Field weed intelligent processing method and system based on unmanned aerial vehicle

Abstract

The invention provides a field weed intelligent processing method based on an unmanned aerial vehicle, which is based on an unmanned aerial vehicle image gridding weed density estimation and quantitative spraying control system and is characterized in that the unmanned aerial vehicle image resolution is improved through a super-resolution enhancement technology (ESRGAN), so that small-scale weed features are clearer, and the weed density is improved. Therefore, the weed identification accuracy of the model is improved. Meanwhile, statistics of the number of weeds in different grids is achieved through a density estimation algorithm, the spraying amount is dynamically adjusted according to an estimation result, and quantitative accurate spraying is achieved. According to the method, the weed monitoring precision is improved, the pesticide consumption is greatly reduced, and the environmental protection benefit and the cost benefit of the system are improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer vision, remote sensing technology and agricultural science, and particularly relates to an intelligent processing method and system for field weeds based on an unmanned aerial vehicle (UAV). Background Art

[0002] With the development of precision agriculture and smart agriculture, how to manage farmland weed problems more efficiently has become one of the key points in current agricultural research. Traditional agricultural weed monitoring and management methods usually rely on manual observation or simple image processing techniques. These methods not only consume a large amount of manpower and time, but also it is difficult to achieve grid-based weed monitoring in large-scale farmland. Therefore, it is impossible to flexibly control the spraying amount of pesticides according to different weed distributions. That is, traditional pesticide spraying equipment generally adopts a uniform spraying method and cannot be dynamically adjusted according to the distribution of weeds. This results in pesticide waste, and it is impossible to effectively avoid the problems of excessive or insufficient pesticides, and it may have a negative impact on the soil and crops in low-density areas. Therefore, there is a lack of a system that can estimate the grid-based weed density based on UAV images and perform precise pesticide spraying control in combination with the weed distribution situation. Summary of the Invention

[0003] Technical Solution: To solve the above technical problems, the present invention provides an intelligent processing method for field weeds based on an unmanned aerial vehicle, and the steps of the method are as follows:

[0004] S01 Image Acquisition and Preprocessing

[0005] Set the flight path, flight altitude and speed of the UAV, capture and obtain images with GPS coordinate information, and then perform preprocessing on the images, including noise removal, color adjustment and contrast optimization;

[0006] S02 Super-Resolution Enhancement

[0007] Construct a model with the ESRGAN architecture, use the images in S01 as the input set and training set, and through the training and production process of the model, output images with a high resolution ratio;

[0008] S03 Grid Division

[0009] Divide the images output in S02 into small blocks, each small block corresponding to a specific area in the farmland. Use the image database to cut the images one by one, and save the corresponding pixel information for each grid block; each cut small block image inherits the GPS data of the original image, and through coordinate transformation, calculate the geographical location and boundary of each grid block; then generate a unique identifier for each image block, and store all grid blocks in the form of a list or dictionary;

[0010] S04 Weed Density Estimation

[0011] Calculate through the weed density estimation model to obtain the weed density map;

[0012] S05 Spraying control

[0013] Characterize the number of weeds in each grid through the density map, calculate the spraying amount; then set the spraying path planning route to complete the synchronous update and feedback record of the spraying data;

[0014] S06 Data recording and analysis

[0015] Store all data after each inspection in S01 - S05, construct a database, and file and store it according to date or region; generate a heat map to display the weed density in different regions, use the inspection data for multiple times to draw a trend chart of the change in weed density, and perform visual display and traceable retrieval of the data.

[0016] As an improvement, in S01, the method of image pre - processing includes adjusting the input original image to a format that conforms to the input size of the model; then performing format conversion, channel processing, and size adaptation processing;

[0017] Among them, the format conversion is to load the image from the image format to a tensor and normalize it to [0, 1];

[0018] If the channel - processed image has a transparent channel Alpha, it will be removed, and only the RGB channels will be retained;

[0019] The size adaptation processing is to crop the large image into multiple overlapping small image blocks in a fixed - block manner and set the required input size of the small - block image.

[0020] As an improvement, in S02, the specific steps for constructing the ESRGAN architecture are as follows:

[0021] S21 Network design of the generator

[0022] The generator adopts the ResNet structure of the residual network, which includes multiple residual blocks and multiple skip connections; each residual block contains a convolutional layer, a batch normalization layer, and an activation layer;

[0023] S22 Network design of the discriminator

[0024] Use the PatchGAN structure. The discriminator extracts features through convolutional layers to distinguish the detailed differences between the generated image and the real image, and continuously optimizes the output of the generator through adversarial training;

[0025] S23 Feature extraction and perceptual loss:

[0026] Using the intermediate layer features of the VGG network, calculate the difference between the generated image and the original image in the feature space through the perceptual loss formula, and then calculate the perceptual loss and pixel loss through the pixel loss formula in the feature extraction layer to evaluate the visual effect of the generated image and optimize the loss function under the generator through the joint loss to optimize the detail quality.

[0027] As an improvement, in S21, the residual block is expressed as H(x) = F(x, {W i}), + x, where H(x): the output of the residual block; F(x, {W i}): the feature transformation after two layers of convolution; W i : the weight of the i-th layer convolution kernel; the specific calculation formula for each layer is: F(x) = ReLU(BN(Conv2D(x))); the skip connection is H final = x + H(x);

[0028] It also includes an upsampling module and a generator output. The upsampling module uses the PixelShuffle upsampling method, expressed as Y upsample = PixelShuffle(X), X: the input feature map; Y upsample : the upsampled feature map;

[0029] The generator output, after passing through multiple residual blocks and the upsampling module, outputs a high-resolution image, expressed as: I high = G(I low ; θ G ); I high For generating a high-resolution image, I low is the input low-resolution image; θ G : the parameters of the generator network.

[0030] 5. The intelligent field weed processing method based on a drone according to claim 3, wherein: in S22, the PatchGAN discriminator is a process of inputting an image, extracting local features through convolution operations, and outputting the authenticity probability of each Patch, expressed as: D(I) = σ(W * I + b),

[0031] where, W: the convolution kernel; *: the convolution operation; b: the bias; σ: the Sigmoid activation function, outputting the authenticity probability.

[0032] 11. As an improvement, in S23, the perceptual loss formula is φ i : the feature extraction of the i-th layer of the VGG network, H i , W i , C i: The height, width, and number of channels of the i-th layer feature map, which are used for normalizing the loss calculation. The expression of the pixel loss formula is: L pixel = ||I high - I real ||1, I high Input high-quality image, I real : Real image;

[0033] The loss function under the joint loss optimization generator is: L total = λ1L GAN + λ2L perceptual + λ3L pixel , where λ1, λ2, λ3 are weight coefficients, and L total : Total loss function, which is used to optimize the overall objective function of the generator, L GAN : Adversarial loss, which is used to measure the adversarial game between the generator and the discriminator, and encourages the generator to generate more realistic images, L perceptual : Perceptual loss, which is used to calculate the difference between the generated image and the real image in the high-level feature space, L pixel : Pixel-level loss, which is used to measure the direct difference between the generated image and the real image in the pixel space, usually mean squared error or L1 loss.

[0034] As an improvement, in S04, before the grid image patch is input into the density estimation model, preprocessing including standardization and normalization is also included; the normalization is to normalize the input pixel values to the range of [0, 1], and the expression is:

[0035]

[0036] where I(i, j, c) is the value of the c-th channel of the i-th and j-th pixels; I max , I min are the minimum and maximum pixel values of the image;

[0037] The standardization is to further convert the normalized pixel values into a zero-mean, unit-variance distribution, and the expression is: where μ is the mean of the image; σ is the standard deviation of the image.

[0038] As an improvement, in S04, in the weed density map, it includes visualizing the density map through an image library and marking different colors for areas with different density sizes.

[0039] As an improvement, in S05, the spraying amount adopts an adaptive spraying method, and the specific calculation formula is: spraying amount = α × weed density, where α is an adjustment parameter based on the pesticide concentration and the weed density.

[0040] Meanwhile, the present invention also provides an intelligent field weed treatment system based on a drone, including a drone terminal, a ground workstation, a cloud, an image acquisition and preprocessing module, a super-resolution enhancement module, an image grid division and positioning module, a weed density estimation module, a quantitative spraying control module, and a data recording and analysis module;

[0041] Information interaction is carried out among the drone terminal, the ground workstation, and the cloud;

[0042] The drone terminal transmits the images and GPS data obtained by the image acquisition and preprocessing module under the drone path to the ground workstation;

[0043] The ground workstation, after enhancing the image resolution through the super-resolution enhancement module for the obtained data; performs image grid division with GPS positioning data through the image grid division and positioning module; then outputs a weed density map and the number of weeds through the weed density estimation module; and then outputs a spraying instruction and the spraying GPS path to the cloud through the quantitative spraying control module;

[0044] The cloud records, stores, and analyzes all the data of the drone terminal and the ground workstation, and then performs visual output, data analysis reports, and traceable retrieval processing.

[0045] Beneficial effects: The method proposed by the present invention is based on a grid-based weed density estimation and quantitative spraying control system for drone images. The super-resolution enhancement technology (ESRGAN) is used to improve the resolution of drone images, making the small-scale weed features clearer, thereby improving the recognition accuracy of the model for weeds. At the same time, the density estimation algorithm is also used to count the number of weeds in different grids, and the spraying amount is dynamically adjusted according to the estimation results to achieve quantitative and precise spraying. This method not only improves the accuracy of weed monitoring but also significantly reduces the amount of pesticide used, enhancing the environmental and cost benefits of the system.

[0046] Compared with conventional technologies, the present invention has the following advantages:

[0047] (1) Improve the pesticide use efficiency: The system adjusts the spraying amount according to the weed density of each grid, avoiding the waste of pesticides in the traditional uniform spraying method and reducing the production cost.

[0048] (2) Enhance the accuracy of weed estimation: Using the super-resolution enhancement module to improve the image resolution and combining with CBAM to increase the model's attention to weeds, the system can more accurately count the number of weeds, especially in areas with high-density weeds, improving the recognition accuracy.

[0049] (3) Reduce environmental pollution: Reduce unnecessary pesticide use, protect the soil and water sources, and reduce the negative impact on the environment.

[0050] (4) Improve the efficiency of farmland management: The system can achieve fully automated management of large-scale farmland, greatly improving the efficiency of field management and supporting the development of precision agriculture. Description of the Drawings

[0051] Figure 1 It is the technical roadmap of the estimation method of the present invention.

[0052] Figure 2 It is the module schematic structure diagram of the intelligent processing system of the present invention.

[0053] Figure 3 It is the schematic diagram of the original image of a part of the present invention.

[0054] Figure 4 It is the schematic diagram of the high-resolution image of a part of the present invention.

[0055] Figure 5 It is the schematic diagram of the farmland weed grid and the visualization result diagram of the UAV path planning in Embodiment 2 of the present invention. Detailed Embodiments

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below, so that those skilled in the art can better understand the advantages and features of the present invention, and thus more clearly define the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts belong to the protection scope of the present invention.

[0057] See Figure 1 As shown, it is the technical roadmap of the intelligent processing method for field weeds based on UAVs of the present invention, including multiple parts. The specific steps are as follows:

[0058] I. Image acquisition and preprocessing

[0059] 1. UAV inspection and image acquisition

[0060] (1) Flight planning: The UAV covers different areas of the farmland according to the preset flight path. The flight height and speed are set through the ground control system to ensure the clarity of the images and the comprehensiveness of the coverage.

[0061] (2) Image acquisition parameters: The camera parameters (such as resolution, exposure time) are dynamically adjusted according to the environmental light conditions to ensure clear details of the weeds.

[0062] (3) GPS positioning: Each image captured by the UAV camera carries accurate GPS coordinate information, which helps the subsequent grid division module to align with the actual geographical location.

[0063] 2. Image preprocessing

[0064] (1) Noise reduction processing: For the high-frequency noise in the image, denoising algorithms such as Gaussian filtering and bilateral filtering are used to reduce noise interference, making the edges of weeds clearer.

[0065] (2) Color adjustment and contrast optimization: Under different lighting conditions, the image color may deviate. Through white balance adjustment and contrast enhancement, the image has a consistent visual effect in different regions.

[0066] II. Super-resolution enhancement (ESRGAN)

[0067] 1. ESRGAN architecture:

[0068] (1) Generator network design: The generator adopts the residual network (ResNet) structure, including multiple skip connections, so that the edge and detail information of the image can be retained during the resolution improvement process.

[0069] (2) Discriminator network design: The discriminator extracts features through convolutional layers to distinguish the detail differences between the generated image and the real image, and continuously optimizes the output of the generator through adversarial training.

[0070] (3) Feature extraction and perceptual loss: The perceptual loss is calculated using the feature extraction layer of the VGG network to evaluate the visual effect and detail quality of the generated image, ensuring that the super-resolution image meets the actual application requirements.

[0071] 2. Training and generation process:

[0072] (1) The ESRGAN model is pre-trained with a large number of farmland images to ensure that the model can accurately reconstruct high-resolution images in the farmland weed scenario.

[0073] (2) The generator network enlarges the input image to four times the resolution and outputs a high-resolution image.

[0074] (3) The finally output high-resolution image is passed to the grid division module to ensure that the details of small-scale weeds are not missed in the subsequent density estimation.

[0075] III. Grid division

[0076] 1. Image cutting and grid division

[0077] The high-resolution image is segmented into several small pieces according to the set size, and each small piece corresponds to a specific area in the farmland (such as 1x1 meter). Using image processing libraries such as OpenCV or Pillow, the image is cut one by one, and the corresponding pixel information is saved for each grid block.

[0078] 2. GPS Data Mapping

[0079] Each cut small piece of the image inherits the GPS data of the original image. Through coordinate transformation, the geographical location and boundaries of each grid block are calculated. The geographical information is stored together with the image data for precise positioning during subsequent spraying control.

[0080] 3. Image Block Encoding

[0081] Generate a unique identifier (such as a number) for each image block for easy data tracking and identification. Store all grid blocks in the form of a list or dictionary to provide an efficient data structure for subsequent batch processing.

[0082] IV. Weed Density Estimation

[0083] 1. Model Input Preparation

[0084] Each grid image block undergoes preprocessing (standardization, normalization) to ensure that the input format meets the requirements of the density estimation model (CombinedUNet). To improve computational efficiency, image blocks are processed in batches when input, and several image blocks are packed and input into the model.

[0085] 2. CombinedUNet Structure

[0086] (1) Global Branch: The global branch extracts the overall scene information and captures the overall distribution pattern of weeds through deep convolutional layers.

[0087] (2) Local Branch: The local branch focuses on details and obtains the features of small-scale weeds through shallower convolutional layers, enabling the model to accurately count even in complex environments.

[0088] (3) CBAM Attention Mechanism: CBAM adds channel and spatial attention layers to the model encoder to enhance the model's attention to key weed regions, making the recognition effect more accurate in dense areas.

[0089] 3. Generate Density Map

[0090] The model outputs the weed density map for each grid. The pixel value in the density map represents the number of weeds in that area. The density map is visualized through the Python image library, with high-density areas shown in bright colors and low-density areas shown in dark colors for convenient calculation of subsequent spraying amounts.

[0091] V. Spraying Control

[0092] 1. Spraying Amount Calculation

[0093] By calculating the number of weeds in each grid of the density map, a spraying volume distribution map is generated based on the weed density. An adaptive spraying volume formula is used, for example: spraying volume = α × weed density, where α is an adjustment parameter based on the pesticide concentration and weed density, and the spraying volume of different grids is dynamically adjusted.

[0094] 2. Spraying Path Planning and Equipment Control

[0095] The spraying equipment plans the spraying path according to the GPS data to ensure precise spraying within the specified grid area. The spraying system uses an automatic control module to adjust the opening and closing of the nozzles, combined with pressure control and flow regulation, to ensure that the pesticide distribution meets the requirements of the weed density map.

[0096] 3. Spraying Data Synchronization and Feedback

[0097] After each spraying, the equipment records the spraying volume, time, and location of each area and uploads them to the cloud database to provide support for subsequent data analysis.

[0098] VI. Data Recording and Analysis

[0099] 1. Data Storage and Archiving: After each inspection, the system stores all data (such as image block numbers, weed density, spraying volume) in the database for easy retrieval in the future. The historical data of each grid block includes weed density, spraying volume, GPS location, etc., and is archived by date or area.

[0100] 2. Data Analysis and Visualization:

[0101] (1) Weed Distribution Heat Map: A heat map is generated to show the weed density in different areas, helping users visually identify high-density areas.

[0102] (2) Weed Growth Trend Analysis: By using the data from multiple inspections, a change trend graph of weed density is drawn to assist in agricultural decision-making.

[0103] Spraying Effect Evaluation: Analyze the change in weed density, evaluate the spraying effect, and optimize future spraying strategies and pesticide usage.

[0104] Example 2

[0105] This example is based on the estimation of the number of weeds in multiple grids and the path planning of drones, and details how to plan the path of drones and perform precise spraying.

[0106] Farmland Area: 20m × 30m.

[0107] Grid Division: The farmland is divided into 20 × 20 = 400 grids according to each grid covering 1m × 1.5m.

[0108] Data source: High-resolution images are collected by drones and grid cutting is performed to generate 400 image patches, with each patch corresponding to a grid.

[0109] Objective: Based on the estimated weed density of each grid, plan the spraying path of the drone to achieve precise spraying.

[0110] 1. Data input

[0111] 1.1 Weed density estimation:

[0112] Use the CombinedUNet model to generate a density map for each grid image patch. Use the weed quantity calculation formula: Accumulate the pixel values of the density map for each grid to obtain the weed quantity of each grid.

[0113] 1.2 Example input data

[0114]

[0115] Grid classification criteria:

[0116] High-density area: Weed quantity > 3000.

[0117] Medium-density area: 1000 ≤ weed quantity ≤ 3000.

[0118] Low-density area: Weed quantity < 1000.

[0119] Spraying parameter setting Spraying quantity formula: Spraying quantity = α × weed density

[0120] 2. Path planning

[0121] 2.1 Planning principles

[0122] (1) Prioritize high-density areas: First cover the grids in high-density areas.

[0123] (2) Path optimization: Under the priority order, adopt the nearest neighbor path optimization algorithm to reduce flight time.

[0124] (3) Dynamically adjust the spraying quantity: Adjust the nozzle flow rate in real time according to the grid classification.

[0125] 2.2 Path planning algorithm

[0126] Classify and sort all grids according to density (high density > medium density > low density). For the grids within each classification, use the nearest neighbor path algorithm to plan the flight order. Combine the paths of all classifications to generate a complete flight route.

[0127] 1. Basic idea of the algorithm

[0128] Starting point selection: Start from a specified starting point (usually the initial position of the drone or a certain grid).

[0129] Nearest neighbor rule: At each step, starting from the current point, select the unvisited point with the shortest distance.

[0130] Loop to construct the path:

[0131] Repeat the above steps until all points have been visited.

[0132] Return to the starting point (optional):

[0133] If a closed-loop path needs to be formed, return to the starting point at the end.

[0134] 2.3 Example path planning

[0135] The starting point of the drone is the GPS coordinate (30.12340, 120.54310), and the planned path is:

[0136] High-density area:

[0137] Grid numbers: 3 → 7 → 10.

[0138] Path: (30.12365, 120.54330) → (30.12370, 120.54350) → (30.12380, 120.54360).

[0139] Medium-density area:

[0140] Grid numbers: 2 → 4 → 6.

[0141] Path: (30.12355, 120.54325) → (30.12375, 120.54335) → (30.12365, 120.54340).

[0142] Low-density area:

[0143] Grid numbers: 1 → 5 → 8.

[0144] Path: (30.12345, 120.54321) → (30.12360, 120.54327) → (30.12385, 120.54342).

[0145] Nearest neighbor path optimization process

[0146] Step 1: Initialization

[0147] Current point: p start =(30.12340, 120.54310)

[0148] Set of unvisited points: punvisited

[0149] ={(30.12365, 120.54330), (30.12370, 120.54350), (30.12380, 120.54360), (30.12355, 120.54325), (30.12375, 120.54335), (30.12365, 120.54340), (30.12345, 120.54321), (30.12360, 120.54327), (30.12385, 120.54342)}

[0150] Step 2: Select the nearest point

[0151] Starting from p start calculate the distance to each point:

[0152]

[0153] Repeat the above calculation to find the nearest point.

[0154] Step 3: Path iteration

[0155] High-density area:

[0156] Find the nearest points in sequence:

[0157] From (30.12340, 120.54310) to (30.12365, 120.54330).

[0158] Then from (30.12365, 120.54330) to (30.12370, 120.54350).

[0159] Finally from (30.12370, 120.54350) to (30.12380, 120.54360).

[0160] The path is:

[0161] (30.12340, 120.54310) → (30.12365, 120.54330) → (30.12370, 120.54350) → (30.12380, 120.54360).

[0162] Medium-density area:

[0163] Starting from the current point (30.12380, 120.54360):

[0164] The nearest point is (30.12375, 120.54335).

[0165] Then to (30.12365, 120.54340).

[0166] Finally to (30.12355, 120.54325).

[0167] The path is:

[0168] (30.12380, 120.54360) → (30.12375, 120.54335) → (30.12365, 120.54340) → (30.12355, 120.54325).

[0169] Low - density area:

[0170] Starting from the current point (30.12355, 120.54325):

[0171] The nearest point is (30.12345, 120.54321).

[0172] Then to (30.12360, 120.54327).

[0173] Finally to (30.12385, 120.54342).

[0174] The path is:

[0175] (30.12355, 120.54325) → (30.12345, 120.54321) → (30.12360, 120.54327) → (30.12385, 120.54342).

[0176] The full path merged is:

[0177] (30.12340, 120.54310) → (30.12365, 120.54330) → (30.12370, 120.54350) → (30.12380, 120.54360) → (30.12375, 120.54335) → (30.12365, 120.54340) → (30.12355, 120.54325) → (30.12345, 120.54321) → (30.12360, 120.54327) → (30.12385, 120.54342)

[0178] 3. Precise spraying control

[0179] 3.1 Spraying amount calculation

[0180] For each grid, calculate the required spraying amount

[0181]

[0182] Dynamic spraying control

[0183] When the UAV enters the target grid area, classify according to the weed density:

[0184] High-density area: maximum nozzle flow rate to ensure coverage.

[0185] Medium-density area: medium flow rate.

[0186] Low-density area: reduce the flow rate to reduce pesticide waste.

[0187] Spraying path coverage:

[0188] The UAV follows the grid path from left to right.

[0189] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. An intelligent processing method for field weeds based on drones, characterized in that: The steps of the method are as follows: S01 Image acquisition and preprocessing Set the flight path, flight altitude and speed of the drone, capture images with GPS coordinate information, and then perform preprocessing on the images, including denoising, color adjustment and contrast optimization; S02 Super-resolution enhancement Build a model with the ESRGAN architecture. Use the images in S01 as the input set and training set, and through the training and production process of the model, output images with a high resolution ratio; S03 Grid division Divide the images output in S02 into small blocks, each small block corresponding to a specific area in the farmland. Use the image database to cut the images one by one, and save the corresponding pixel information for each grid block; each cut small block image inherits the GPS data of the original image. Through coordinate transformation, calculate the geographical location and boundary of each grid block; then generate a unique identifier for each image block, and store all grid blocks in the form of a list or dictionary; S04 Weed density estimation Calculate through the weed density estimation model to obtain a weed density map; S05 Spraying control Characterize the number of weeds in each grid through the density map, calculate the spraying amount; then set the spraying path planning route to complete the synchronous update and feedback record of the spraying data; S06 Data recording and analysis Store all data after each inspection in S01-S05, build a database, and archive and store it according to date or region; by generating a heat map to display the weed density in different regions, use the inspection data of multiple times to draw a trend chart of the change in weed density for visual display and traceable retrieval of the data.

2. The intelligent field weed treatment method based on an unmanned aerial vehicle according to claim 1, wherein: In S01, the method of image preprocessing includes adjusting the input original image to a format that conforms to the input size of the model; then performing format conversion, channel processing and size adaptation processing; Among them, the format conversion is to load the image from the image format to a tensor and normalize it to [0,1]; If the channel-processed image has a transparent channel Alpha, it will be removed, and only the RGB channels will be retained; The size adaptation processing is to crop the large image into multiple overlapping small image blocks in a fixed block manner, and set the required input size of the small image blocks.

3. The intelligent field weed treatment method based on an unmanned aerial vehicle according to claim 1, characterized in that: In S02, the specific steps for building the ESRGAN architecture are as follows: S21 Network design of the generator The generator adopts the ResNet structure of the residual network, which includes multiple residual blocks and multiple skip connections; each residual block includes a convolutional layer, a batch normalization layer and an activation layer; S22 Network design of the discriminator Use the PatchGAN structure. The discriminator extracts features through convolutional layers to distinguish the detailed differences between the generated image and the real image, and continuously optimizes the output of the generator through adversarial training; S23 Feature extraction and perceptual loss: Use the intermediate layer features of the VGG network, calculate the difference between the generated image and the original image in the feature space through the perceptual loss formula, and then calculate the perceptual loss and pixel loss through the pixel loss formula in the feature extraction layer to evaluate the visual effect of the generated image and optimize the loss function under the generator through the joint loss to optimize the detail quality.

4. The intelligent field weed treatment method based on an unmanned aerial vehicle according to claim 3, wherein: In S21, the residual block is expressed as H(x) = F(x, {W i}) + x, where H(x): the output of the residual block; F(x, {W i}): the feature transformation after two-layer convolution; W i : the weight of the convolutional kernel in the i-th layer; the specific calculation formula for each layer is: F(x) = ReLU(BN(Conv2D(x))); the skip connection is H final = x + H(x); It further includes an upsampling module and the generator output. The upsampling module uses the PixelShuffle upsampling method, denoted as Y upsample = PixelShuffle(X), where X is the input feature map; Y upsample : the feature map after upsampling; The output of the generator, after passing through multiple residual blocks and upsampling modules, outputs a high-resolution image, denoted as: I high = G(I low ; θ G ); I high To generate a high-resolution image, I low is the input low-resolution image; θ G : the parameters of the generator network.

5. The intelligent field weed treatment method based on an unmanned aerial vehicle according to claim 3, wherein: In S22, the PatchGAN discriminator takes the input image, extracts local features through convolutional operations, and outputs the authenticity probability of each patch, which is expressed as: D(I) = σ(W * I + b), where W: convolutional kernel; *: convolutional operation; b: bias; σ: Sigmoid activation function, which outputs the authenticity probability.

6. The intelligent processing method for field weeds based on an unmanned aerial vehicle according to claim 3, characterized in that: In S23, the perceptual loss formula is φ i : Feature extraction of the i-th layer of the VGG network, H i 、W i 、C i : The height, width, and number of channels of the i-th layer feature map, used for normalized loss calculation. The expression of the pixel loss formula is: L pixel =||I high -I real ||1, I high The input high-quality image, I real : The real image; The loss function under the combined loss optimization generator is: L total = λ1L GAN + λ2L perceptual + λ3L pixel , where λ1, λ2, and λ3 are weight coefficients, and L total : the total loss function; L GAN : the adversarial loss function; L perceptual : the perceptual loss function; L pixel : the pixel-level loss function.

7. The intelligent field weed treatment method based on a drone according to claim 1, wherein: In S04, before the grid image patch is input into the density estimation model, preprocessing including standardization and normalization is also performed; the normalization normalizes the input pixel values to the range of [0, 1], and the expression is: where I(i, j, c) is the value of the c-th channel of the i-th and j-th pixels; I max , I min are the minimum and maximum pixel values of the image; The standardization further converts the normalized pixel values into a distribution with zero mean and unit variance, and the expression is: where μ is the mean of the image; σ is the standard deviation of the image.

8. The intelligent field weed processing method based on an unmanned aerial vehicle according to claim 1, wherein: In S04, in the weed density map, it includes visualizing the density map through an image library and identifying different colors for regions with different density sizes.

9. The intelligent field weed processing method based on an unmanned aerial vehicle according to claim 1, characterized in that: In S05, the spraying amount adopts an adaptive spraying method, and the specific calculation formula is: spraying amount = α × weed density, where α is an adjustment parameter based on the pesticide concentration and weed density.

10. An intelligent field weed processing system based on a drone, characterized in that it includes a drone terminal, a ground workstation, a cloud, an image acquisition and preprocessing module, a super-resolution enhancement module, an image grid division and positioning module, a weed density estimation module, a quantitative spraying control module, and a data recording and analysis module; Information interaction is carried out among the drone terminal, the ground workstation, and the cloud; The drone terminal transmits the image and GPS data obtained by the image acquisition and preprocessing module under the drone path to the ground workstation; The ground workstation, after enhancing the image resolution through the super-resolution enhancement module for the obtained data; performs image grid division with GPS positioning data through the image grid division and positioning module; then outputs the weed density map and the number of weeds through the weed density estimation module; and then outputs the spraying instruction and the GPS path of spraying to the cloud through the quantitative spraying control module; The cloud performs visualization output, data analysis report, and traceable retrieval processing after recording, storing, and analyzing all the data of the drone terminal and the ground workstation.

Images