A method for detecting grassland empty spot and supplementing pasture based on unmanned aerial vehicle multi-spectral image

Abstract

This invention relates to the field of multispectral UAV grassland inspection and machine vision deep learning detection and segmentation algorithms. Specifically, it is a method for grassland spot detection and forage reseeding based on UAV multispectral imagery. The method involves a UAV equipped with a multispectral camera to acquire multispectral image data of grassland pastures. Professional remote sensing software is used to analyze and extract information on grassland cover and health status. Multiple deep learning algorithms are combined to detect and segment grassland spot areas, extracting information on bareness, coordinates, and size. Reseeding prescription data is generated based on the analysis results and optimized by considering soil conditions and climate factors. The prescription data is imported into the control terminal of a large agricultural UAV, which carries seeding and fertilization equipment, loading appropriate amounts of seeds and fertilizers for precise reseeding and fertilization. This method can effectively detect grassland spots, implement targeted forage reseeding, and promote the restoration and stability of the grassland ecosystem.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment for grassland inspection and deep learning algorithm technology, and in particular to a multispectral UAV grassland inspection and machine vision deep learning detection and segmentation algorithm. Background Technology

[0002] The statements in this section are merely background information relevant to this invention and do not necessarily constitute prior art.

[0003] Grasslands, as one of the most important ecosystems on Earth, not only provide abundant livestock resources for humankind but also play a vital role in maintaining biodiversity, regulating climate, and conserving soil and water. However, due to factors such as overgrazing, climate change, and land degradation, many grasslands around the world are facing serious ecological problems, such as grassland degradation, biodiversity loss, and declining productivity. These problems not only threaten the health of grassland ecosystems but also severely impact communities and economies that depend on them.

[0004] With the rapid development of unmanned aerial vehicle (UAV) technology, its application in agricultural monitoring is becoming increasingly widespread. High-resolution cameras carried by UAVs can quickly and efficiently acquire image data of large areas of pasture, providing strong technical support for precision agriculture. Multispectral imaging technology is a remote sensing technology capable of capturing the reflected or radiated energy of objects in different spectral bands. By analyzing this spectral data, important information such as vegetation type, growth status, and biomass can be obtained. In grassland monitoring, multispectral cameras combined with UAVs can quickly acquire detailed information about grassland cover and vegetation health, helping to identify grassland degradation and patchy areas.

[0005] The inventors discovered that conventional grassland spot detection relies solely on standard RGB remote sensing imagery and common deep learning algorithms for identification and analysis, resulting in relatively low accuracy. Furthermore, the acquired spectral imagery data is limited and lacks integration with spectral information from other bands, making it difficult to accurately identify and segment significantly degraded grassland spots from stitched remote sensing images. It is also difficult to distinguish the degree of grassland degradation within these spots, hindering managers' ability to promptly understand the actual condition of the grassland and implement appropriate management measures. This ultimately threatens the health of the grassland ecosystem and negatively impacts communities and economies that depend on the grassland. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for grassland patch detection and pasture reseeding based on UAV multispectral imagery. The method utilizes a multispectral camera mounted on a UAV to acquire image data of grassland pastures. Specialized remote sensing software and deep learning algorithms are used to analyze and extract grassland cover and health information, enabling the detection and segmentation of patchy grassland areas. Reseeding prescription data is generated based on the analysis results and optimized by incorporating soil conditions and climatic factors. Finally, the prescription data is imported into the control terminal of an agricultural UAV, allowing the UAV to perform precise reseeding and fertilization, thereby achieving grassland ecological protection and restoration.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for grassland spot detection and pasture reseeding based on UAV multispectral imagery. Based on the geographical environment information of the target ranch on that day, plan the multispectral UAV to select a suitable flight route, select the optimal terrain-following flight altitude, flight speed, forward overlap rate, and lateral overlap rate to ensure the coverage and quality of multispectral image data; Near-low altitude image data of grassland pastures are acquired by multispectral cameras and transmitted to a ground processing system. Professional remote sensing software is used to stitch, correct and analyze the acquired multispectral image data to extract information on grassland cover and distribution. The image data was processed using a deep learning algorithm that combines U-Net and DeepLabV3+ to detect and segment grassland vacancy areas, extract information on the degree of bareness, coordinates, and size of the vacancy areas, generate reseeding prescription data, and optimize the data by combining soil conditions and climate factors. QGIS Desktop software was used to extract multispectral reflectance information of segmented spots. Combined with the acquired NDVI and EVI information, the exposure degree of the spots was analyzed and mapped, and their coordinate location and area information were output to generate a reseeding prescription map. The data on reseeding is optimized by combining soil conditions and climate factors, and the prescription information is imported into a large plant protection drone. The plant protection drone carries sowing and fertilization equipment, loads an appropriate amount of seeds and fertilizer, and performs precise reseeding and fertilization.

[0008] As a further limitation of the first aspect of the present invention, based on the geographical environment information of the target ranch on that day, a multispectral UAV is planned to select a suitable flight path, and the optimal terrain-following flight altitude, flight speed, forward overlap rate, and lateral overlap rate are selected to ensure the coverage and quality of the multispectral image data, including: Use a handheld RTK device to locate the boundary corners of the ranch fence and import the latitude and longitude coordinate data into the DJI Mavic 3 Multispectral Edition drone remote controller; Based on the wind speed, wind direction and atmospheric visibility of the day, the flight speed and aerial photography altitude were modified, and the forward overlap rate and lateral overlap rate were adjusted. Plan drone flight routes based on the ranch terrain to minimize the number of times the drone turns around, thereby increasing the proportion of straight-line flight time; As a further limitation of the first aspect of the present invention, near-low-altitude image data of grassland pastures are acquired using a multispectral camera and transmitted to a ground processing system. Specialized remote sensing software is used to stitch, correct, and analyze the acquired multispectral image data, extracting information on grassland cover and distribution, including: Before the drone lands, manually control the drone to take multiple photos of the CH-MXDJ255075 P4M diffuse reflector to obtain the radiometric correction parameters of the images. Pix4Dmapper was used to perform alignment checks on the acquired multispectral remote sensing images, and spectral correction parameters for the green, red, near-infrared, and red edge bands were input respectively to complete radiometric correction during the stitching process. The latitude and longitude coordinates of the corner points obtained by handheld RTK are used as ground control points for the stitched image, and are also imported into Pix4Dmapper to correct the distortion of the stitched image and improve the image processing accuracy.

[0009] In ENVI 5.3.1, the band combination tool was used to calculate the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI), and the threshold filtering tool was used to extract the raster attribute values ​​with NDVI values ​​greater than 0.3 and EVI values ​​greater than 0.2, respectively. The NDVI value was assigned a weight of 0.6 and the EVI value a weight of 0.4 respectively. The raster blocks extracted in the two trials were then weighted and superimposed to generate an N-EVI map. By calculating the proportion of grid cells with pixel values ​​greater than 0.2 in the N-EVI map to the total number of grid cells in the pasture, the healthy grassland cover information of the pasture can be obtained, and the corresponding area is the distribution area of ​​healthy pasture. To further define the first aspect of this invention, a deep learning algorithm combining U-Net and DeepLabV3+ is applied to process image data, enabling the detection and segmentation of grassland patchy areas, extracting information on the degree of bareness, coordinate location, and size of the patches, generating reseeding prescription data, and optimizing the data in conjunction with soil conditions and climatic factors, including: In ENVI 5.3.1, by combining the pasture and grassland cover information in the generated N-EVI map, some regions of interest are manually created, and some representative empty and healthy areas in the mosaic image are selected using polygonal fine outlines.

[0010] The selected empty areas are labeled with pixel-level classification labels using the LabelImg tool, and OpenCV is used to crop, scale, and normalize the image to meet the input format of the subsequent model. The selected empty area is used as the original input image of the U-Net and DeepLabV3+ deep learning models. The encoders of U-Net and DeepLabV3+ are used to extract the low-level features of the image, and the feature maps are fused by concatenation and weighted summation in the encoder stage. A cascaded model is constructed, and the decoder part of U-Net is used to upsample the fused feature map to gradually restore the spatial resolution. The corresponding low-level feature map in the U-Net encoder is concatenated with the feature map in the decoder to retain more detailed information; The decoder section of DeepLabV3+ is used to further upsample and refine the features of the U-Net decoder output; In DeepLabV3+, the Atrous Spatial Pyramid Pooling module is used to capture multi-scale contextual information. In the output layer, multiple convolutional layers are added to the output of the DeepLabV3+ decoder, and the Softmax activation function is applied to output the probability of each pixel belonging to each category, while generating the final spot region segmentation map. The encoder using U-Net and DeepLabV3+ has a convolution kernel size of 3×3, a number of 32 kernels, a stride of 1, a feature map padding strategy of "same", and an activation function of the non-linear activation function "ReLU". The pooling type of the encoder pooling layer using U-Net and DeepLabV3+ is max pooling to highlight edge and texture features in the sparse image. The pooling window is 5×5. The pooling layer achieves dimensionality reduction of the image while retaining the most significant features. The Atrous Spatial Pyramid Pooling module in the DeepLabV3+ model contains four parallel dilated convolutional layers: a 3x3 convolutional kernel with a dilation rate of 6, a 3x3 convolutional kernel with a dilation rate of 12, a 3x3 convolutional kernel with a dilation rate of 18, and a 3x3 convolutional kernel with a dilation rate of 24. This expands the receptive field of the convolutional kernel without increasing the number of parameters. During training, the constructed U-Net and DeepLabV3+ fusion model uses cross-entropy loss function and Dice loss function to optimize the model, and selects Adam and SGD optimizers for model training; The constructed U-Net and DeepLabV3+ fusion model was finally evaluated using IoU (Intersection over Union) and accuracy metrics. As a further limitation of the first aspect of the present invention, QGIS Desktop software is used to extract the multispectral reflectance information of segmented empty spots, and the exposure degree of the empty spots is analyzed and mapped by combining the acquired NDVI and EVI information, and the coordinate position and area information are output to generate a reseeding prescription map, including: Soil samples were collected using soil testing equipment to analyze key parameters such as soil nutrient content, pH value, and organic matter content. Local temperature, humidity, rainfall, wind speed, and climate data are obtained through weather stations and meteorological data services. By integrating soil data, meteorological data, and remote sensing imagery data on the extent and location of exposed patches, optimized and detailed reseeding and fertilization prescriptions are generated, including information on forage seed type, fertilizer type, and dosage. The formulated prescription information is imported into the control system of the agricultural drone, and the drone's flight path, altitude, speed, and parameters for sowing and fertilization are further set. By loading an appropriate amount of seeds and fertilizer onto an agricultural drone, the drone flies along a preset path and parameters to achieve precise reseeding and fertilization. The operation process is monitored in real time by cameras and radar sensors on the drone to ensure the quality of reseeding and fertilization. As a further limitation of the first aspect of the invention, the reseeding data is optimized by combining soil conditions and climate factors, and the prescription information is imported into a large-scale plant protection drone. The plant protection drone carries sowing and fertilization equipment, loads an appropriate amount of seeds and fertilizer, and performs precise reseeding and fertilization, including: Soil samples were collected using soil testing equipment to analyze key parameters such as soil nutrient content, pH value, and organic matter content. Local temperature, humidity, rainfall, wind speed, and climate data are obtained through weather stations and meteorological data services. By integrating soil data, meteorological data, and remote sensing imagery data on the extent and location of exposed patches, optimized and detailed reseeding and fertilization prescriptions are generated, including information on forage seed type, fertilizer type, and dosage. The formulated prescription information is imported into the control system of the agricultural drone, and the drone's flight path, altitude, speed, and parameters for sowing and fertilization are further set. By loading an appropriate amount of seeds and fertilizer onto an agricultural drone, the drone flies along a preset path and parameters to achieve precise reseeding and fertilization. The operation process is monitored in real time by cameras and radar sensors on drones to ensure the quality of reseeding and fertilization.

[0011] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention will first utilize UAV multispectral imagery technology to acquire remote sensing data of grasslands, combining spectral information from multiple bands including RGB, red, green, near-infrared, infrared, and red-edge, in order to improve the accuracy of acquiring grassland cover, health status, and patch distribution information from multiple information sources as much as possible; 2. This invention innovatively introduces a deep learning model algorithm that integrates U-Net and DeepLabV3+ to automatically identify and segment empty areas in grassland remote sensing images. By training the deep learning model, the location and size information of bare surfaces and empty areas can be accurately extracted from complex multispectral images. This algorithm greatly improves the accuracy and efficiency of empty area detection, providing strong technical support for the precision of grassland management.

[0012] 3. This invention applies the concept of precision agriculture to grassland reseeding. By analyzing multispectral image data and extracting patch information using deep learning algorithms, it generates precise reseeding prescription data. It considers not only the specific location and size of grassland patches but also soil conditions and climatic factors, ensuring the scientific rigor and effectiveness of the reseeding work. Furthermore, it utilizes agricultural drones for precise reseeding and fertilization, further improving resource utilization efficiency and reseeding effectiveness.

[0013] 4. This invention not only focuses on the detection and reseeding of grassland patches, but also emphasizes the entire process management from data collection and analysis to reseeding implementation. Data is acquired through drones and multispectral imaging technology, analyzed using deep learning multi-algorithm fusion, and finally reseeded using precision agriculture technology, forming a complete management process for grassland ecosystem restoration and stability. This comprehensive management approach ensures the systematic and continuous nature of grassland ecological restoration work, providing a guarantee for the long-term stability of the grassland ecosystem.

[0014] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 This invention provides a roadmap for grassland spot detection technology based on UAV multispectral imagery; Figure 2 This invention provides a schematic diagram of the reseeding technology process based on grassland patch segmentation. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0020] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0021] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings.

[0022] A method for grassland spot detection and pasture reseeding based on UAV multispectral imagery, characterized by the following steps: S10. Based on the geographical environment information of the target ranch on that day, plan the multispectral UAV to select a suitable flight route, select the optimal terrain-following flight altitude, flight speed, forward overlap rate, and lateral overlap rate to ensure the coverage and quality of multispectral image data. S20. Acquire near-low-altitude image data of grassland pastures through multispectral cameras and transmit them to the ground processing system. Use professional remote sensing software to stitch, correct and analyze the acquired multispectral image data, and extract information on grassland cover and distribution. S30. The image data is processed using a deep learning algorithm that combines U-Net and DeepLabV3+ to detect and segment grassland vacant areas, extract information on the degree of bareness, coordinates, and size of the vacant areas, generate reseeding prescription data, and optimize the data by combining soil conditions and climate factors. S40. Use QGIS Desktop software to extract the multispectral reflectance information of the segmented spots, combine it with the acquired NDVI and EVI information to analyze and map the exposure degree of the spots, and output their coordinate position and area information to generate a reseeding prescription map. The S50 combines soil conditions and climate factors to optimize reseeding data and imports prescription information into a large plant protection drone. The drone carries sowing and fertilization equipment, loads an appropriate amount of seeds and fertilizer, and performs precise reseeding and fertilization.

[0023] Step S10 also includes: S11. Use a handheld RTK device to locate the boundary corners of the ranch fence and import the latitude and longitude coordinate data into the DJI Mavic 3 Multispectral Edition drone remote controller; S12. Based on the wind speed, wind direction and atmospheric visibility of the day, modify the forward flight speed and aerial photography altitude, and adjust the forward overlap rate and lateral overlap rate. S13. Plan the drone flight path according to the ranch terrain, and minimize the number of times the drone turns around to increase the proportion of straight flight time.

[0024] Step S20 also includes: S21. Before the drone lands, manually control the drone to take multiple photos of the CH-MXDJ255075 P4M diffuse reflector to obtain the radiometric correction parameters of the images. S22. Use Pix4Dmapper to perform alignment checks on the acquired multispectral remote sensing images, and input the spectral correction parameters for the green, red, near-infrared, and red edge bands respectively, and complete the radiometric correction during the stitching process. S23. Use the latitude and longitude coordinates of the corner points obtained by the handheld RTK as the ground control points of the stitched image, and import them into Pix4Dmapper to correct the distortion of the stitched image and improve the image processing accuracy.

[0025] S24. In ENVI 5.3.1, use the band combination tool to calculate the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), and use the threshold filtering tool to extract the raster block attribute values ​​with NDVI values ​​greater than 0.3 and EVI values ​​greater than 0.2, respectively. S25. Assign a weight of 0.6 to the NDVI value and a weight of 0.4 to the EVI value, and then weight and superimpose the two extracted raster blocks to generate an N-EVI map. S26. Calculate the proportion of grid blocks with pixel values ​​greater than 0.2 in the N-EVI map to the total number of grid blocks in the pasture. This will give you the information on the healthy grassland cover of the pasture. The corresponding area is the distribution area of ​​healthy pasture.

[0026] Step S30 also includes: S31. In ENVI 5.3.1, based on the pasture and grassland cover information in the generated N-EVI map, manually create some regions of interest, and use polygonal fine outlines to select some representative empty and healthy areas in the mosaic image.

[0027] S32. Use the LabelImg tool to create pixel-level classification labels for the selected empty areas, and use OpenCV to crop, scale, and normalize the image to meet the input format of the subsequent model. S33. The selected empty area is used as the original input image of the U-Net and DeepLabV3+ deep learning models. The encoders of U-Net and DeepLabV3+ are used to extract the low-level features of the image, and the feature maps are fused by concatenation and weighted summation in the encoder stage. S34. Construct a cascaded model and use the decoder part of U-Net to upsample the fused feature map to gradually restore the spatial resolution. S35. Concatenate the corresponding low-level feature maps in the U-Net encoder with the feature maps in the decoder to retain more detailed information; S36. Use the decoder part of DeepLabV3+ to further upsample and refine the output of the U-Net decoder; S37. Use the Atrous Spatial Pyramid Pooling module in DeepLabV3+ to capture multi-scale contextual information; In S38, the output layer adds multiple convolutional layers to the output of the DeepLabV3+ decoder and applies the Softmax activation function to output the probability of each pixel belonging to each category, while generating the final spot region segmentation map.

[0028] Step S40 further includes: S41. Import the red, green, near-infrared, red-edge, and RGB band stitched images, along with the NDVI and EVI band calculated images, into QGIS Desktop software. Simultaneously import the segmented empty spot areas as separate layers into the software and place them at the top layer. S42. Extract the latitude and longitude coordinates of the pixel points within the segmented area from the geometric field of the vector layer of the empty spot area, and use the field calculator to calculate and output the area of ​​each empty spot based on the latitude and longitude coordinate information. S43. Using the empty spot region layer segmented by the algorithm as the extraction boundary, extract the reflectance of the empty spot regions of 7 groups of images respectively, and perform weighted summation on their mapping values. Divide the degree of exposure of the empty spots according to the size of the result. S44. Map the green gradient color depth according to the weighted result, fill all empty areas in the entire pasture, and generate a reseeding prescription map. Step S50 further includes: S51. Collect soil samples using soil testing equipment and analyze key parameters such as soil nutrient content, pH value, and organic matter content. S52. Obtain local temperature, humidity, rainfall, wind speed and climate data through weather stations and meteorological data services; S53. Integrate soil data, meteorological data, and remote sensing imagery data on the degree and location of exposed patches to generate optimized and detailed reseeding and fertilization prescriptions, including information on forage seed type, fertilizer type, and dosage. S54. Import the formulated prescription information into the control system of the plant protection drone, and further set the drone's flight path, altitude, speed, and parameters for sowing and fertilization. S55. Load an appropriate amount of seeds and fertilizer onto the plant protection drone, and the drone will fly according to the preset path and parameters to achieve precise reseeding and fertilization. S56. The operation process is monitored in real time by cameras and radar sensors on the drone to ensure the quality of reseeding and fertilization.

[0029] Step S60 also includes: S61. The encoder using U-Net and DeepLabV3+ has a convolution kernel size of 3×3, a number of convolution kernels of 32, a stride of 1, a feature map padding strategy of "same", and an activation function of the non-linear activation function "ReLU". S62. The pooling type of the encoder pooling layer of U-Net and DeepLabV3+ is max pooling to highlight the edge and texture features in the sparse image. The pooling window is 5×5. The pooling layer achieves dimensionality reduction of the image while retaining the most significant features. S63. The Atrous Spatial Pyramid Pooling module in the DeepLabV3+ model contains four parallel dilated convolutional layers: a 3x3 convolutional kernel with a dilation rate of 6, a 3x3 convolutional kernel with a dilation rate of 12, a 3x3 convolutional kernel with a dilation rate of 18, and a 3x3 convolutional kernel with a dilation rate of 24. This expands the receptive field of the convolutional kernel without increasing the number of parameters. S64. During training, the constructed U-Net and DeepLabV3+ fusion model uses the cross-entropy loss function and the Dice loss function to optimize the model, and selects the Adam and SGD optimizers for model training. S65. The constructed U-Net and DeepLabV3+ fusion model is finally evaluated using IoU (Intersection over Union) and accuracy metrics.

[0030] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

Claims

1. A method for grassland spot detection and pasture reseeding based on UAV multispectral imagery, characterized in that, Includes the following steps: S10. Based on the geographical environment information of the target ranch on that day, plan the multispectral UAV to select a suitable flight route, and select the optimal terrain-following flight altitude, flight speed, forward overlap rate, and lateral overlap rate to ensure the coverage and quality of the multispectral image data. S20. Acquire near-low-altitude image data of grassland pastures through multispectral cameras and transmit them to the ground processing system. Use professional remote sensing software to stitch, correct and analyze the acquired multispectral image data, and extract information on grassland cover and distribution. S30. A deep learning algorithm combining U-Net and DeepLabV3+ is applied to process the image data, enabling the detection and segmentation of grassland patch areas. Information on the degree of bareness, coordinate location, and size of the patches is extracted to generate reseeding prescription data. Data optimization is then performed based on soil conditions and climatic factors. Step S30 also includes: S31. In ENVI 5.3.1, based on the pasture and grassland cover information in the generated N-EVI map, manually create some regions of interest, and use polygonal fine outlines to select some representative empty and healthy areas in the mosaic image. S32. Use the LabelImg tool to create pixel-level classification labels for the selected empty areas, and use OpenCV to crop, scale, and normalize the image to meet the input format of the subsequent model. S33. The selected empty area is used as the original input image of the U-Net and DeepLabV3+ deep learning models. The encoders of U-Net and DeepLabV3+ are used to extract the low-level features of the image, and the feature maps are fused by concatenation and weighted summation in the encoder stage. S34. Construct a cascaded model and use the decoder part of U-Net to upsample the fused feature map to gradually restore the spatial resolution. S35. Concatenate the corresponding low-level feature maps in the U-Net encoder with the feature maps in the decoder to retain more detailed information; S36. Use the decoder part of DeepLabV3+ to further upsample and refine the output of the U-Net decoder; S37. Use the Atrous Spatial Pyramid Pooling module in DeepLabV3+ to capture multi-scale contextual information; S38. In the output layer, multiple convolutional layers are added to the output of the DeepLabV3+ decoder, and the Softmax activation function is applied to output the probability of each pixel belonging to each category, while generating the final spot region segmentation map. S40. Use QGIS Desktop software to extract the multispectral reflectance information of the segmented spots, combine it with the acquired NDVI and EVI information to analyze and map the exposure degree of the spots, and output their coordinate position and area information to generate a reseeding prescription map. The S50 combines soil conditions and climate factors to optimize reseeding data and imports prescription information into a large plant protection drone. The drone carries sowing and fertilization equipment, loads an appropriate amount of seeds and fertilizer, and performs precise reseeding and fertilization.

2. The method for grassland spot detection and pasture reseeding based on UAV multispectral imagery according to claim 1, characterized in that, Step S10 also includes: S11. Use a handheld RTK device to locate the boundary corners of the ranch fence and import the latitude and longitude coordinate data into the DJI Mavic 3 Multispectral Edition drone remote controller; S12. Based on the wind speed, wind direction and atmospheric visibility of the day, modify the forward flight speed and aerial photography altitude, and adjust the forward overlap rate and lateral overlap rate. S13. Plan the drone flight path according to the ranch terrain, and minimize the number of times the drone turns around to increase the proportion of straight flight time.

3. The method for grassland spot detection and pasture reseeding based on UAV multispectral imagery according to claim 1, characterized in that, Step S20 also includes: S21. Before the drone lands, manually control the drone to take multiple photos of the CH-MXDJ255075 P4M diffuse reflector to obtain the radiometric correction parameters of the images. S22. Use Pix4Dmapper to perform alignment checks on the acquired multispectral remote sensing images, and input the spectral correction parameters for the green, red, near-infrared, and red edge bands respectively, and complete the radiometric correction during the stitching process. S23. Use the latitude and longitude coordinates of the corner points obtained by the handheld RTK as the ground control points of the stitched image, and import them into Pix4Dmapper to correct the distortion of the stitched image and improve the image processing accuracy. S24. In ENVI 5.3.1, use the band combination tool to calculate the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), and use the threshold filtering tool to extract the raster attribute values ​​with NDVI values ​​greater than 0.3 and EVI values ​​greater than 0.2, respectively. S25. Assign a weight of 0.6 to the NDVI value and a weight of 0.4 to the EVI value, and then weight and superimpose the two extracted raster blocks to generate an N-EVI map. S26. Calculate the proportion of grid blocks with pixel values ​​greater than 0.2 in the N-EVI map to the total number of grid blocks in the pasture. This will give you the information on the healthy grassland cover of the pasture. The corresponding area is the distribution area of ​​healthy pasture.

4. The method for grassland spot detection and pasture reseeding based on UAV multispectral imagery according to claim 1, characterized in that, Step S40 further includes: S41. Import a total of 7 images, including red, green, near-infrared, red-edge, RGB band mosaic images, NDVI, and EVI band calculation maps, into QGIS Desktop software. Simultaneously import the segmented empty spot areas as separate layers into the software and place them at the top layer. S42. Extract the latitude and longitude coordinates of the pixel points within the segmented area from the geometric field of the vector layer of the empty spot area, and use the field calculator to calculate and output the area of ​​each empty spot based on the latitude and longitude coordinate information; S43. Using the empty spot region layer segmented by the algorithm as the extraction boundary, extract the reflectance of the empty spot regions of 7 groups of images respectively, and perform weighted summation on the mapping values ​​of the empty spot regions. Divide the degree of exposure of the empty spots according to the size of the result. S44. Map the green gradient color depth according to the weighted result size, fill all empty areas in the pasture, and generate a reseeding prescription map.

5. The method for grassland spot detection and pasture reseeding based on UAV multispectral imagery according to claim 1, characterized in that, Step S50 further includes: S51. Collect soil samples using soil testing equipment and analyze key parameters such as soil nutrient content, pH value, and organic matter content. S52. Obtain local temperature, humidity, rainfall, wind speed and climate data through weather stations and meteorological data services; S53. Integrate soil data, meteorological data, and remote sensing imagery data on the degree and location of exposed patches to generate optimized and detailed reseeding and fertilization prescriptions, including information on forage seed type, fertilizer type, and dosage. S54. Import the formulated prescription information into the control system of the plant protection drone, and further set the drone's flight path, altitude, speed, and parameters for sowing and fertilization. S55. Load an appropriate amount of seeds and fertilizer onto the plant protection drone, and the drone will fly according to the preset path and parameters to achieve precise reseeding and fertilization. S56. The operation process is monitored in real time by cameras and radar sensors on the drone to ensure the quality of reseeding and fertilization.

6. The method for grassland spot detection and pasture reseeding based on UAV multispectral imagery according to claim 1, characterized in that, It also includes step S60, which includes: S61. The encoder using U-Net and DeepLabV3+ has a convolution kernel size of 3×3, a number of convolution kernels of 32, a stride of 1, a feature map padding strategy of "same", and an activation function of the non-linear activation function "ReLU". S62. The pooling type of the encoder pooling layer of U-Net and DeepLabV3+ is max pooling to highlight the edge and texture features in the sparse image. The pooling window is 5×5. The pooling layer achieves dimensionality reduction of the image while retaining the most significant features. S63. The Atrous Spatial Pyramid Pooling module in the DeepLabV3+ model contains four parallel dilated convolutional layers: a 3×3 convolutional kernel with a dilation rate of 6, a 3×3 convolutional kernel with a dilation rate of 12, a 3×3 convolutional kernel with a dilation rate of 18, and a 3×3 convolutional kernel with a dilation rate of 24, which expands the receptive field of the convolutional kernel without increasing the number of parameters. S64. During training, the constructed U-Net and DeepLabV3+ fusion model uses the cross-entropy loss function and the Dice loss function to optimize the model, and selects the Adam and SGD optimizers for model training. S65. The constructed U-Net and DeepLabV3+ fusion model is finally evaluated using IoU (Intersection over Union) and accuracy metrics.

Images