A vegetation coverage rapid measurement method and system based on image recognition

Abstract

The application relates to the technical field of ecological intelligent monitoring, in particular to a vegetation coverage rapid measurement method and system based on image recognition, which comprises the following steps: constructing a multi-source vegetation image acquisition device, obtaining vegetation image initial data of a target area based on the multi-source vegetation image acquisition device; performing image preprocessing and light self-adaptive correction on the vegetation image initial data to obtain vegetation image standardized data; establishing an improved lightweight semantic segmentation model, segmenting the vegetation image standardized data to obtain vegetation image segmentation results; preliminarily measuring the vegetation coverage of the target area based on the vegetation image segmentation results to obtain preliminary vegetation coverage; and constructing a geometric distortion dynamic compensation model to dynamically compensate the preliminary vegetation coverage, so that the vegetation coverage can be rapidly measured. The application effectively improves the efficiency and accuracy of vegetation coverage measurement.

Description

Technical Field

[0001] This invention relates to the field of ecological intelligent monitoring technology, and in particular to a method and system for rapid calculation of vegetation coverage based on image recognition. Background Technology

[0002] In existing technologies, vegetation cover measurement largely relies on remote sensing image analysis, ground sampling surveys, or a combination of both. For remote sensing imagery, multispectral images acquired by satellites or drones are commonly used. Vegetation areas are delineated by calculating indicators such as the Normalized Difference Vegetation Index (NDVI), and the coverage is obtained by statistically analyzing the percentage of vegetation pixels. Some methods combine traditional machine learning models to classify images, further refining the boundaries between vegetation and non-vegetation areas. Ground surveys involve setting up quadrats, manually measuring the vegetation cover area within each quadrat, and extrapolating the overall regional situation, or using handheld spectrometers to collect vegetation reflectance data to assist in the calculation. Furthermore, some technical solutions fuse remote sensing imagery with ground sampling data, using interpolation methods to optimize the spatial distribution of coverage results, thus balancing the measurement range and accuracy.

[0003] However, existing technologies have significant limitations, including: First, remote sensing images are limited by resolution, resulting in insufficient accuracy in identifying low vegetation and sparse areas, and are easily affected by cloud and atmospheric scattering. Second, traditional machine learning models perform poorly in complex scenarios, leading to low accuracy in vegetation coverage measurement. Third, while ground sampling offers high accuracy, it is time-consuming and labor-intensive, making it difficult to meet the needs of rapid measurement over large areas. Fourth, existing solutions do not fully consider changes in lighting during image acquisition, geometric distortion caused by equipment posture, and misclassification of dead vegetation and living vegetation, resulting in systematic biases in the measurement results. Fifth, existing segmentation models are mostly heavy networks with low computational efficiency, making it difficult to meet the practical application requirements of real-time and rapid measurement, thus restricting the immediate application of vegetation coverage data in fields such as ecological monitoring and agricultural management. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for rapid measurement of vegetation coverage based on image recognition.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for rapid calculation of vegetation cover based on image recognition, the method comprising the following steps: A multi-source vegetation image acquisition device is constructed, and initial vegetation image data of the target area is acquired based on the multi-source vegetation image acquisition device; The initial vegetation image data is preprocessed and subjected to adaptive illumination correction to obtain standardized vegetation image data. An improved lightweight semantic segmentation model is established to segment the standardized vegetation image data to obtain vegetation image segmentation results. Based on the vegetation image segmentation results, a preliminary vegetation coverage is calculated for the target area to obtain a preliminary vegetation coverage. A dynamic compensation model for geometric distortion is constructed to dynamically compensate for the preliminary vegetation coverage, thereby enabling rapid calculation of the vegetation coverage.

[0006] This invention improves data comprehensiveness through multi-source acquisition, ensures data quality through preprocessing and illumination correction, improves the lightweight model to balance segmentation accuracy and speed, combines preliminary calculation with feature optimization, and eliminates distortion effects through dynamic compensation, thereby achieving rapid and accurate calculation of vegetation coverage.

[0007] Optionally, the construction of the multi-source vegetation image acquisition device, and the acquisition of initial vegetation image data of the target area based on the multi-source vegetation image acquisition device, includes: The multi-source vegetation image acquisition device is constructed by combining drones and ground-based imaging equipment; The drone acquires distant vegetation image data of the target area; The ground-based imaging equipment is used to collect close-up vegetation image data of the target area; The distant vegetation image data and the close-up vegetation image data are used as the initial data for the vegetation image.

[0008] This invention combines drones and ground equipment to construct a multi-source acquisition device, which simultaneously acquires distant and close-up vegetation images. This not only covers the overall situation of a large area but also captures local vegetation details, improving the completeness and richness of the initial data and laying the foundation for subsequent accurate calculations.

[0009] Optionally, the step of performing image preprocessing and illumination adaptive correction on the initial vegetation image data to obtain standardized vegetation image data includes: The initial vegetation image data is preprocessed to obtain preprocessed vegetation image data. The image preprocessing includes image format and resolution standardization, image denoising, edge enhancement, and image registration. The vegetation image data of the target area is obtained by filtering the non-vegetated areas of the preprocessed vegetation image data. A light unevenness quantification factor is obtained to construct a light dynamic correction function, and the vegetation image data and the light dynamic correction function are combined to obtain the vegetation image standardized data.

[0010] This invention improves data consistency through preprocessing such as standardization and noise reduction, filters non-vegetation areas to focus on effective information, and combines a dynamic illumination correction function to eliminate illumination interference, thereby obtaining standardized data and providing high-quality input for subsequent segmentation and calculation.

[0011] Optionally, the establishment of the improved lightweight semantic segmentation model includes: A dynamic fusion gating mechanism is constructed, and a dual-branch heterogeneous backbone model structure is established by combining spatial branches and spectral branches; A vegetation image density prior index is obtained based on the vegetation image standardized data, and a pruning rate control function is constructed based on the vegetation image density prior index. A deformable convolutional attention unit is established, and the improved lightweight semantic segmentation model is constructed by combining the dual-branch heterogeneous backbone model structure and the pruning rate control function.

[0012] This invention extracts multi-dimensional features through a dual-branch structure, optimizes feature integration through dynamic fusion gating mechanism, improves efficiency through pruning rate control function, and enhances detail perception through deformable convolutional attention unit. The constructed model improves the segmentation accuracy and objectivity of images.

[0013] Optionally, the step of segmenting the standardized vegetation image data to obtain the vegetation image segmentation result includes: Based on the standardized vegetation image data, the spatial morphological features of vegetation are obtained according to the spatial branch, and the spectral response features of vegetation are obtained according to the spectral branch. Based on the dynamic fusion gating mechanism, the vegetation spatial morphological features and the vegetation spectral response features are dynamically weighted and fused to obtain a vegetation enhancement joint feature map; Adaptive feature pruning is performed on the vegetation enhancement joint feature map according to the pruning rate control function to obtain a vegetation pruning feature map; By combining the deformable convolutional attention unit with the vegetation pruning feature map, a refined vegetation morphology feature map is obtained; The vegetation image segmentation result is obtained by binarizing the refined vegetation morphology feature map.

[0014] This invention extracts spatial and spectral features through dual-branch extraction, dynamically fuses and enhances feature representation, improves efficiency through adaptive pruning, strengthens detail perception through deformable convolutional attention, and significantly improves segmentation accuracy and processing speed through binarization segmentation, thereby enhancing the accuracy and efficiency of vegetation image segmentation results.

[0015] Optionally, the step of performing a preliminary calculation of the vegetation coverage of the target area based on the vegetation image segmentation results to obtain a preliminary vegetation coverage includes: A three-dimensional overlap correction mechanism is established, and the quantification value of leaf overlap is obtained based on the vegetation image segmentation results. Construct a litter condition criterion, and obtain the litter quantization value of the vegetation image segmentation result based on the litter condition criterion; The preliminary vegetation coverage is obtained by combining the vegetation image segmentation results, the leaf overlap quantification value, and the litter quantification value, and by performing differential weight allocation based on the vegetation type.

[0016] This invention corrects leaf overlap errors through three-dimensional calibration, quantifies and removes fallen leaves according to criteria, and combines vegetation type-differentiated weighting to optimize the calculation logic in multiple dimensions, effectively improving the accuracy and scene adaptability of preliminary coverage.

[0017] Optionally, the criteria for constructing litter conditions include: Live vegetation samples were collected under standard illumination conditions, and the texture features of the live vegetation samples and the baseline range of near-infrared reflectance were obtained. The near-infrared band reflectance threshold is obtained by correcting the near-infrared band reflectance baseline range based on the living vegetation texture characteristics. The average reflectance in the near-infrared band of the vegetation image segmentation result is obtained, and the litter condition criterion is constructed by combining the near-infrared band reflectance threshold.

[0018] This invention collects samples under standard illumination, combines the average reflectance in the near-infrared band of vegetation image segmentation results with texture features to correct the near-infrared reflectance threshold, and accurately constructs a litter criterion. This can effectively distinguish between living vegetation and litter, reduce misjudgments, and improve the accuracy of vegetation coverage measurement.

[0019] Optionally, the construction of the geometric distortion dynamic compensation model to dynamically compensate for the preliminary vegetation coverage, thereby enabling rapid calculation of the vegetation coverage, includes: Obtain the multi-parameter distortion factor of the multi-source vegetation image acquisition device, and construct the geometric distortion dynamic compensation model based on the multi-parameter distortion factor; Based on the aforementioned geometric distortion dynamic compensation model, the edge distortion rate is obtained by adaptive compensation of the edge region of the vegetation image segmentation result. The vegetation coverage is obtained by dynamically compensating the initial vegetation coverage based on the edge distortion rate.

[0020] This invention is based on the multi-parameter distortion factor of a multi-source vegetation image acquisition device. It obtains the multi-parameter distortion factor to construct a dynamic compensation model, and adaptively compensates the edge region of the segmentation result to obtain the edge distortion rate, so as to correct the initial coverage, effectively eliminate the influence of geometric distortion, realize the rapid and accurate measurement of vegetation coverage, and effectively improve the efficiency and accuracy of vegetation coverage measurement.

[0021] Optionally, obtaining the multi-parameter distortion factor of the multi-source vegetation image acquisition device includes: An inertial measurement unit is deployed on the multi-source vegetation image acquisition device, and the multi-parameter distortion factor is obtained through the inertial measurement unit. The multi-parameter distortion factor includes pitch angle, yaw angle and relative flight altitude.

[0022] This invention uses an inertial measurement unit to accurately acquire multi-parameter distortion factors, providing reliable data support for dynamic compensation of geometric distortion and improving the accuracy and adaptability of the compensation model.

[0023] Secondly, the present invention provides a rapid vegetation cover measurement system based on image recognition, comprising: The initial data module constructs a multi-source vegetation image acquisition device and acquires initial vegetation image data of the target area based on the multi-source vegetation image acquisition device. The standardized data module performs image preprocessing and adaptive illumination correction on the initial vegetation image data to obtain standardized vegetation image data. The segmentation result module establishes an improved lightweight semantic segmentation model to segment the standardized vegetation image data and obtain the vegetation image segmentation result. The preliminary vegetation coverage module calculates the preliminary vegetation coverage of the target area based on the vegetation image segmentation results. The rapid calculation module constructs a dynamic compensation model for geometric distortion to dynamically compensate for the preliminary vegetation coverage, thereby realizing the rapid calculation of the vegetation coverage.

[0024] Thirdly, the present invention provides a rapid vegetation coverage calculation system based on image recognition. The system executes the rapid vegetation coverage calculation method based on image recognition provided by the present invention. The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions.

[0025] This invention utilizes high-performance hardware to store and execute calculation methods, enabling efficient invocation of instructions for rapid and accurate vegetation coverage calculation. This improves operational convenience and result reliability, and adapts to the application needs of vegetation coverage calculation in various scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for rapid calculation of vegetation coverage based on image recognition, according to an embodiment of the present invention. Figure 2 This is a block diagram of a rapid vegetation coverage measurement system based on image recognition, according to an embodiment of the present invention. Figure 3This is a schematic diagram of a rapid vegetation coverage measurement system based on image recognition, according to an embodiment of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0028] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0029] Please see Figure 1 One embodiment of the present invention provides a method for rapid calculation of vegetation cover based on image recognition, the method comprising the following steps: Step S1: Construct a multi-source vegetation image acquisition device, and acquire initial vegetation image data of the target area based on the multi-source vegetation image acquisition device; Step S2: Perform image preprocessing and illumination adaptive correction on the initial vegetation image data to obtain standardized vegetation image data; Step S3: Establish an improved lightweight semantic segmentation model to segment the standardized vegetation image data to obtain vegetation image segmentation results; Step S4: Based on the vegetation image segmentation results, perform a preliminary calculation of the vegetation coverage of the target area to obtain a preliminary vegetation coverage. Step S5: Construct a geometric distortion dynamic compensation model to dynamically compensate for the preliminary vegetation coverage, thereby realizing rapid calculation of the vegetation coverage.

[0030] The image recognition-based rapid vegetation cover calculation method and system according to embodiments of the present invention can construct a multi-source vegetation image acquisition device to acquire initial vegetation image data of a target area, perform image preprocessing and adaptive illumination correction to obtain standardized vegetation image data, then segment the standardized vegetation image data to obtain vegetation image segmentation results, and perform preliminary vegetation cover calculation of the target area to obtain preliminary vegetation cover. Furthermore, a geometric distortion dynamic compensation model is constructed to dynamically compensate for the preliminary vegetation cover, achieving rapid calculation of vegetation cover. Multi-source acquisition improves data comprehensiveness, preprocessing and illumination correction ensure data quality, an improved lightweight model balances segmentation accuracy and speed, preliminary calculation is combined with feature optimization, and dynamic compensation eliminates distortion effects, achieving rapid and accurate vegetation cover calculation, effectively improving the efficiency and accuracy of vegetation cover calculation.

[0031] Step S1: Construct a multi-source vegetation image acquisition device, and acquire initial vegetation image data of the target area based on the multi-source vegetation image acquisition device.

[0032] In this embodiment, a multi-source vegetation image acquisition device is constructed by combining a drone and a ground-based imaging device; distant vegetation image data of the target area is acquired by the drone; close-up vegetation image data of the target area is acquired by the ground-based imaging device; and the distant vegetation image data and close-up vegetation image data are used as the initial vegetation image data.

[0033] In this embodiment, the multi-source vegetation image acquisition device is constructed based on the core design principle of "distant macroscopic coverage + close-up detail capture" to achieve collaborative linkage between the UAV and ground-based imaging equipment. The UAV is a quadcopter equipped with a high-resolution camera and a multispectral sensor. The camera resolution is no less than 20 megapixels, and the multispectral sensor covers the blue (450nm±10nm), green (550nm±10nm), red (650nm±10nm), and near-infrared (850nm±10nm) bands to ensure simultaneous acquisition of vegetation morphological characteristics and spectral response information. The positioning module and inertial measurement unit on the UAV support centimeter-level positioning accuracy and attitude recording, outputting parameters such as flight altitude, pitch angle, and yaw angle in real time, providing raw data for subsequent geometric distortion compensation. The ground-based imaging equipment uses a portable multispectral camera and tripod combination. The camera's focal length is adjustable (24mm-70mm), and it supports autofocus and manual exposure control. It is equipped with a detachable supplementary lighting device to cope with complex lighting environments. Its data interface must be compatible with the UAV ground station system to achieve synchronous storage and transmission of image data.

[0034] In this embodiment, a wireless communication link is established between the UAV and the ground equipment to achieve real-time data interaction, ensuring that the spatial coordinate matching error of the two data collection areas does not exceed 0.5 meters. The data collection route and shooting parameters of the UAV are preset, and manual triggering of shooting function is supported to meet the flexible data collection needs of special terrains. In addition, the device needs to be equipped with a portable power storage power supply and protective devices to adapt to long-term operation in complex field environments.

[0035] In this embodiment, image acquisition based on a multi-source acquisition device follows a process of "macroscopic to microscopic, overall to local" to ensure the completeness and relevance of data coverage. Before collecting distant view data by the UAV, the boundary of the target area is delineated using satellite maps, and a gridded flight path is planned using ground station software. The flight path is parallel to the main terrain slope of the area to avoid insufficient image overlap due to terrain undulations. During flight, the UAV collects data in layers at a preset altitude. After each shooting group is completed, a real-time stitched preview image is automatically generated. If any missed areas are found (such as shadows or cloud cover), a reshoot mechanism is immediately triggered. The collected distant view images are the distant vegetation image data, and this image data must include the distribution of vegetation communities, topography, and large non-vegetation targets (such as buildings and roads) within the area. The ground resolution of a single image is controlled at 5cm-10cm to ensure that the canopy outline of the vegetation canopy can be identified.

[0036] In this embodiment, the ground-based close-range data acquisition must precisely correspond to the drone's acquisition area. A combination of "sample plot layout + random sampling" is used. For example, within the target area designated by the drone, standard sample plots are laid out in a 100m × 100m grid. Within each sample plot, 3-5 representative sub-regions (such as herbaceous layer, shrub layer, and understory of trees) are selected. Ground-based imaging equipment is used to capture images from multiple angles (3-5 images per sub-region, covering front, side, and top views). During shooting, the lens axis must be kept horizontal or slightly tilted downwards to avoid excessive leaf overlap due to excessive shooting angle. Simultaneously, supplementary lighting is used to balance the brightness of backlit or shadowed areas. The captured close-range images are the close-range vegetation image data. These images must clearly show leaf texture, stem structure, and the contact boundary between vegetation and the ground surface, with a resolution of no less than 5 megapixels and a vegetation coverage of no less than 60% in a single image, providing detailed support for subsequent leaf overlap correction and litter identification. Furthermore, the distant vegetation image data and the close-up vegetation image data can be used as the initial data for the vegetation image.

[0037] In this embodiment, data storage adopts a "local caching + cloud backup" model. Raw images are stored in real-time on a portable hard drive at the acquisition site, organized into folders based on "date + region + device type". After the day's acquisition is completed, the data is synchronized to a cloud database via encrypted transmission. The database supports multi-user access control and fast retrieval, allowing users to search for corresponding image data based on keywords such as acquisition time, region coordinates, and vegetation type. Furthermore, the system detects image sharpness, exposure, and the presence of motion blur, marking substandard images as "pending processing" and prompting for re-inspection, providing reliable data input for subsequent preprocessing and model segmentation.

[0038] In this way, a multi-source acquisition device can be built by combining drones and ground equipment to simultaneously acquire distant and close-up vegetation images. This not only covers the overall situation of a large area but also captures local vegetation details, improving the completeness and richness of the initial data and laying the foundation for subsequent accurate calculations.

[0039] Step S2: Perform image preprocessing and adaptive illumination correction on the initial vegetation image data to obtain standardized vegetation image data.

[0040] Specifically, S2 includes the following steps: Step S21: Perform image preprocessing on the initial vegetation image data to obtain preprocessed vegetation image data. The image preprocessing includes image format and resolution standardization, image denoising, edge enhancement, and image registration.

[0041] In this embodiment, when preprocessing the initial vegetation image data, firstly, image format and resolution standardization is performed. For example, UAV multispectral images and images acquired by ground equipment are uniformly converted to Tagged Image File Format (TIFF) to retain the original spectral information and shooting parameters. Bicubic interpolation is then used to unify the resolution of all images, ensuring that each pixel corresponds to the actual ground size consistently, avoiding scale differences from affecting subsequent analysis. Next, image denoising is performed. For example, adaptive median filtering is used for images containing Gaussian noise, with the window size dynamically adjusted according to noise density. Bilateral filtering is used for images dominated by salt-and-pepper noise, smoothing noise while preserving edge details and improving the image signal-to-noise ratio. Subsequently, edge enhancement is implemented. For example, vegetation contours are detected by first smoothing the image with Gaussian filtering, then calculating the gradient magnitude and direction, and extracting continuous edges through non-maximum suppression and double thresholding to enhance the clarity of leaf edges and canopy boundaries. Finally, image registration is completed, using the distant UAV image as a reference, and employing a scale-invariant feature transformation algorithm. The SIFT transform is used to extract feature points from the close-up image. The Random Sample Consensus (RANSAC) algorithm is used to remove mismatched points. Coordinate alignment is achieved through affine transformation. After registration, the average error of the overlapping area is controlled within 1 pixel, forming spatially consistent vegetation image preprocessing data.

[0042] Step S22: Filter the non-vegetated areas of the preprocessed vegetation image data to obtain the vegetation image data of the target area.

[0043] In this embodiment, non-vegetation areas are filtered based on preprocessed vegetation image data. Multi-feature fusion is used to distinguish vegetation from the background by combining spectral and texture features. For example, utilizing the high reflectivity of vegetation in the near-infrared band and low reflectivity in the red band, a normalized vegetation index (NVI) is calculated. Areas with an NVI < 0.2 are marked as potential non-vegetation areas (including soil, rocks, etc.). Next, image texture features are extracted, and contrast and entropy values ​​are calculated for auxiliary filtering. Hard backgrounds such as rocks typically have high contrast and low entropy values, and typical background interference areas such as soil and rocks are accurately labeled. Finally, overlapping non-vegetation areas are filtered, and blurred edges are fine-tuned using color features. The filtered non-vegetation areas are filled with a black background while retaining complete vegetation areas, thus obtaining the vegetation image data for the target area.

[0044] Step S23: Obtain the light unevenness quantification factor to construct the light dynamic correction function, and combine the vegetation image data and the light dynamic correction function to obtain the vegetation image standardized data.

[0045] In this embodiment, when obtaining the quantification factor for uneven illumination, the vegetation image data of the target area is first converted from the RGB (red, green, blue) color space to the HSV (hue, saturation, brightness) space, and the brightness channel that specifically reflects brightness information is extracted. By calculating the overall brightness average value and brightness distribution standard deviation of the brightness channel, and combining the ratio of the two with an exponential function adjustment term based on the brightness average value, the quantification factor for uneven illumination is obtained, thereby measuring the uniformity of image illumination and the brightness bias. The quantification factor for uneven illumination satisfies the following formula (1): (1) in, This is a quantification factor for uneven illumination. The standard deviation of the brightness channel. This represents the average value of the brightness channel. It is a natural constant.

[0046] In this embodiment, when constructing the dynamic illumination correction function, the illumination non-uniformity quantification factor is used as input, and dynamically adjusted correction parameters are designed: when the illumination non-uniformity quantification factor indicates that the illumination is too dim and uneven, the parameter is automatically adjusted to a value less than 1 to enhance the brightness of the weak light area; when the illumination non-uniformity quantification factor indicates that the illumination is too strong and unevenly distributed, the parameter is adjusted to a value greater than 1 to suppress the overexposure of the strong light area; under medium illumination conditions, the parameter remains in a balanced state. The dynamic illumination correction function satisfies the following formula (2): (2) in, This is the illumination dynamic correction function. This is a quantification factor for uneven illumination.

[0047] In this embodiment, the dynamic correction function is applied to the brightness channel of the vegetation image data, and targeted adjustments are made according to the brightness characteristics of different regions. Then, the corrected brightness channel and the original hue and saturation channels are converted back to RGB space to finally obtain the standardized vegetation image data. This effectively eliminates the brightness difference caused by uneven lighting, keeps the brightness of the vegetation area consistent, and provides a stable and reliable image foundation for subsequent semantic segmentation. The corrected output satisfies the following formula (3): (3) in, These are the corrected brightness channel values. The maximum quantization value of the image brightness. These are the original brightness channel values ​​before correction. This is the dynamic correction function for illumination.

[0048] In this way, data consistency can be improved through preprocessing such as standardization and noise reduction, non-vegetation areas can be filtered to focus effective information, and light interference can be eliminated by combining the dynamic light correction function to obtain standardized data, providing high-quality input for subsequent segmentation and calculation.

[0049] Step S3: Establish an improved lightweight semantic segmentation model to segment the standardized vegetation image data to obtain the vegetation image segmentation result.

[0050] Step S3 specifically includes the following steps: Step S31: Establish an improved lightweight semantic segmentation model.

[0051] In this embodiment, a dynamic fusion gating mechanism is constructed, and a dual-branch heterogeneous backbone model structure is established by combining spatial and spectral branches; a vegetation image density prior index is obtained based on vegetation image standardized data, and a pruning rate control function is constructed based on the vegetation image density prior index; a deformable convolutional attention unit is established, and an improved lightweight semantic segmentation model is constructed by combining the dual-branch heterogeneous backbone model structure and the pruning rate control function.

[0052] In this embodiment, a dual-branch heterogeneous backbone model structure is established. The spatial branch adopts a lightweight convolutional neural network with 3×3 depthwise separable convolutions as the basic unit. It extracts spatial features of vegetation images (e.g., leaf contours, canopy textures, etc.) through progressive downsampling. The number of channels in the output feature map at each stage is controlled between 16 and 128 to balance feature extraction capability and computational efficiency. The spectral branch is designed for multispectral vegetation data and adopts a structure combining 1×1 convolutions and channel attention. It focuses on capturing the spectral response differences in near-infrared, red, and other bands to enhance the spectral distinction between vegetation and non-vegetation areas. The above spatial branch satisfies the following formula (4): (4) in, The spatial morphological characteristics of vegetation output by spatial branches. This represents a 3×3 depthwise separable convolution. For activation function, For batch normalization, Standardize the data for vegetation images.

[0053] The above spectral branches satisfy the following formula (5): (5) in, The vegetation spectral response characteristics output by the spectral branch. This indicates the channel attention module. Represents a 1×1 convolution. Standardize the data for vegetation images.

[0054] In this embodiment, the dynamic fusion gating mechanism serves as the core of the dual-branch feature interaction. First, global average pooling is performed on the vegetation spatial morphology features output by the spatial branch to extract its global average information. Simultaneously, global max pooling is performed on the vegetation spectral response features output by the spectral branch to extract its global maximum response information. The two results are then concatenated along the channel dimension to form a fused global feature vector. Further, this vector is input into a multilayer perceptron for processing. Feature importance scores are generated through a fully connected layer, and then the scores are mapped to the 0-1 interval by an activation function to obtain dynamic fusion weights. Finally, the vegetation spatial morphology features and vegetation spectral response features are weighted using the dynamic fusion weights to obtain the fused vegetation enhancement joint feature map. This achieves adaptive dynamic fusion of spatial and spectral features, enabling the model to flexibly adjust the contribution ratio of the two branches according to feature importance. The dynamic fusion gating mechanism satisfies the following formulas (6) and (7): (6) (7) in, For dynamic weight fusion, For activation function, It is a multilayer perceptron. Indicates global average pooling. The spatial morphological characteristics of vegetation output by spatial branches. Indicates global max pooling. The vegetation spectral response characteristics output by the spectral branch. This is a combined feature map of vegetation enhancement.

[0055] In this embodiment, when calculating the vegetation image density prior index based on the standardized vegetation image data, the image is first divided into several sliding windows of 16×16 pixels. The number of windows is determined by rounding down the product of the image width and height ratios to 16. For each window, the difference between the near-infrared and red light band reflectance is calculated, and then divided by the maximum value of this difference in the entire image to obtain the spectral response coefficient. At the same time, the difference between the average brightness value within the window and the global brightness expectation value (default 128) is calculated. Combined with the brightness standard deviation, an illumination compensation term is generated through an exponential function to reduce the error in highlight or shadow areas. The spectral response coefficient of each window is multiplied by the illumination compensation term, and the average value of all window results is taken as the vegetation image density prior index, which satisfies the following formula (8): (8) in, The prior index for vegetation image density. The number of sliding windows, The index variable for the sliding window. This represents the difference in reflectance between the near-infrared band and the red band. This indicates taking the maximum reflectivity difference over the entire range. It is an exponential function. For the first The average brightness within each sliding window The expected value of global brightness. This represents the standard deviation of the brightness channel.

[0056] In this embodiment, when constructing the pruning rate control function, a piecewise function is first established to determine the basic pruning rate based on the density prior index; then, a hyperbolic tangent function is established using the density sensitivity coefficient (default 3.0) and the density threshold (default 0.4) to adjust the basic pruning rate and obtain the final pruning rate. The above piecewise function satisfies the following formula (9): (9) in, Based on the pruning rate, is the prior index of vegetation image density.

[0057] The above pruning rate control function satisfies the following formula (10): (10) in, For the final pruning rate, Based on the pruning rate, It is the hyperbolic tangent function. Density sensitivity coefficient The prior index for vegetation image density. This is the density threshold.

[0058] In this embodiment, when establishing the deformable convolutional attention unit, the vegetation enhancement joint feature map output by the dual-branch heterogeneous backbone model is first processed, and the offset of each sampling point of the convolution kernel is learned through a 1×1 convolutional layer, so that the 3×3 convolution kernel can adaptively adjust its spatial position to fit irregular shapes such as leaf curling and branch tilting. At the same time, the gradient direction histogram of the feature map is extracted, the peak value is matched with the preset direction angle, and the contribution of sampling points consistent with the vegetation growth direction is strengthened through weighting. The deformable convolutional attention unit satisfies the following formulas (11) and (12): (11) (12) in, This is the offset. This represents the convolution operation. For vegetation enhancement joint feature map, This is a refined vegetation morphology feature map after deformable convolution processing. The index variable for the sampling point. The weights of the sampling points, This represents the original location of the sampling point.

[0059] In this embodiment, deformable convolutional attention units are embedded into the feature extraction stage of the dual-branch model to replace traditional convolutional layers and improve morphological adaptability. Subsequently, the model structure is dynamically optimized using a pruning rate control function. During model training and inference, the vegetation image density prior index is calculated in real time, applying a higher base pruning rate to sparse regions and a fixed base pruning rate to dense regions. Furthermore, the constructed improved lightweight semantic segmentation model retains the ability to finely capture complex vegetation morphologies, achieving efficient inference while maintaining segmentation accuracy.

[0060] In this way, multi-dimensional features can be extracted through a dual-branch structure, the feature integration can be optimized by a dynamic fusion gating mechanism, the efficiency can be improved by a pruning rate control function, and the detail perception can be enhanced by deformable convolutional attention units. The constructed model improves the segmentation accuracy and objectivity of images.

[0061] Step S32: Segment the standardized vegetation image data to obtain the vegetation image segmentation result.

[0062] In this embodiment, based on standardized vegetation image data, spatial morphological features of vegetation are obtained according to spatial branches, and spectral response features of vegetation are obtained according to spectral branches. A dynamic weighted fusion of the spatial morphological features and spectral response features of vegetation is performed using a dynamic fusion gating mechanism to obtain a vegetation enhancement joint feature map. Adaptive feature pruning is performed on the vegetation enhancement joint feature map according to a pruning rate control function to obtain a vegetation pruning feature map. A refined vegetation morphological feature map is obtained by combining deformable convolutional attention units to perform feature perception on the vegetation pruning feature map. Finally, the refined vegetation morphological feature map is binarized and segmented to obtain the vegetation image segmentation result.

[0063] In this embodiment, when segmenting the standardized vegetation image data, the data is first processed using a dual-branch heterogeneous backbone model. The spatial branch extracts spatial morphological features of the vegetation through depthwise separable convolution stacking, capturing details such as leaf edges and canopy texture. The spectral branch uses 1×1 convolution and channel attention mechanisms to obtain spectral response features in the near-infrared and red light bands, enhancing the spectral difference between vegetation and the background. Next, a dynamic fusion gating mechanism is used to dynamically weight and fuse the two types of features to obtain a vegetation-enhanced joint feature map. Subsequently, adaptive pruning is performed on this joint feature map according to a pruning rate control function. A higher pruning rate is used for areas with low vegetation density, while the pruning intensity is reduced for areas with high density, preserving key features and reducing redundant computation, thus obtaining the vegetation pruning feature map. Further, deformable convolutional attention units are used to process the pruned feature map. By learning the convolution kernel offset to adapt to morphologies such as leaf curling and branch tilting, and by allocating sampling weights according to the gradient direction, the feature response of the vegetation growth direction is enhanced, resulting in a refined vegetation morphological feature map. Finally, the refined vegetation morphology feature map is binarized and segmented, and an appropriate threshold is set to distinguish between vegetation and non-vegetation areas, and the vegetation image segmentation result is output.

[0064] In this way, spatial and spectral features can be extracted through dual-branch extraction, dynamically fused to enhance feature representation, improved efficiency through adaptive pruning, enhanced detail perception through deformable convolutional attention, and significantly improved segmentation accuracy and processing speed through binarization segmentation, thereby improving the accuracy and efficiency of vegetation image segmentation results.

[0065] Step S4: Based on the vegetation image segmentation results, perform a preliminary calculation of the vegetation coverage of the target area to obtain the preliminary vegetation coverage.

[0066] In this embodiment, a three-dimensional overlap correction mechanism is established to obtain the quantification value of leaf overlap based on the vegetation image segmentation results; a litter condition criterion is constructed to obtain the litter quantification value of the vegetation image segmentation results based on the litter condition criterion; and the preliminary vegetation coverage is obtained by combining the leaf overlap quantification value and the litter quantification value and performing differentiated weight allocation based on the vegetation type.

[0067] In this embodiment, when establishing a three-dimensional overlap correction mechanism to obtain the quantification value of leaf overlap, firstly, leaf instance segmentation is performed on the vegetation image segmentation results. By identifying each leaf and generating a pixel-level mask, the contour boundary and pixel area of ​​each leaf are marked. Then, a stereo matching algorithm is used to construct the three-dimensional shape of the leaf. The depth information of each pixel of the leaf is calculated by combining the disparity map of the binocular camera to distinguish the hierarchical relationship of different leaves in space (such as upper leaves occluding lower leaves). Next, the area of ​​visible leaves is counted. The pixels of the area not occluded by other leaves are extracted in the two-dimensional image and accumulated to obtain the total area of ​​visible leaves. At the same time, the complete shape of the leaf is restored by morphological dilation operation. The mask of each leaf is moderately dilated (the dilation coefficient is set according to the average thickness of the leaf) to estimate the complete pixel area including the occluded part. The total area of ​​all leaves is obtained by accumulating. Finally, the quantification value of overlap is calculated. The ratio of the visible leaf area to the complete area of ​​all leaves is used as the non-overlap coefficient to obtain the quantification value of leaf overlap. The above quantification value of leaf overlap satisfies the following formula (13): (13) in, This is a quantification value for the degree of leaf overlap. This represents the total pixel area of ​​visible leaves in the vegetation image segmentation results. This represents the total number of leaves identified in the vegetation image segmentation results. For the index variable of the blade, For the first The pixel area of ​​the blade.

[0068] In this embodiment, live vegetation samples are collected under standard lighting conditions, and the texture features of the live vegetation samples and the near-infrared reflectance baseline range are obtained. The near-infrared reflectance baseline range is corrected based on the texture features of the live vegetation to obtain the near-infrared reflectance threshold. The average near-infrared reflectance of the vegetation image segmentation results is obtained, and the litter condition criterion is constructed by combining the near-infrared reflectance threshold.

[0069] In this embodiment, various typical live vegetation samples (such as fresh leaves of herbs, shrubs, and trees) were collected under standard lighting conditions. Texture feature data of the samples were acquired using a high-resolution imaging device, including the density of wrinkles and the distribution pattern of veins on the leaf surface. Simultaneously, the reflectance of the samples in the near-infrared band was measured using a spectrometer, and a baseline range of near-infrared reflectance for the live vegetation was statistically determined (typically, due to the presence of chlorophyll, the reflectance in this range is relatively high, generally between 40% and 60%). Subsequently, the baseline range was modified based on the texture characteristics of the live vegetation: for samples with fine textures (such as broad-leaved plants), whose near-infrared reflectance is more stable, the upper limit of the baseline range could be increased by 5% as a subdivision threshold. For samples with coarse textures (such as coniferous plants), whose reflectance fluctuates more, the lower limit of the baseline range was decreased by 3% to broaden the threshold adaptability, ultimately determining a near-infrared reflectance threshold applicable to multiple vegetation types. Furthermore, near-infrared reflectance data of the target area in the vegetation image segmentation results are extracted, and its average reflectance is calculated. If the average value is lower than the above-mentioned corrected near-infrared reflectance threshold, and irregular textures (such as broken edges and no obvious veins) unique to litter are detected in the area in combination with texture features, it is determined to be a litter area. Thus, a litter condition criterion that integrates spectral and texture features is constructed.

[0070] In this way, samples can be collected under standard lighting conditions. The average reflectance in the near-infrared band of vegetation image segmentation results is combined with texture features to correct the near-infrared reflectance threshold, accurately constructing a litter criterion. This can effectively distinguish between living vegetation and litter, reduce misjudgments, and improve the accuracy of vegetation coverage measurement.

[0071] In this embodiment, the litter quantification value is obtained based on the litter condition criterion. First, the pixel data of the target area is extracted from the vegetation image segmentation result, and the number of pixels identified as litter by the criterion is counted. Then, the total number of all vegetation pixels in the target area is calculated. Next, the green-red band reflectance difference is introduced for correction, and then the mean of the reflectance difference in the litter area is calculated. Compared with the maximum value of the difference in the whole image, the spectral attenuation coefficient is obtained. Finally, the ratio of the number of litter pixels to the total number of vegetation pixels is multiplied by the exponential function value corresponding to the spectral attenuation coefficient (this function decreases as the reflectance difference decreases, strengthening the distinction between litter and living vegetation) to obtain the litter quantification value. The higher the value, the greater the proportion of litter, which can accurately reflect the actual distribution ratio of litter in the target area. The above litter quantification value satisfies the following formula (14): (14) in, This represents the quantitative value of litter. This represents the number of pixels identified as litter in the vegetation image segmentation results. This represents the total number of pixels in all vegetation regions of the vegetation image segmentation result. It is an exponential function. This represents the difference in reflectivity between the green and red light bands. This indicates taking the maximum reflectance difference across the entire image.

[0072] In this embodiment, when calculating the preliminary vegetation coverage, firstly, the ratio of the total pixel area of ​​the vegetation region to the total pixel area of ​​the target region in the vegetation image segmentation result is calculated as the basic coverage ratio. Then, the basic ratio is corrected according to the quantization value of leaf overlap. Next, differentiated weights can be assigned based on vegetation type: herbaceous plants are assigned 0.6, shrubs 0.7, and trees 0.8. The higher the weight, the greater the influence of leaf overlap on coverage. The effective coverage ratio is multiplied by the corresponding weight to obtain the first part of coverage after overlap correction. At the same time, the living vegetation coefficient is obtained by subtracting the litter quantization value from 1, and then the second part of coverage is obtained by eliminating the influence of litter. Finally, the two parts of coverage are added together to obtain the preliminary vegetation coverage. It satisfies the following formula (15): (15) in, For preliminary vegetation coverage, Differentiated weights for vegetation types, This represents the total pixel area of ​​the vegetation region. The total pixel area of ​​the target region. This is a quantification value for the degree of leaf overlap. This represents the quantitative value of litter.

[0073] In this way, leaf overlap errors can be corrected through three-dimensional calibration, litter can be quantified and removed according to criteria, and vegetation type differentiation weighting can be combined to optimize the calculation logic in multiple dimensions, effectively improving the accuracy of preliminary coverage and scene adaptability.

[0074] Step S5: Construct a geometric distortion dynamic compensation model to dynamically compensate for the preliminary vegetation coverage, thereby realizing rapid calculation of the vegetation coverage.

[0075] In this embodiment, multi-parameter distortion factors from a multi-source vegetation image acquisition device are obtained, and a dynamic compensation model for geometric distortion is constructed based on these factors. Edge distortion rate is obtained by adaptively compensating the edge regions of the vegetation image segmentation results according to the dynamic compensation model. The initial vegetation coverage is then dynamically compensated based on the edge distortion rate to obtain the vegetation coverage. The multi-parameter distortion factors, including pitch angle, yaw angle, and relative altitude, are obtained through an inertial measurement unit.

[0076] In this embodiment, when acquiring the multi-parameter distortion factor of the multi-source vegetation image acquisition device, the flight attitude parameters, including pitch angle, yaw angle, and relative flight altitude, are collected in real time by the inertial measurement unit carried by the UAV. Based on the attitude parameters, a dynamic compensation model for geometric distortion is constructed, which integrates three types of compensation terms: pitch compensation term, which corrects the projection stretching caused by the tilt of the viewing angle by attenuating the pitch angle of the UAV by the cosine square value; yaw attenuation term, which introduces an exponential function to nonlinearly suppress the horizontal distortion caused by the yaw angle; and flight altitude correction term, which dynamically adjusts the compensation intensity according to the ratio of the actual flight altitude to the reference flight altitude, with a greater compensation intensity at lower flight altitudes. These three terms are combined in a weighted manner to output a three-dimensional distortion factor that comprehensively reflects the influence of flight attitude and altitude, thereby realizing a quantitative description of spatial projection distortion. The above dynamic compensation model for geometric distortion satisfies the following formula (16): (16) in, For three-dimensional distortion factor, The pitch angle, Yaw angle For relative flight altitude, It is a cosine function. It is a natural constant. This is the azimuth attenuation coefficient. As a highly weighted factor, This serves as the baseline flight altitude.

[0077] In this way, multi-parameter distortion factors can be accurately obtained through the inertial measurement unit, providing reliable data support for dynamic compensation of geometric distortion and improving the accuracy and adaptability of the compensation model.

[0078] In this embodiment, when performing edge region adaptive compensation on the vegetation image segmentation result based on the geometric distortion dynamic compensation model, the image is first divided into a central region (radius less than 40% of the image width) and an edge region, and a differentiated compensation strategy is adopted for different regions. When calculating the edge distortion rate, the distance from the pixel to the image center is used as a variable, combined with half the length of the image diagonal, and combined with the distortion intensity coefficient to construct the calculation formula for the edge distortion rate. For edge region pixels, the formula amplifies the distortion rate through the square term to enhance the compensation intensity; while the central region pixels maintain a lower distortion rate to reduce overcorrection. By traversing the edge pixels of the vegetation image segmentation result, substituting them into the formula to calculate the distortion rate of each pixel, and then taking the average distortion rate of the edge region as the overall edge distortion rate, the precise quantification of the geometric distortion of the image edge is achieved. The above edge distortion rate satisfies the following formula (17): (17) in, For edge distortion rate, The distance from a pixel to the center of the image. The distortion intensity coefficient, This represents the length of the image's diagonal.

[0079] In this embodiment, the initial vegetation coverage is dynamically compensated based on the edge distortion rate. First, the vegetation type compensation coefficient is determined according to the vegetation type: herbaceous plants do not require additional compensation due to their vertical projection characteristics, so the coefficient is 1.0; shrubs require moderate compensation due to the tilting of their lateral branches, so the coefficient is 0.9; and trees require enhanced compensation due to the spherical structure of their crowns, so the coefficient is 0.7. Then, the three-dimensional distortion factor, the edge distortion rate, and the vegetation type compensation coefficient are multiplied together, and the square root of the product is taken as the comprehensive compensation factor. Finally, the initial vegetation coverage is multiplied by this comprehensive compensation factor to obtain the final vegetation coverage. The following formula (18) is satisfied: (18) in, For vegetation coverage, For preliminary vegetation coverage, This is the vegetation type compensation coefficient. For three-dimensional distortion factor, This represents the edge distortion rate.

[0080] In this way, based on the multi-parameter distortion factors of the multi-source vegetation image acquisition device, a dynamic compensation model can be constructed using the multi-parameter distortion factors. The edge distortion rate is obtained by adaptive compensation of the edge region of the segmentation result to correct the initial coverage, effectively eliminate the influence of geometric distortion, and realize the rapid and accurate calculation of vegetation coverage, which effectively improves the efficiency and accuracy of vegetation coverage calculation.

[0081] Please see Figure 2 In an optional embodiment, the present invention provides a rapid vegetation cover measurement system based on image recognition, the system comprising: The initial data module constructs a multi-source vegetation image acquisition device and acquires initial vegetation image data of the target area based on the multi-source vegetation image acquisition device. The standardized data module performs image preprocessing and adaptive illumination correction on the initial vegetation image data to obtain standardized vegetation image data. The segmentation result module establishes an improved lightweight semantic segmentation model to segment the standardized vegetation image data and obtain the vegetation image segmentation result. The preliminary vegetation coverage module calculates the preliminary vegetation coverage of the target area based on the vegetation image segmentation results. The rapid calculation module constructs a dynamic compensation model for geometric distortion to dynamically compensate for the preliminary vegetation coverage, thereby realizing the rapid calculation of the vegetation coverage.

[0082] Please see Figure 3 In an optional embodiment, the present invention provides a rapid vegetation coverage calculation system based on image recognition, including an input device, an output device, a processor, and a memory. These hardware facilities are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute specific steps as described in the relevant embodiments of the rapid vegetation coverage calculation method based on image recognition provided by the present invention. The rapid vegetation coverage calculation system based on image recognition provided by the present invention has a complete structure, is objective and stable, and improves the overall applicability and practical application capability of the present invention.

[0083] In summary, the present invention provides a rapid vegetation cover calculation method and system based on image recognition. This method involves constructing a multi-source vegetation image acquisition device to obtain initial data, followed by preprocessing and adaptive illumination correction to obtain standardized data. An improved lightweight semantic segmentation model is then established to segment the data, yielding segmentation results. Based on the segmentation results, vegetation cover is initially calculated. Finally, dynamic compensation for geometric distortion is applied to obtain the final vegetation cover. The method is easy to understand, computationally simple, requires minimal workload, and is suitable for engineering applications, providing a theoretical foundation and technical support for the further development of the field of intelligent ecological monitoring technology.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for rapid measurement of vegetation cover based on image recognition, characterized in that, Includes the following steps: A multi-source vegetation image acquisition device is constructed, and initial vegetation image data of the target area is acquired based on the multi-source vegetation image acquisition device; The initial vegetation image data is preprocessed and subjected to adaptive illumination correction to obtain standardized vegetation image data. An improved lightweight semantic segmentation model is established to segment the standardized vegetation image data to obtain vegetation image segmentation results. Based on the vegetation image segmentation results, a preliminary vegetation coverage is calculated for the target area to obtain a preliminary vegetation coverage. A dynamic compensation model for geometric distortion is constructed to dynamically compensate for the preliminary vegetation coverage, thereby enabling rapid calculation of the vegetation coverage. The establishment of the improved lightweight semantic segmentation model includes: A dynamic fusion gating mechanism is constructed, and a dual-branch heterogeneous backbone model structure is established by combining spatial branches and spectral branches. The spatial branches satisfy the formula... ,in, The spatial morphological characteristics of vegetation output by spatial branches. This represents a 3×3 depthwise separable convolution. For activation function, For batch normalization, For vegetation image data standardization, the spectral branches satisfy the formula ,in, The vegetation spectral response characteristics output by the spectral branch. This indicates the channel attention module. Representing a 1×1 convolution, the dynamic fusion gating mechanism satisfies the formula. and formula ,in, For dynamic weight fusion, For activation function, It is a multilayer perceptron. Indicates global average pooling. Indicates global max pooling. For vegetation enhancement joint feature map; Based on the standardized vegetation image data, a priori index of vegetation image density is obtained. A pruning rate control function is then constructed based on this priori index, wherein the priori index of vegetation image density satisfies the formula... ,in, The prior index for vegetation image density. The number of sliding windows, The index variable for the sliding window. This represents the difference in reflectance between the near-infrared band and the red band. This indicates taking the maximum reflectivity difference over the entire range. It is an exponential function. For the first The average brightness within each sliding window The expected value of global brightness. The standard deviation of the brightness channel; A deformable convolutional attention unit is established, and the improved lightweight semantic segmentation model is constructed by combining the dual-branch heterogeneous backbone model structure and the pruning rate control function. The improved model is achieved through a piecewise function. Obtain the basic pruning rate In the formula The pruning rate control function is obtained, where, For the final pruning rate, Based on the basic pruning rate, It is the hyperbolic tangent function. Density sensitivity coefficient The prior index for vegetation image density. Given a density threshold, deformable convolutional attention units satisfy the formula... and ,in, This is the offset. This represents the convolution operation. For vegetation enhancement joint feature map, This is a refined vegetation morphology feature map after deformable convolution processing. For the index variable of the sampling point, The weights of the sampling points, This represents the original location of the sampling point.

2. The method for rapid measurement of vegetation cover based on image recognition according to claim 1, characterized in that, The construction of the multi-source vegetation image acquisition device, and the acquisition of initial vegetation image data of the target area based on the multi-source vegetation image acquisition device, includes: The multi-source vegetation image acquisition device is constructed by combining drones and ground-based imaging equipment; The drone acquires distant vegetation image data of the target area; The ground-based imaging equipment is used to collect close-up vegetation image data of the target area; The distant vegetation image data and the close-up vegetation image data are used as the initial data for the vegetation image.

3. The method for rapid calculation of vegetation cover based on image recognition according to claim 1, characterized in that, The process of performing image preprocessing and adaptive illumination correction on the initial vegetation image data to obtain standardized vegetation image data includes: The initial vegetation image data is preprocessed to obtain preprocessed vegetation image data. The image preprocessing includes image format and resolution standardization, image denoising, edge enhancement, and image registration. The vegetation image data of the target area is obtained by filtering the non-vegetated areas of the preprocessed vegetation image data. A light unevenness quantification factor is obtained to construct a light dynamic correction function, and the vegetation image data and the light dynamic correction function are combined to obtain the vegetation image standardized data.

4. The method for rapid measurement of vegetation cover based on image recognition according to claim 1, characterized in that, The step of segmenting the standardized vegetation image data to obtain the vegetation image segmentation result includes: Based on the standardized vegetation image data, the spatial morphological features of vegetation are obtained according to the spatial branch, and the spectral response features of vegetation are obtained according to the spectral branch. Based on the dynamic fusion gating mechanism, the vegetation spatial morphological features and the vegetation spectral response features are dynamically weighted and fused to obtain a vegetation enhancement joint feature map; Adaptive feature pruning is performed on the vegetation enhancement joint feature map according to the pruning rate control function to obtain a vegetation pruning feature map; By combining the deformable convolutional attention unit with the vegetation pruning feature map, a refined vegetation morphology feature map is obtained; The vegetation image segmentation result is obtained by binarizing the refined vegetation morphology feature map.

5. The method for rapid calculation of vegetation cover based on image recognition according to claim 1, characterized in that, The preliminary calculation of vegetation coverage in the target area based on the vegetation image segmentation results to obtain preliminary vegetation coverage includes: A three-dimensional overlap correction mechanism is established, and the quantification value of leaf overlap is obtained based on the vegetation image segmentation results. Construct a litter condition criterion, and obtain the litter quantization value of the vegetation image segmentation result based on the litter condition criterion; The preliminary vegetation coverage is obtained by combining the vegetation image segmentation results, the leaf overlap quantification value, and the litter quantification value, and by performing differential weight allocation based on the vegetation type.

6. The method for rapid calculation of vegetation cover based on image recognition according to claim 5, characterized in that, The criteria for constructing litter include: Live vegetation samples were collected under standard illumination conditions, and the texture features of the live vegetation samples and the baseline range of near-infrared reflectance were obtained. The near-infrared band reflectance threshold is obtained by correcting the near-infrared band reflectance baseline range based on the living vegetation texture characteristics. The average reflectance in the near-infrared band of the vegetation image segmentation result is obtained, and the litter condition criterion is constructed by combining the near-infrared band reflectance threshold.

7. The method for rapid calculation of vegetation cover based on image recognition according to claim 1, characterized in that, The construction of a dynamic compensation model for geometric distortion is used to dynamically compensate for the initial vegetation coverage, thereby enabling rapid calculation of the vegetation coverage. This includes: Obtain the multi-parameter distortion factor of the multi-source vegetation image acquisition device, and construct the geometric distortion dynamic compensation model based on the multi-parameter distortion factor; Based on the aforementioned geometric distortion dynamic compensation model, the edge distortion rate is obtained by adaptive compensation of the edge region of the vegetation image segmentation result. The vegetation coverage is obtained by dynamically compensating the initial vegetation coverage based on the edge distortion rate.

8. The method for rapid calculation of vegetation cover based on image recognition according to claim 7, characterized in that, The acquisition of the multi-parameter distortion factor of the multi-source vegetation image acquisition device includes: An inertial measurement unit is deployed on the multi-source vegetation image acquisition device, and the multi-parameter distortion factor is obtained through the inertial measurement unit. The multi-parameter distortion factor includes pitch angle, yaw angle and relative flight altitude.

9. A rapid vegetation cover measurement system based on image recognition, used to execute the method as described in any one of claims 1-8, characterized in that, include: The initial data module constructs a multi-source vegetation image acquisition device and acquires initial vegetation image data of the target area based on the multi-source vegetation image acquisition device. The standardized data module performs image preprocessing and adaptive illumination correction on the initial vegetation image data to obtain standardized vegetation image data. The segmentation result module establishes an improved lightweight semantic segmentation model to segment the standardized vegetation image data and obtain the vegetation image segmentation result. The preliminary vegetation coverage module calculates the preliminary vegetation coverage of the target area based on the vegetation image segmentation results. The rapid calculation module constructs a dynamic compensation model for geometric distortion to dynamically compensate for the preliminary vegetation coverage, thereby realizing the rapid calculation of the vegetation coverage.

Images