemporally registered multispectral image to generate a high-resolution digital orthophoto.

[0009] Process the 3D point cloud data to generate a canopy height model;

[0010] High-resolution digital orthophotos and height data from canopy height models are stacked and fused at the pixel level to generate multi-channel input tensors.

[0011] A deep learning semantic segmentation model is used to segment the multi-channel input tensor to generate a baseline vegetation coverage map of the test area.

[0012] In one embodiment, the three-dimensional point cloud data is processed to generate a canopy height model, including:

[0013] The 3D point cloud data is separated into ground points and non-ground points using a ground point filtering algorithm;

[0014] A digital elevation model is generated by interpolating ground points.

[0015] A digital surface model is generated by interpolating non-ground points.

[0016] The true height of ground features is calculated by subtracting the digital surface model from the digital elevation model pixel by pixel on the raster surface. This true height is then used to construct a canopy height model. The true height of ground features is calculated using the following formula:

[0017]

[0018] in, For pixels The actual height of vegetation or ground features; For pixels The altitude at the highest point on the Earth's surface; For pixels The altitude of the exposed surface.

[0019] In one embodiment, the training steps of the deep learning semantic segmentation model include:

[0020] For each pixel in the training set, obtain the corresponding canopy height value from the canopy height model;

[0021] Calculate the height weighting factor based on the canopy height value using the following formula:

[0022]

[0023] in, Highly weighted factors; In pixels; This represents the canopy height value corresponding to each pixel. The maximum height value in the training set; and To adjust the hyperparameters affecting height;

[0024] Calculate the height-weighted mixed loss function based on the height weighting factor using the following formula:

[0025]

[0026]

[0027] in, It is a highly weighted mixed loss function; and These are the weighting coefficients; For the Focalloss function; An improved Diceloss function; For pixels The true label; For the model to pixels The probability of it being predicted as vegetation; It is a very small smoothing constant;

[0028] By minimizing the highly weighted mixed loss function through backpropagation, the parameters of the deep learning semantic segmentation model are optimized, resulting in a trained deep learning semantic segmentation model.

[0029] In one embodiment, a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) further includes:

[0030] Acquire satellite imagery covering the area to be measured;

[0031] Feature extraction is performed on satellite imagery to obtain satellite feature vectors that include texture and spectral features;

[0032] The baseline vegetation cover map is aggregated to the pixel scale of satellite imagery using the spatial averaging method to obtain the aggregated training labels.

[0033] Based on satellite feature vectors and aggregated training labels, a nonlinear machine learning regression model is trained to obtain the trained nonlinear machine learning regression model.

[0034] The trained nonlinear machine learning regression model is applied to the satellite image features of the target area to generate a large-scale thematic map of vegetation coverage in the target area.

[0035] Furthermore, feature extraction is performed on the satellite imagery to obtain satellite feature vectors that include texture and spectral features, including:

[0036] The gray-level co-occurrence matrix is ​​calculated within a preset sliding window; wherein, the sliding window is set on the PCA1 image of the satellite imagery;

[0037] Second-order statistical texture features are extracted from the gray-level co-occurrence matrix to obtain texture features; texture features include entropy, contrast, and homogeneity.

[0038] The original spectral bands of the satellite imagery are extracted, and the vegetation index is calculated to obtain the spectral characteristics.

[0039] By combining texture features and spectral features, satellite feature vectors are generated.

[0040] Furthermore, the baseline vegetation cover map is aggregated to the pixel scale of the satellite imagery using spatial averaging to obtain the aggregated training labels, including:

[0041] Acquire the pixel grid of satellite imagery;

[0042] Spatially align the baseline vegetation cover map with the pixel grid to obtain the spatially aligned baseline vegetation cover map.

[0043] The average value of all pixels in the spatially aligned baseline vegetation cover ground truth map is calculated to obtain the average value corresponding to each pixel, which is used as the aggregated training label.

[0044] Furthermore, based on satellite feature vectors and aggregated training labels, a nonlinear machine learning regression model is trained to obtain the trained nonlinear machine learning regression model, including:

[0045] Satellite feature vectors are used as training features, and the aggregated training labels are input into a nonlinear machine learning regression model.

[0046] The training process of a nonlinear machine learning regression model is performed to fit the nonlinear mapping relationship between training features and training labels, thereby obtaining the nonlinear mapping relationship;

[0047] Based on the nonlinear mapping relationship, the output is a nonlinear machine learning regression model that fits the nonlinear mapping relationship, which is then used as the trained nonlinear machine learning regression model.

[0048] Secondly, this application also provides an ecological vegetation coverage measurement device based on unmanned aerial vehicles (UAVs), comprising:

[0049] The acquisition module is used to acquire multispectral images and 3D point cloud data of the area to be tested collected by the UAV platform;

[0050] The digital orthophoto generation module is used to perform spatiotemporal registration on multispectral images to obtain spatiotemporally registered multispectral images, and to perform orthophoto processing on the spatiotemporally registered multispectral images to generate high-resolution digital orthophotos.

[0051] The canopy height model generation module is used to process 3D point cloud data and generate a canopy height model.

[0052] The fusion module is used to stack and fuse high-resolution digital orthophotos and height data from the canopy height model at the pixel level to generate a multi-channel input tensor.

[0053] The baseline coverage module is used to segment the multi-channel input tensor using a deep learning semantic segmentation model to generate a baseline vegetation coverage map of the area to be tested.

[0054] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform any of the above steps.

[0055] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs any of the above steps.

[0056] The aforementioned method, apparatus, equipment, and medium for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) acquire multispectral images and 3D point cloud data collected by UAVs, generating high-resolution digital orthophotos and canopy height models. These are then stacked and fused at the pixel level to generate a multi-channel input tensor, incorporating the vertical structure information of ground features into the semantic segmentation process. This fusion method utilizes deep learning models to simultaneously extract the spectral and 3D structural features of ground features, reducing the adverse effects of changes in illumination conditions and shading on vegetation extraction accuracy when relying solely on 2D spectral information. Segmenting the fused multi-channel data using a deep learning semantic segmentation model effectively distinguishes ground features with similar spectral characteristics but different heights, thereby improving the accuracy and robustness of generating a baseline vegetation coverage ground map for the measured area. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of the present invention;

[0059] Figure 2 This is a flowchart illustrating the process of a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) in one embodiment of the present invention.

[0060] Figure 3 This is a structural block diagram of an ecological vegetation coverage measurement device based on a drone in one embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] The method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) provided in this application can be applied to, for example... Figure 1 In the application environment shown, the drone 101 communicates with the processing device 102 via wireless image transmission or a wired connection. The satellite 103 communicates with the processing device 102 via a wireless network. The data storage system can be integrated into the processing device 102 or placed in the cloud or on other network servers. The processing device 102 can be, but is not limited to, various personal computers, laptops, etc.

[0063] In one exemplary embodiment, such as Figure 2 As shown, a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) is provided, and this method is applied to... Figure 1 Taking the processing equipment in the middle as an example, it can include:

[0064] Step 201: Acquire multispectral images and 3D point cloud data of the area to be tested collected by the UAV platform.

[0065] Multispectral imagery, captured by a multispectral camera mounted on a drone, contains reflectance information of the target area in different spectral bands, typically including spectral data in the red, green, blue, and near-infrared bands. The multispectral image data is stored as a two-dimensional image matrix, recording the spectral reflectance characteristics of the land cover. Three-dimensional point cloud data is obtained by scanning with a lidar sensor mounted on the drone, or calculated using photogrammetric reconstruction algorithms on highly overlapping multispectral images. The three-dimensional point cloud data contains a massive number of discrete spatial coordinate points, each recording the X, Y, and Z coordinates of the ground surface in a Cartesian coordinate system, as well as reflectance intensity information. The processing equipment receives raw observation data from the drone platform via wired or wireless data transmission interfaces, storing the received multispectral imagery and three-dimensional point cloud data in internal memory or a connected database as the basic data source for subsequent processing.

[0066] Step 202: Perform spatiotemporal registration on the multispectral image to obtain a spatiotemporally registered multispectral image. Perform orthophoto processing on the spatiotemporally registered multispectral image to generate a high-resolution digital orthophoto image.

[0067] Digital orthophotos contain geocoded information, with the coordinates of any point on the image corresponding to its real-world coordinates. They also include radiance or reflectance values ​​from multiple spectral channels, including red, green, blue, and near-infrared. The processing equipment reads the multispectral image and its associated inertial navigation system and GPS data, resolving the exterior orientation elements at the time of image capture, including spatial coordinates and attitude angles. Using these exterior orientation elements and camera intrinsic parameters, the equipment performs geometric correction on the multispectral image, eliminating geometric distortions caused by sensor imaging mechanisms, terrain undulations, and changes in flight attitude. The corrected multispectral image is then projected onto a unified geographic coordinate system, completing spatiotemporal registration to obtain a spatiotemporally registered multispectral image. Based on this registration, the equipment executes an orthorectification algorithm, using a digital elevation model to perform differential correction on the image, eliminating projection differences and converting the centrally projected perspective image into an orthophoto image. Finally, the equipment mosaics, evens out colors, and crops multiple orthophotos to generate a high-resolution digital orthophoto covering the entire area under test.

[0068] Step 203: Process the 3D point cloud data to generate a canopy height model.

[0069] In the canopy height model, each pixel value represents the vertical height of a feature relative to the ground at its corresponding location. This eliminates the influence of terrain undulations and directly represents the physical height information of the vegetation canopy. The processing device reads discrete 3D point cloud data and uses a point cloud filtering algorithm to classify the data points. Specifically, the processing device analyzes the spatial geometric features of the point cloud and divides the point cloud set into a ground point set and a non-ground point set. The ground point set represents the spatial location of exposed ground surfaces, while the non-ground point set includes the spatial locations of ground objects such as vegetation canopies and buildings. The processing device constructs continuous digital elevation models and digital surface models, and performs pixel-by-pixel subtraction operations on the digital surface model and digital elevation model at the raster level. The resulting difference matrix is ​​the canopy height model.

[0070] Step 204: Stack and fuse the height data of the high-resolution digital orthophoto and the canopy height model at the pixel level to generate a multi-channel input tensor.

[0071] The multi-channel input tensor simultaneously contains the spectral reflectance and vertical structural features of the region under test. The processing device first checks whether the spatial resolution and coordinate system of the high-resolution digital orthophoto and the canopy height model are consistent. If inconsistent, the device performs a resampling operation to adjust the spatial resolution of the canopy height model to match that of the digital orthophoto, ensuring that their raster grids are strictly aligned in geographic space. The device extracts data matrices for red, green, blue, and near-infrared spectral bands from the digital orthophoto, and simultaneously extracts a data matrix for the height band from the canopy height model. The device then stitches these two-dimensional data matrices along the channel dimension. For example, four spectral band matrices and one height band matrix are stacked in the depth direction to construct a three-dimensional array structure with five channels. The device then normalizes or standardizes this three-dimensional array, mapping spectral and height values ​​of different dimensions to a unified numerical range, eliminating the impact of data magnitude differences on subsequent calculations. The data structure after stacking and standardization is the multi-channel input tensor.

[0072] Step 205: Use a deep learning semantic segmentation model to segment the multi-channel input tensor and generate a baseline vegetation coverage ground map of the area to be tested.

[0073] The baseline vegetation cover map accurately identifies the land cover attribute of each pixel within the test area, providing high-precision spatial distribution information of vegetation for subsequent vegetation cover calculations or as training labels. The deep learning semantic segmentation model has the ability to extract features and classify pixels at the pixel level from multi-channel data. The processing device feeds the generated multi-channel input tensor into the deep learning semantic segmentation model. The model's internal convolutional layers, pooling layers, and activation functions perform layer-by-layer operations on the input tensor, extracting a deep feature map that integrates spectral and height information. The model's decoder restores the deep feature map to the original image size and outputs a probability map indicating whether each pixel belongs to the vegetation or non-vegetation category. The processing device converts the probability map into a binary segmentation mask based on a preset probability threshold or the maximum probability principle. In the segmentation mask, a specific value (e.g., 1) represents a vegetation pixel, and another value (e.g., 0) represents a non-vegetation pixel. The processing device outputs this binary segmentation mask as the baseline vegetation cover map.

[0074] The UAV-based method for calculating ecological vegetation coverage provided in this embodiment acquires multispectral images and 3D point cloud data collected by the UAV, generates high-resolution digital orthophotos and a canopy height model, and stacks and fuses these two at the pixel level to generate a multi-channel input tensor, incorporating the vertical structure information of ground features into the semantic segmentation process. This fusion method can utilize a deep learning model to simultaneously extract the spectral and 3D structural features of ground features, reducing the adverse effects of changes in illumination conditions and shading on the accuracy of vegetation extraction when relying solely on 2D spectral information. By segmenting the fused multi-channel data using a deep learning semantic segmentation model, ground features with similar spectral features but different heights can be effectively distinguished, thereby improving the accuracy and robustness of generating a baseline vegetation coverage ground map for the area to be measured.

[0075] In one embodiment, processing three-dimensional point cloud data to generate a canopy height model may include:

[0076] Step 301: Separate the 3D point cloud data into ground points and non-ground points using a ground point filtering algorithm.

[0077] The processing device loads point cloud filtering algorithms, such as a cloth simulation filtering algorithm or a progressive irregular triangular mesh encryption algorithm. In the cloth simulation filtering algorithm, the processing device flips the point cloud, assuming a piece of cloth is placed on the flipped point cloud surface under gravity. Ground points are determined by analyzing the interaction positions between the cloth nodes and the point cloud. In the progressive irregular triangular mesh encryption algorithm, the processing device selects the lowest point as the initial ground seed point, constructs an initial triangular mesh, and gradually adds points that meet the conditions to the ground point set based on iterative angle and distance parameters. The processing device traverses all point cloud data and, based on the geometric feature thresholds set by the algorithm, determines whether each point belongs to bare ground or to overlying cover. Points identified as ground are labeled as ground points, while points identified as vegetation, buildings, or other obstacles are labeled as non-ground points.

[0078] Step 302: Use ground points for interpolation to generate a digital elevation model.

[0079] In the digital elevation model (DEM), each pixel value represents the absolute elevation of the exposed surface at that geographic coordinate, forming the baseline terrain surface for subsequent calculations. The processing device extracts the classified set of ground point data and sets the spatial resolution and coordinate range of the target raster image. Since the spatial distribution of ground points is non-uniform, the processing device uses spatial interpolation algorithms to estimate the elevation values ​​of unsampled locations. Specifically, the processing device uses inverse distance weighted interpolation, calculating the weighted average elevation of known ground points surrounding the point to be interpolated to determine the elevation of that point. Alternatively, the processing device uses Kriging interpolation, utilizing a variogram to analyze spatial correlation and perform unbiased estimation.

[0080] Step 303: Use non-ground points for interpolation to generate a digital surface model.

[0081] The digital surface model generated by the processing equipment is presented in the form of a raster image, where the pixel values ​​record the absolute elevation of the topmost surface cover. The processing equipment reads classified non-ground point data or raw point cloud data containing all echo information, and constructs a regular two-dimensional planar grid within the area to be measured. The processing equipment traverses and searches the point cloud data within each grid cell, comparing the vertical coordinate values ​​of each point, and identifies and extracts the point data with the highest elevation value within each grid. The highest point data represents the physical location of the topmost surface cover, such as the top of a tree canopy or the roof of a building. Based on the extracted set of highest point data from each grid, the processing equipment uses a spatial interpolation algorithm to calculate the continuous surface elevation of the entire area. Specifically, the processing equipment uses an irregular triangular mesh linear interpolation algorithm or an inverse distance weighted interpolation algorithm to calculate the elevation values ​​at the grid nodes, filling in the gaps between discrete points. The processing equipment organizes the elevation data generated by the interpolation calculation into a raster data format to generate the digital surface model.

[0082] Step 304: Subtract the digital surface model from the digital elevation model pixel by pixel on the raster to calculate the true height of the ground features. Use the true height of the ground features to construct the canopy height model; calculate the true height of the ground features using the following formula:

[0083]

[0084] in, For pixels The actual height of vegetation or ground features; For pixels The altitude at the highest point on the Earth's surface; For pixels The altitude of the exposed surface.

[0085] In the canopy height model, non-zero value regions visually represent the height distribution of vegetation, while zero value regions represent bare ground without vegetation. The processing device first calibrates the spatial reference frames of the digital surface model and the digital elevation model to ensure that the number of rows and columns and the pixel size are completely consistent. The processing device then traverses each row and column of the raster matrix, locating specific pixel coordinates. The processing device reads the pixel value of the digital surface model at that coordinate. This represents the elevation of the top of a feature; simultaneously, it reads the pixel value of the digital elevation model at that coordinate. This represents the Earth's surface elevation. The processing device performs a subtraction operation, and the result... This represents the relative vertical height of features at that location, eliminating the influence of terrain undulations on height measurement. The processing device writes the calculated height values ​​of all pixels into a new raster file, which becomes the canopy height model.

[0086] The UAV-based method for calculating vegetation cover provided in this embodiment separates point clouds using a ground point filtering algorithm and constructs digital elevation models and digital surface models respectively. Then, a canopy height model is generated through raster interpolation. This processing method eliminates the interference of terrain undulations on the measurement of ground feature height, restoring the true vertical height of vegetation relative to the ground surface. The canopy height model constructed using this true height provides clear three-dimensional structural features for subsequent semantic segmentation networks, helping to distinguish ground features with similar spectral characteristics but different heights (such as grassland and tree canopy), thereby improving the accuracy of vegetation cover calculation.

[0087] In one embodiment, the training steps of the deep learning semantic segmentation model may include:

[0088] Step 401: For each pixel in the training set, obtain the corresponding canopy height value from the canopy height model.

[0089] The processing device reads the training dataset from the storage medium. The training dataset contains multispectral image data and corresponding spatially relevant canopy height model raster data. The processing device constructs an index of training samples in memory, traversing the pixel coordinates of each training sample. Specifically, for each coordinate position, the processing device accesses the canopy height model data matrix and reads the stored floating-point value. This floating-point value represents the relative vertical height of the ground feature at that pixel. The processing device temporarily stores the read height value in a computational cache as the basic input data for subsequent weight calculations. Simultaneously, the processing device calculates the maximum height value in the current training set or training batch for subsequent normalization calculations, ensuring that the numerical range of the height data meets the algorithm requirements.

[0090] Step 402: Calculate the height weighting factor based on the canopy height value using the following formula:

[0091]

[0092] in, Highly weighted factors; In pixels; This represents the canopy height value corresponding to each pixel. The maximum height value in the training set; and To adjust the hyperparameters that affect height.

[0093] The height weight factor is a value greater than or equal to 1, which increases with the height of the ground feature. It is used to quantify the importance of the pixel corresponding to the height weight factor in the model training. and This is used to control the strength and non-linearity of the influence of height on the weights. The processing device divides the canopy height value of each pixel by the maximum height value in the training set to normalize the height data, obtaining a relative height value between 0 and 1. The processing device then performs... The exponentiation is performed to adjust the sensitivity to height differences. The processing device multiplies the result of the exponentiation by a hyperparameter. The product is then added to the baseline value of 1. The processing device performs the above operation on each pixel to generate a height weighting factor.

[0094] Step 403: Calculate the height-weighted mixed loss function based on the height weighting factor using the following formula:

[0095]

[0096]

[0097] in, It is a highly weighted mixed loss function; and These are the weighting coefficients; For the Focalloss function; An improved Diceloss function; For pixels The true label; For the model to pixels The probability of it being predicted as vegetation; It is a very small smoothing constant.

[0098] The height-weighted hybrid loss function comprehensively reflects the difference between the model's predictions and the actual situation, and this difference measure is explicitly weighted by pixel height. The processing device performs forward propagation of the deep learning model to generate a predicted probability map. The processing device first calculates the Focalloss function value, using this function to reduce the weight of easily classified samples and focus on difficult-to-classify samples. Subsequently, the processing device calculates the improved Diceloss function value. Specifically, the processing device performs a dot product operation on the matrix, combining the predicted probability matrix, the true label matrix, and the previously generated height-weighted factor matrix. Element-by-element multiplication is performed. The processing device calculates the weighted intersection (numerator) and weighted union (denominator) separately, by introducing a smoothing constant. To prevent division by zero errors, a highly weighted Diceloss value is obtained. Finally, the processing device utilizes preset weighting coefficients. and The Focalloss and the height-weighted Diceloss values ​​are summed in a weighted manner to obtain the final height-weighted mixed loss function value. .

[0099] Step 404: Minimize the highly weighted mixed loss function through backpropagation to optimize the parameters of the deep learning semantic segmentation model, and obtain the trained deep learning semantic segmentation model.

[0100] The processing device, based on the calculated highly weighted mixed loss function value, initiates an automatic differentiation engine for backpropagation. Specifically, following the chain rule, the processing device calculates the gradient vector of the loss function with respect to the weights and biases of each layer in the model, layer by layer. These gradient vectors indicate the direction and magnitude of parameter updates. The processing device employs optimization algorithms such as stochastic gradient descent or Adam to update the network parameters of the model using the calculated gradients, thereby reducing the value of the loss function. Through multiple iterations of this process, the processing device continuously adjusts the model parameters until the loss function converges or reaches a preset number of training epochs. The processing device saves the optimized model state to memory; this optimized model is the highly sensitive, trained deep learning semantic segmentation model, used for subsequent inference tasks.

[0101] The UAV-based ecological vegetation coverage measurement method provided in this embodiment introduces a height weight factor dynamically calculated based on the canopy height model during the training process of the deep learning model, and constructs a height-weighted hybrid loss function. This approach explicitly incorporates three-dimensional structural information into the loss calculation, forcing the model to impose a higher penalty on segmentation errors of tall vegetation (such as trees) during training, thereby making the model pay more attention to vegetation features with a certain height. This can effectively solve the "same spectral, different objects" confusion problem that easily occurs when relying solely on two-dimensional spectral information (e.g., difficulty in distinguishing between spectrally similar grasslands and tree canopies), and can improve the model's segmentation accuracy and generalization ability for vegetation with complex vertical structures.

[0102] In one embodiment, a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs) may further include:

[0103] Step 501: Obtain satellite imagery covering the area to be tested.

[0104] The satellite imagery covers a wider geographic area, including the region collected by the UAV. The satellite imagery data contains spectral information across multiple bands, recording the reflectivity of ground features to solar radiation. The processing equipment receives satellite imagery data files from a satellite ground station or remote sensing data server via a data communication interface. The processing equipment reads the metadata information of the imagery data to confirm its spatial resolution, projection coordinate system, and acquisition time, ensuring consistency with the UAV data in terms of spatiotemporal reference. The processing equipment performs radiometric correction on the satellite imagery, converting the original quantized grayscale values ​​into radiance or top-atmosphere reflectance. The processing equipment further performs atmospheric correction calculations, using an atmospheric transport model to eliminate the scattering and absorption of light by atmospheric molecules and aerosols, obtaining accurate surface reflectance data. The processing equipment loads the preprocessed satellite imagery data into a memory cache, ready for subsequent feature calculations.

[0105] Step 502: Perform feature extraction on the satellite imagery to obtain a satellite feature vector that includes texture features and spectral features.

[0106] The processing equipment performs mathematical operations on the spectral bands of satellite imagery. It selects the near-infrared and red light bands to calculate the normalized difference vegetation index (NDVI) to reflect chlorophyll content and growth vitality. It also calculates an enhanced vegetation index (EDI) by introducing the blue light band to reduce atmospheric influence. Furthermore, it calculates a soil-regulated vegetation index (SRI) to correct for soil background interference with the spectral signal using soil regulation factors. These calculated indices constitute the spectral feature component. Simultaneously, the equipment defines sliding windows on specific bands or principal component components of the satellite imagery. Within these windows, it statistically analyzes the co-occurrence probability of pixel gray levels, constructing a gray-level co-occurrence matrix. Based on this matrix, it calculates second-order statistics such as entropy, contrast, homogeneity, and correlation to quantify the spatial texture structure of land cover. These calculated statistics constitute the texture feature component. Finally, the equipment combines and stitches the extracted spectral and texture feature values ​​according to pixel location to form a satellite feature vector that includes both texture and spectral features.

[0107] Step 503: The baseline vegetation cover true map is aggregated to the pixel scale of the satellite image using the spatial averaging method to obtain the aggregated training labels.

[0108] The processing equipment establishes a spatial correspondence between satellite image pixels and ground truth map pixels. Specifically, it determines the geographic boundary of each satellite pixel, retrieves all ground truth map pixels falling within that boundary, counts the number of pixels marked as vegetation, and calculates the proportion of these pixels to the total number of pixels within that boundary. This proportion represents the vegetation cover abundance of the corresponding land area for that satellite pixel. The processing equipment performs this aggregation calculation on all satellite pixels, generating a series of continuous coverage values. These coverage values ​​are used as aggregated training labels for the corresponding satellite pixels, reflecting the true vegetation information at the sub-pixel scale.

[0109] Step 504: Based on the satellite feature vectors and the aggregated training labels, train a nonlinear machine learning regression model to obtain the trained nonlinear machine learning regression model.

[0110] The processing device constructs a training dataset containing input features and a target variable, where the input features are satellite feature vectors and the target variable is the aggregated training label. The processing device selects a nonlinear machine learning regression algorithm, such as a random forest regression algorithm or an extreme gradient boosting tree algorithm. The processing device inputs the training dataset into the algorithm model. During training, the model seeks the nonlinear mapping between spectral features, texture features, and vegetation cover labels by constructing a decision tree group or iteratively optimizing the loss function. The processing device adjusts the model's hyperparameters through cross-validation to reduce the prediction error on the validation set. The processing device continuously corrects the node splitting rules or leaf node weights within the model using the training data until the model converges. The processing device saves the model with fixed parameters; this saved model is the trained nonlinear machine learning regression model.

[0111] Step 505: Apply the trained nonlinear machine learning regression model to the satellite image features of the target area to generate a large-scale thematic map of vegetation coverage in the target area.

[0112] The processing equipment acquires satellite imagery of a large target area and extracts spectral and texture features following the same process described above. It then calls upon a pre-trained nonlinear machine learning regression model stored in memory, inputting the feature vector of each pixel in the target area into the model for inference. Based on the learned nonlinear relationship, the model outputs a predicted vegetation cover value for each pixel. The processing equipment arranges these predicted values ​​according to their geospatial location and reconstructs them into a two-dimensional raster image. The pixel values ​​of this raster image directly represent the percentage of vegetation cover at the corresponding location. Finally, the processing equipment geocodes and encapsulates this image to generate a large-scale thematic map of vegetation cover. This thematic map provides macro-scale information on ecological vegetation distribution for regional ecological environment monitoring or assessment.

[0113] The UAV-based method for calculating ecological vegetation cover provided in this embodiment effectively expands the scale from point to area by aggregating ground truth maps generated by the UAV into training labels at the satellite pixel scale and training a nonlinear regression model using spectral and texture features of satellite imagery. By fitting the relationship between satellite features and vegetation cover using a nonlinear model, the problem of vegetation index saturation, which often occurs in densely vegetated areas, and the difficulty of distinguishing complex land cover using single spectral features can be solved by traditional linear models. Introducing texture features increases the model's ability to perceive spatial heterogeneity, thereby improving the accuracy and reliability of large-scale vegetation cover retrieval and providing more accurate data support for regional ecological monitoring.

[0114] Furthermore, feature extraction is performed on the satellite imagery to obtain a satellite feature vector that includes texture and spectral features, which may include:

[0115] Step 601: Calculate the gray-level co-occurrence matrix within a preset sliding window; wherein the sliding window is set on the PCA1 image of the satellite imagery.

[0116] The gray-level co-occurrence matrix (GLCM) reflects the spatial correlation of gray levels within a local area of ​​the image, recording the microstructural information of the texture. The processing device first performs principal component analysis (PCA) on the multi-band satellite image data. Specifically, by calculating the covariance matrix and eigenvectors between bands, the processing device linearly transforms the original highly correlated spectral bands into uncorrelated orthogonal variables, i.e., principal components. The processing device selects the first principal component (PCA1), which contains the most information and has the highest variance contribution rate, as the base image for texture calculation. The processing device defines a sliding window of fixed size on the PCA1, such as a 3×3 or 7×7 pixel rectangular area. The processing device controls this sliding window to perform row-by-row and column-by-column scanning on the image according to a set step size. For each position where the sliding window stops on the image, the processing device analyzes the spatial distribution of pixel gray values ​​within the window's coverage area. The processing device counts the frequency of two pixels with specific gray values ​​appearing simultaneously at a specific direction and distance within that local sliding window. The processing device fills the statistically obtained frequencies into a square matrix, the number of rows and columns of which equals the number of gray levels in the image. The processing device then normalizes this square matrix so that the sum of all its elements is 1, thereby generating a gray-level co-occurrence matrix.

[0117] Step 602: Extract second-order statistical texture features based on the gray-level co-occurrence matrix to obtain texture features; texture features include entropy, contrast, and homogeneity.

[0118] Among these parameters, the entropy value is higher when the texture within the window is chaotic and disordered, and lower when the texture is smooth or regular. Contrast reflects the degree to which the matrix values ​​are distributed near the main diagonal; the deeper the texture grooves, the clearer the visual effect, the greater the contrast, and the higher the contrast value. When the local variations in image texture are small and the gray-level distribution is uniform, the homogeneity value is large. The processing device reads the generated gray-level co-occurrence matrix and uses statistical formulas to calculate second-order texture statistics. The processing device calculates the entropy value by summing the products of matrix elements and their logarithmic values ​​and taking the negative, quantifying the randomness or complexity of the image texture. The processing device calculates the contrast value by summing the matrix elements multiplied by the square of the difference between their row and column indices, quantifying the sharpness and groove depth of the image texture. The processing device calculates the homogeneity value, also known as the inverse difference moment, by summing the matrix elements divided by 1 and the square of the difference between their row and column indices, quantifying the local uniformity of the image texture. The processing device stores the entropy, contrast, and homogeneity values ​​calculated for each pixel location to form corresponding texture features.

[0119] Step 603: Extract the original spectral bands of the satellite image and calculate the vegetation index to obtain spectral features.

[0120] The processing equipment reads raw spectral band data from satellite imagery files, including reflectance values ​​for the red, green, blue, and near-infrared bands, and performs algebraic operations using a band arithmetic unit. Specifically, the equipment calculates the Normalized Difference Vegetation Index (NDVI) by dividing the difference between the near-infrared and red bands by their sum to enhance the response to vegetation growth status and eliminate some radiation errors. It also calculates the Enhanced Vegetation Index (EVI) by incorporating the blue band into the formula to correct for atmospheric aerosol scattering and by introducing soil adjustment parameters to reduce soil background interference, thereby improving sensitivity to high biomass areas. Finally, it calculates the Soil-Regulated Vegetation Index (SVI) by setting specific soil adjustment factors to correct the reflectance of the red and near-infrared bands, making it suitable for areas with low vegetation cover. The processing equipment then combines the extracted raw band numerical matrix with the calculated vegetation index matrices. These data directly reflect the spectral reflectance characteristics and biophysical parameters of ground features, constituting the spectral features of the satellite imagery.

[0121] Step 604: Combine texture features and spectral features to generate satellite feature vectors.

[0122] The satellite feature vector, which integrates spectral and spatial texture information of ground features, serves as high-dimensional input data for subsequent machine learning regression models, establishing a mapping relationship between surface features and vegetation cover. The processing device acquires the texture and spectral features generated in the above steps. It checks the spatial reference system and raster size of all layers to ensure they are strictly aligned in geospatial space. For each pixel location, the processing device extracts its corresponding texture feature values ​​(including entropy, contrast, and homogeneity) and spectral feature values ​​(including raw band reflectance and various vegetation indices). These values ​​are arranged in a predetermined order to construct a one-dimensional array or vector structure. The processing device performs numerical standardization on each component of the feature vector, such as scaling the data to the 0-1 range or performing Z-score standardization, to eliminate differences in dimensions and orders of magnitude between different features. The processing device defines the processed vector as the satellite feature vector for that pixel location.

[0123] The UAV-based method for calculating ecological vegetation cover provided in this embodiment calculates the gray-level co-occurrence matrix on the first principal component of satellite imagery and extracts second-order statistical texture features (entropy, contrast, homogeneity). These texture features are then combined with spectral features to construct a feature vector containing multi-dimensional "spectral-spatial" information of land features. Using PCA1 for texture calculation concentrates the main information of the imagery and reduces computational load. Introducing texture features can compensate for the shortcomings of single spectral features in reflecting surface spatial structure and roughness, helping machine learning models distinguish land features with similar spectral responses but different surface textures (e.g., dense forests and fragmented grasslands with the same spectral values), thereby improving the accuracy of vegetation cover inversion in complex surface environments.

[0124] Furthermore, the baseline vegetation cover map is aggregated to the pixel scale of the satellite imagery using a spatial averaging method to obtain aggregated training labels, which may include:

[0125] Step 701: Obtain the pixel grid of the satellite image.

[0126] The processing device reads satellite imagery data files from the storage medium, parses the header information of the data files to extract spatial reference coordinate system parameters, geographic boundary coordinates, and raster resolution values. Based on the extracted resolution and geographic extent, the processing device constructs a virtual two-dimensional raster matrix structure in computer memory. The number of rows and columns of this raster matrix is ​​consistent with the pixel arrangement of the satellite imagery, and each matrix cell represents an independent imaging unit on the satellite imagery. The processing device determines the corner coordinates and center coordinates of each matrix cell in geographic space, establishing a mapping relationship between image pixel indices and geographic spatial locations.

[0127] Step 702: Spatially align the baseline vegetation cover ground map with the cell grid to obtain the spatially aligned baseline vegetation cover ground map.

[0128] The processing device reads the spatial reference information of the baseline vegetation cover ground truth map and compares it with the spatial reference information of the satellite imagery cell grid. If the coordinate systems of the two are different, the processing device calls a geographic coordinate transformation algorithm to reproject the ground truth map to the same geographic coordinate system as the cell grid. Under the unified coordinate system, the processing device overlays the ground truth map onto the cell grid, calculates the position index of each pixel in the ground truth map relative to the cell grid, and determines which satellite cell grid unit each pixel falls within. Through the above spatial position matching calculation, the processing device establishes the spatial hierarchy between pixels and cell grids, generating a spatially aligned baseline vegetation cover ground truth map with spatial mapping relationships.

[0129] Step 703: Calculate the average value of all pixel values ​​in the spatially aligned benchmark vegetation cover ground truth map to obtain the average value corresponding to each pixel, which is used as the aggregated training label.

[0130] The processing device iterates through each cell in the pixel grid. For each cell, it retrieves all ground truth pixels falling within the geographic boundary of that cell, based on spatial alignment. Since the ground truth image is a binary image, pixel values ​​represent vegetation (e.g., 1) or non-vegetation (e.g., 0). The processing device sums these pixel values ​​to calculate the total number of pixels marked as vegetation within that cell. Simultaneously, it counts the total number of pixels falling within that cell. The processing device then divides the total number of vegetation pixels by the total number of pixels and performs an arithmetic mean to obtain the area proportion of vegetation within the current satellite pixel. This statistical operation is repeated for all cells in the grid, assigning the average value to the corresponding satellite pixel and labeling this set of coverage values, one-to-one with each satellite pixel, as the aggregated training label.

[0131] The UAV-based ecological vegetation cover measurement method provided in this embodiment aggregates high-resolution UAV binary segmentation results to low-resolution satellite pixel scale using spatial averaging, achieving a conversion from discrete classification results to continuous coverage values. This processing method leverages the spatial detail advantage of UAV data to provide accurate "sub-pixel level" vegetation abundance true values ​​for mixed pixels at the satellite scale, solving the problem of lacking large-scale, high-precision ground validation data in satellite remote sensing inversion. Using this aggregated data as training labels, machine learning models can learn the nonlinear mapping relationship between satellite spectral features and surface vegetation cover, thereby improving the training effect and prediction accuracy of large-scale vegetation cover inversion models.

[0132] Furthermore, based on the satellite feature vectors and the aggregated training labels, a nonlinear machine learning regression model is trained to obtain the trained nonlinear machine learning regression model, which may include:

[0133] Step 801: The satellite feature vectors are used as training features, and together with the aggregated training labels, they are input into the nonlinear machine learning regression model.

[0134] The nonlinear machine learning regression model is either an Extreme Gradient Boosting (XGBoost) model or a Random Forest (RF) model. The processing device reads the sample dataset generated in the previous steps from memory, extracts the combined spectral and texture feature vectors corresponding to each pixel location to construct a feature matrix, and extracts the corresponding aggregated UAV ground truth coverage values ​​to construct a label vector. The processing device initializes an ensemble tree model instance. For example, if an RF model is selected, the processing device sets hyperparameters such as the number of decision trees, maximum depth, and feature sampling ratio, and feeds the feature matrix and label vectors into the initialized model algorithm through a data interface, establishing a connection between the input data and the model algorithm.

[0135] Step 802: Perform the training process of the nonlinear machine learning regression model to fit the nonlinear mapping relationship between the training features and the training labels, and obtain the nonlinear mapping relationship.

[0136] The processing device utilizes a bootstrap sampling method to generate multiple training subsets, independently grows decision trees on each subset, and randomly selects feature subsets when splitting nodes. It then aggregates the prediction results of all trees through averaging or voting mechanisms. During training, the device leverages multidimensional information from satellite feature vectors, particularly spatial structure cues provided by texture features, to fit the variation patterns of vegetation cover across different numerical ranges (especially high-coverage ranges), capturing the complex nonlinear dependencies between features and labels.

[0137] Step 803: Based on the nonlinear mapping relationship, output the nonlinear machine learning regression model that fits the nonlinear mapping relationship, as the trained nonlinear machine learning regression model.

[0138] The processing device monitors the error change curve or validation set metrics during training. Training terminates when the preset convergence condition or maximum number of iterations is reached. The processing device then solidifies all tree structures, splitting rules, leaf node weights, and related hyperparameters within the model. This solidified mathematical model object is serialized into a file or binary data stream and saved to non-volatile memory, serving as the completed non-linear machine learning regression model.

[0139] The UAV-based method for calculating ecological vegetation cover provided in this embodiment effectively overcomes the limitations of traditional linear regression models in describing the relationship between vegetation cover and remote sensing features by employing ensemble tree models such as Extreme Gradient Boosting (XGBoost) or Random Forest (RF) as nonlinear regressors. Specifically, linear models struggle to handle the "saturation" phenomenon of vegetation indices (such as NDVI) in medium-to-high vegetation cover areas, where vegetation index growth stagnates at higher cover levels, leading to underestimation. Nonlinear machine learning models, however, can fit this complex nonlinear mapping relationship and utilize texture features to help distinguish spectrally similar but structurally different regions, thereby significantly improving the calculation accuracy and model generalization ability in high vegetation cover areas and complex surface environments.

[0140] The preferred embodiment of the present invention provides a method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs), which includes the following steps:

[0141] Step 1: Acquire multispectral images and 3D point cloud data of the area to be tested from the UAV platform.

[0142] Step 2: Perform spatiotemporal registration on the multispectral image and the 3D point cloud data to obtain spatiotemporally registered multispectral image and spatiotemporally registered 3D point cloud data.

[0143] Step 3: Perform orthophoto processing on the spatiotemporally registered multispectral images to generate digital orthophotos.

[0144] Step 4: Using the spatiotemporally registered 3D point cloud data, separate it into ground points and non-ground points using a ground point filtering algorithm. Interpolate using ground points to generate a digital elevation model; interpolate using non-ground points or all points to generate a digital surface model. Subtract the digital surface model from the digital elevation model pixel-by-pixel on the raster to obtain the canopy height model, which represents the true vertical height of ground features.

[0145] Step 5: Extract spectral bands from the digital orthophoto; extract height bands from the canopy height model. Align the extracted spectral and height bands pixel-wise and perform a stacking operation to generate channel input tensors.

[0146] Step 6: During the training phase of the deep learning semantic segmentation model, for each pixel in the training set, obtain its corresponding canopy height value from the canopy height model.

[0147] Step 7: Calculate the height weighting factor for each pixel based on the canopy height value.

[0148] Step 8: Calculate the height-weighted mixed loss function based on the height weighting factor.

[0149] Step 9: Minimize the highly weighted mixed loss function through backpropagation to optimize the parameters of the deep learning semantic segmentation model, and obtain the trained deep learning semantic segmentation model.

[0150] Step 10: Input the channel input tensor into the deep learning semantic segmentation model for inference. The deep learning semantic segmentation model outputs a pixel-level segmentation mask, which is the baseline vegetation cover ground truth map.

[0151] Step 11: Acquire satellite imagery covering the area to be measured. Extract the original spectral bands from the satellite imagery and calculate the vegetation index to obtain spectral characteristics.

[0152] Step 12: Set a sliding window on the near-infrared band of the satellite imagery or on the first component of the principal component imagery. Calculate the gray-level co-occurrence matrix within the sliding window.

[0153] Step 13: Extract second-order statistical texture features based on the gray-level co-occurrence matrix to obtain texture features.

[0154] Step 14: Combine spectral features and texture features to generate satellite feature vectors.

[0155] Step 15: Obtain the satellite imagery pixel grid. Spatially align the baseline vegetation cover ground truth map with the pixel grid to obtain the spatially aligned baseline vegetation cover ground truth map.

[0156] Step 16: For each pixel in the pixel grid, calculate the average value of all pixel values ​​in the baseline vegetation coverage map after spatial alignment, and obtain the average value corresponding to each pixel.

[0157] Step 17: Use the average value corresponding to each pixel as the training label after pixel aggregation. Use the satellite feature vector as the training feature and the aggregated training label as the training label, and input them into the nonlinear machine learning regression model.

[0158] Step 18: Perform the nonlinear machine learning regression model training process to fit the nonlinear mapping relationship between training features and training labels. After the training process is completed, output the nonlinear machine learning regression model that fits the nonlinear mapping relationship, as the trained nonlinear machine learning regression model.

[0159] Step 19: Apply the trained nonlinear machine learning regression model to the satellite image features of the target area to generate a thematic map of vegetation coverage over a large area of ​​the target area.

[0160] This embodiment incorporates vertical structure information of ground features into the semantic segmentation process by stacking and fusing digital orthophotos and canopy height models at the pixel level. This fusion method utilizes a deep learning model to simultaneously extract spectral and 3D structural features of ground features, reducing the adverse effects of changes in illumination conditions and shading on vegetation extraction accuracy when relying solely on 2D spectral information. A height-weighted hybrid loss function is introduced during training, using a height weight factor to penalize segmentation errors of tall vegetation, making the model more focused on vegetation features with height information and reducing confusion between similar and dissimilar spectral features. A deep learning semantic segmentation model is used to segment the fused multi-channel data, distinguishing ground features with similar spectral features but different heights, improving the accuracy and robustness of generating a baseline vegetation cover ground truth map for the tested area. Furthermore, this embodiment achieves scale expansion by aggregating ground truth maps generated by UAVs into satellite pixel-scale training labels and combining satellite image spectral and texture features to train a nonlinear regression model. A nonlinear model is used to fit the relationship between satellite features and vegetation cover, addressing the vegetation index saturation problem of linear models in densely vegetated areas and the difficulty of distinguishing land cover using a single spectral feature. Introducing texture features enhances the model's ability to perceive spatial heterogeneity on the land surface, improves the accuracy and reliability of large-scale vegetation cover inversion, and provides data support for regional ecological monitoring.

[0161] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0162] Based on the same inventive concept, this application also provides an apparatus for implementing the above-mentioned method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs). The solution provided by this apparatus is similar to the implementation described in the above method. Therefore, the specific limitations of one or more UAV-based ecological vegetation coverage calculation apparatus embodiments provided below can be found in the limitations of the UAV-based ecological vegetation coverage calculation method described above, and will not be repeated here.

[0163] In one exemplary embodiment, such as Figure 3 As shown, an ecological vegetation coverage measurement device based on unmanned aerial vehicles (UAVs) is provided, which may include:

[0164] The acquisition module 910 is used to acquire multispectral images and three-dimensional point cloud data of the area to be tested collected by the UAV platform.

[0165] The digital orthophoto generation module 920 is used to perform spatiotemporal registration on multispectral images to obtain spatiotemporally registered multispectral images, and to perform orthophoto processing on the spatiotemporally registered multispectral images to generate high-resolution digital orthophotos.

[0166] The canopy height model generation module 930 is used to process 3D point cloud data and generate a canopy height model.

[0167] The fusion module 940 is used to stack and fuse high-resolution digital orthophotos and height data from the canopy height model at the pixel level to generate a multi-channel input tensor.

[0168] The baseline coverage module 950 is used to segment the multi-channel input tensor using a deep learning semantic segmentation model to generate a baseline vegetation coverage ground map of the area to be tested.

[0169] In one embodiment, the canopy height model generation module 930 may include:

[0170] The separation unit is used to separate 3D point cloud data into ground points and non-ground points using a ground point filtering algorithm.

[0171] Elevation model unit, used to interpolate ground points to generate a digital elevation model.

[0172] The surface model unit uses non-ground points for interpolation to generate a digital surface model.

[0173] The canopy height model unit is used to subtract the digital surface model from the digital elevation model pixel by pixel on the grid to calculate the true height of the ground features, and then use the true height of the ground features to construct the canopy height model.

[0174] In one embodiment, the drone-based ecological vegetation coverage measurement device may further include:

[0175] The training set module is used to obtain the corresponding canopy height value from the canopy height model for each pixel in the training set.

[0176] The height weighting factor module is used to calculate the height weighting factor based on the canopy height value.

[0177] The Highly Weighted Hybrid Loss Function module is used to calculate the highly weighted hybrid loss function based on the high weight factor.

[0178] The model optimization module is used to optimize the parameters of the deep learning semantic segmentation model by minimizing the highly weighted mixture loss function through backpropagation, thus obtaining the trained deep learning semantic segmentation model.

[0179] In one embodiment, the drone-based ecological vegetation coverage measurement device may further include:

[0180] The satellite imagery module is used to acquire satellite images covering the area to be measured.

[0181] The feature extraction module is used to extract features from satellite imagery to obtain satellite feature vectors that include texture features and spectral features.

[0182] The aggregation label module is used to aggregate the baseline vegetation cover ground truth map to the pixel scale of satellite imagery using the spatial averaging method, thereby obtaining the aggregated training labels.

[0183] The nonlinear model training module is used to train a nonlinear machine learning regression model based on satellite feature vectors and aggregated training labels, resulting in a trained nonlinear machine learning regression model.

[0184] The thematic map module is used to apply the trained nonlinear machine learning regression model to the satellite image features of the target area to generate a large-scale thematic map of vegetation coverage of the target area.

[0185] Furthermore, the feature extraction module may include:

[0186] The matrix calculation unit is used to calculate the gray-level co-occurrence matrix within a preset sliding window.

[0187] The matrix extraction unit is used to extract second-order statistical texture features based on the gray-level co-occurrence matrix to obtain texture features.

[0188] The satellite image extraction unit is used to extract the original spectral bands of satellite images and calculate the vegetation index to obtain spectral features.

[0189] The feature combination unit is used to combine texture features and spectral features to generate satellite feature vectors.

[0190] Furthermore, the aggregation tag module may include:

[0191] Pixel acquisition unit, used to acquire the pixel grid of satellite imagery.

[0192] Alignment cells are used to spatially align the baseline vegetation cover map with the cell grid to obtain a spatially aligned baseline vegetation cover map.

[0193] The aggregation label unit is used to calculate the average value of all pixel values ​​in the spatially aligned benchmark vegetation cover ground truth map, and obtain the average value corresponding to each pixel as the aggregated training label.

[0194] Furthermore, the nonlinear model training module may include:

[0195] The input unit is used to input satellite feature vectors as training features, along with aggregated training labels, into a nonlinear machine learning regression model.

[0196] The fitting unit is used to perform the training process of a nonlinear machine learning regression model to fit the nonlinear mapping relationship between training features and training labels, thereby obtaining the nonlinear mapping relationship.

[0197] The output unit is used to output a nonlinear machine learning regression model that fits the nonlinear mapping relationship, which serves as the trained nonlinear machine learning regression model.

[0198] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement any step of the above-described method for calculating ecological vegetation coverage based on unmanned aerial vehicles.

[0199] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0200] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0201] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: Acquire multispectral images and 3D point cloud data of the area to be tested, collected by the UAV platform; The multispectral image is spatiotemporally registered to obtain a spatiotemporally registered multispectral image. The spatiotemporally registered multispectral image is then subjected to orthophoto processing to generate a high-resolution digital orthophoto image. The three-dimensional point cloud data is processed to generate a canopy height model; The high-resolution digital orthophoto and the height data of the canopy height model are stacked and fused at the pixel level to generate a multi-channel input tensor. The multi-channel input tensor is segmented using a deep learning semantic segmentation model to generate a baseline vegetation coverage map of the area to be tested.

2. The method according to claim 1, characterized in that, The process of processing the three-dimensional point cloud data to generate a canopy height model includes: The 3D point cloud data is separated into ground points and non-ground points using a ground point filtering algorithm; The ground points are used for interpolation to generate a digital elevation model; The non-ground points are used for interpolation to generate a digital surface model; The true height of the ground feature is calculated by subtracting the digital surface model from the digital elevation model pixel by pixel on the grid, and the canopy height model is constructed using the true height of the ground feature. The true height of the ground feature is calculated using the following formula: in, For pixels The actual height of vegetation or ground features; For pixels The altitude at the highest point on the Earth's surface; For pixels The altitude of the exposed surface.

3. The method according to claim 1, characterized in that, The training steps of the deep learning semantic segmentation model include: For each pixel in the training set, the corresponding canopy height value is obtained from the canopy height model; Calculate the height weighting factor based on the canopy height value using the following formula: in, Highly weighted factors; In pixels; This represents the canopy height value corresponding to each pixel. The maximum height value in the training set; and To adjust the hyperparameters affecting height; Calculate the height-weighted mixed loss function based on the height weighting factor using the following formula: in, It is a highly weighted mixed loss function; and These are the weighting coefficients; For the Focalloss function; An improved Diceloss function; For pixels The true label; For the model to pixels The probability of it being predicted as vegetation; It is a very small smoothing constant; The parameters of the deep learning semantic segmentation model are optimized by minimizing the highly weighted mixed loss function through backpropagation, resulting in the trained deep learning semantic segmentation model.

4. The method according to claim 1, characterized in that, The method further includes: Acquire satellite imagery covering the area to be measured; Feature extraction is performed on the satellite imagery to obtain a satellite feature vector that includes texture features and spectral features; The baseline vegetation cover true map is aggregated to the pixel scale of the satellite image using the spatial averaging method to obtain the aggregated training labels; Based on the satellite feature vectors and the aggregated training labels, a nonlinear machine learning regression model is trained to obtain the trained nonlinear machine learning regression model. The trained nonlinear machine learning regression model is applied to the satellite image features of the target area to generate a large-scale thematic map of vegetation coverage in the target area.

5. The method according to claim 4, characterized in that, The feature extraction process for the satellite imagery, yielding a satellite feature vector including texture and spectral features, includes: The gray-level co-occurrence matrix is ​​calculated within a preset sliding window; wherein the sliding window is set on the PCA1 image of the satellite imagery; The texture features are obtained by extracting second-order statistical texture features from the gray-level co-occurrence matrix; the texture features include entropy, contrast, and homogeneity. The original spectral bands of the satellite image are extracted, and the vegetation index is calculated to obtain the spectral features; The texture features and the spectral features are combined to generate the satellite feature vector.

6. The method according to claim 5, characterized in that, The step of aggregating the baseline vegetation cover ground truth map to the pixel scale of the satellite image using a spatial averaging method to obtain the aggregated training labels includes: Obtain the pixel grid of the satellite image; The baseline vegetation coverage map is spatially aligned with the pixel grid to obtain a spatially aligned baseline vegetation coverage map. The average value of all pixel values ​​in the spatially aligned benchmark vegetation cover ground truth map is calculated to obtain the average value corresponding to each pixel, which is used as the aggregated training label.

7. The method according to claim 6, characterized in that, The step of training a nonlinear machine learning regression model based on the satellite feature vectors and the aggregated training labels to obtain the trained nonlinear machine learning regression model includes: The satellite feature vectors are used as training features, and the aggregated training labels are input into a nonlinear machine learning regression model. The training process of the nonlinear machine learning regression model is performed to fit the nonlinear mapping relationship between the training features and the training labels, thereby obtaining the nonlinear mapping relationship; Based on the nonlinear mapping relationship, a nonlinear machine learning regression model that fits the nonlinear mapping relationship is output as the trained nonlinear machine learning regression model.

8. A device for calculating ecological vegetation coverage based on unmanned aerial vehicles (UAVs), characterized in that, The device includes: The acquisition module is used to acquire multispectral images and 3D point cloud data of the area to be tested collected by the UAV platform; The digital orthophoto generation module is used to perform spatiotemporal registration on the multispectral image to obtain a spatiotemporally registered multispectral image, and to perform orthophoto processing on the spatiotemporally registered multispectral image to generate a high-resolution digital orthophoto image. The canopy height model generation module is used to process the three-dimensional point cloud data and generate a canopy height model. The fusion module is used to stack and fuse the high-resolution digital orthophoto and the height data of the canopy height model at the pixel level to generate a multi-channel input tensor. The baseline coverage module is used to segment the multi-channel input tensor using a deep learning semantic segmentation model to generate a baseline vegetation coverage map of the area to be tested.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images