A corn drought automatic monitoring method and system based on a large sprinkler

Abstract

This invention discloses an automatic monitoring method and system for corn drought conditions based on a large sprinkler irrigation machine. The invention uses the large sprinkler irrigation machine as a mobile monitoring platform, leveraging its wide travel range and the ease with which its truss structure can accommodate sensors. A 4G camera with FTP functionality and a solar power system are deployed on the sprinkler irrigation machine to automatically acquire corn images. The images are automatically uploaded to a cloud server via a 4G network. A deep learning model integrating EfficientNet / ResNet and an attention mechanism is constructed to automatically identify the images and obtain the drought probability distribution. The identification results are displayed intuitively through a visual webpage, supporting prediction of single or batch images. This method automates the entire process from image acquisition, uploading, analysis to feedback, solving the problems of low efficiency in manual inspections and limited coverage of fixed equipment. It improves the intelligence and real-time performance of corn drought monitoring, providing a scientific basis for precision irrigation.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent monitoring technology, specifically to an automatic monitoring method and system for corn drought conditions based on a large sprinkler irrigation machine. Background Technology Traditional corn field monitoring relies heavily on manual periodic inspections or fixed monitoring equipment. Manual inspections are time-consuming, labor-intensive, and inefficient, making them unsuitable for large-scale agricultural production. Fixed equipment has limited coverage and requires regular maintenance. Large sprinkler irrigation machines, commonly used in modern large-scale farmland irrigation, have wide-ranging travel paths, strong adaptability to different soil types, are easy to operate, and can be remotely controlled via software. Their truss-like structure also facilitates the mounting of sensors, solar panels, cameras, and other equipment. Combining image acquisition with deep learning technology can achieve efficient, automated, and large-scale monitoring of corn growth status. Currently, in the field of crop growth status monitoring, there is no mature system that organically integrates a sprinkler irrigation machine mobile platform, automatic image uploading, cloud-based model inference, and a user interface to achieve real-time intelligent monitoring of corn drought conditions. Summary of the Invention

[0002] Based on the current trend of agricultural informatization and intelligentization, this invention provides an automatic monitoring method and system for corn drought based on a large sprinkler irrigation machine, which realizes automatic acquisition, uploading, intelligent analysis and visualization feedback of corn plant images, improves monitoring efficiency and automation, and can be extended to the field of crop growth monitoring.

[0003] To achieve the above objectives, the present invention provides an automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine, comprising the following steps: S1. Deploy a camera on a large sprinkler irrigation machine to control the sprinkler irrigation machine to move along a preset path and collect images; S2. Automatically upload the acquired images to the specified directory on the cloud server via FTP protocol; S3. Organize the dataset according to the specified path and format, and build and train the deep learning model.

[0004] S4. In the cloud server, drought analysis is performed on the uploaded images based on the constructed deep learning model to obtain prediction results; S5. Output and store the prediction results; S6. Based on the prediction results, launch a visual webpage to realize real-time monitoring of corn drought through a user interface.

[0005] Preferably, S1 includes: Install a 4G camera with FTP transfer capability on the truss of a large sprinkler machine and configure it with timed image capture or motion detection functions. Based on monitoring needs, control the sprinkler machine to move and trigger the camera to collect images of corn.

[0006] Preferably, S2 includes: Enable the FTP service on the cloud server and configure the relevant FTP parameters; The camera establishes an FTP connection with the cloud server via a 4G network and automatically uploads the captured images to a designated directory.

[0007] Preferably, S3 includes: Deep learning models are built based on EfficientNet or ResNet model architectures, combined with attention mechanism modules. The trained model is used to perform forward propagation calculations on the uploaded images to obtain the probability distribution of drought category corresponding to the images.

[0008] Preferably, the steps for building a deep learning model based on the EfficientNet or ResNet model architecture, combined with an attention mechanism module, include: Set the training parameters for the model, including the main model type, attention module type, dataset path, training batch size, and training cycle. Run the training program to train the model and output the trained model weights and configuration file.

[0009] Preferably, S5 includes: The softmax function is used to convert the raw data output by the model into probability distributions for each category; The image name, original path, prediction time, prediction result, and probability of each category are encapsulated into a JSON file and stored in the specified directory.

[0010] Preferably, S6 includes: Run the identification program and provide an API link to access the visualization webpage of the corn drought monitoring system; Import the image to be predicted into the user interface, display the preview image and prediction results, and support single or batch prediction operations.

[0011] This invention also provides an automatic drought monitoring system for maize based on a large sprinkler irrigation machine. The system is used to implement the above method and includes: The data acquisition module consists of cameras deployed on large sprinkler irrigation machines, which acquire images as the sprinkler irrigation machine travels along a preset path. The upload module is used to automatically upload the captured images to a specified directory on the cloud server via the FTP protocol; The build module is used to organize datasets according to specified paths and formats, and to build and train deep learning models.

[0012] The prediction module, located on a cloud server, performs drought analysis on uploaded images based on a constructed deep learning model to obtain prediction results. The output module is used to output and store the prediction results; The interactive module is used to launch a visual webpage based on the prediction results, enabling real-time monitoring of corn drought conditions through a user interface.

[0013] Compared with the prior art, the beneficial effects provided by the present invention are as follows: This invention utilizes a large sprinkler irrigation machine as a mobile monitoring platform. Leveraging its wide travel range and truss structure that easily accommodates sensors, the machine is equipped with a 4G camera with FTP functionality and a solar power system, enabling automatic image acquisition of corn crops. Images are automatically uploaded to a cloud server via a 4G network. A deep learning model integrating EfficientNet / ResNet and attention mechanisms is then used to automatically identify the images and obtain the drought probability distribution. The identification results are displayed intuitively on a webpage, supporting prediction of single or batch images. This method automates the entire process from image acquisition, uploading, analysis to feedback, solving the problems of low efficiency in manual inspections and limited coverage of fixed equipment. It improves the intelligence and real-time performance of corn drought monitoring, providing a scientific basis for precision irrigation. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the EfficientNet-B0 network structure in an embodiment of the present invention; Figure 3 This is a schematic diagram of the MBConv module structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the SEAttention module structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the ECAA attention module structure in an embodiment of the present invention; Figure 6 This is a schematic diagram of the CBAMAttention module structure in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] S1. Deploy cameras on large sprinkler irrigation machines to control the machines to travel along preset paths and collect images.

[0019] The large-scale sprinkler irrigation machine in this embodiment includes, but is not limited to, clockwise sprinkler irrigation machine and horizontal displacement sprinkler irrigation machine. Preparation work in S11-S14 is required before S1 execution: S11: Connect a 4G camera with functions such as FTP transfer, timed image capture, and motion detection (this proposal uses the Hikvision iDS-2DE6C240MW-D / GLT / XM camera) to any computer (with an RJ45 port or a USB-to-Ethernet adapter) via a network cable. Activate the camera according to the manufacturer's instructions, log in to the camera's backend, and set the relevant FTP parameters, including the FTP server IP address, username and password, and file storage location.

[0020] S12: Set timed image capture and motion detection parameters, specify the conditions under which the camera will capture image data, and manually trigger tests.

[0021] S13: Cloud servers or other devices with public IP addresses (i.e., not private IP addresses such as 10., 192., 172., etc.) need to enable FTP transfer services. For easier management, download FileZillaSrever software to configure it, including enabling passive mode, setting account passwords, and setting firewall inbound rules for specified ports.

[0022] S14: When activating the 4G camera, it needs to be bound to the APP provided by the manufacturer. Consult customer service to apply for a physical card and targeted IP address. The targeted IP address is the public IP address of the cloud server. At this time, the camera can communicate with the cloud server via FTP protocol under the 4G network.

[0023] S15: Based on monitoring requirements, install the camera, power supply battery, and solar panel on the truss of the large sprinkler machine (clockwise or horizontal type) and secure them. Connect the camera to the solar panel, etc., to ensure the camera can be successfully connected online. Start the sprinkler machine, and the camera should be adjusted to overhead shooting mode. At this time, the camera will automatically take pictures if the preset conditions are met. Users can also control the camera to collect images of corn growth.

[0024] S2. Automatically upload the images acquired by S1 to the specified directory on the cloud server via FTP protocol.

[0025] In this embodiment, the camera can be set to capture images at set times and detect motion. It can take pictures while the sprinkler is moving or at a specified time each day. The captured images can be automatically uploaded to the central / NAS / FTP server (this requires the camera's own functionality).

[0026] S3. Organize the dataset according to the specified path and format, and build and train the deep learning model.

[0027] The purpose of each subfolder under the S31 and EfficientNet folders is explained as follows: the config folder stores configuration files for the prediction model, including service configuration, model configuration, file monitoring configuration, log configuration, API configuration, etc.; the uploads folder in the data folder is used to receive image data transmitted from the camera (the path setting for FileZillaServer is similar); the deployment folder contains files related to cloud server deployment; the logs folder stores logs recorded by the running program; the models folder stores files related to the trained model; the scripts folder contains script files that control the automated operation of the program; and the training and classifier folders provide the code required to train the model.

[0028] S32. Download Anaconda and PyCharm Community Edition software on your local computer. Use Anaconda to configure a Python environment to run the program. The Python version used in this patent is 3.8.20. The dependencies that need to be installed when training the model are specified in the requirements.txt file.

[0029] S33. Open the EfficientNet folder using PyCharm Community Edition. The training_config.yaml file in the training folder is the model's parameter configuration file. It mainly includes: ① Main model: EfficientNet or ResNet. ② Additional attention mechanism modules. ③ Setting the parameters of the corresponding attention modules, such as the dimensionality reduction ratio of the SE module, and the convolution kernel size of the ECA and CBAM modules. ④ Dataset path and the ratio of training, validation, and test datasets. ⑤ Batch size and training epochs during training. ⑥ Data augmentation configuration. The main part of the model is located in the model_factory.py and model_architecture.py files in the classifier folder. The main structure of the SE, ECA, and CBAM modules is stored in attention_modules.py.

[0030] The training configuration is as follows: training: Epoch: Training period. One epoch trains all training samples once. batch_size: Batch size, the number of data samples processed by the neural network in one training iteration, usually set to a power of 2 (32, 64, etc.). learning_rate: A crucial hyperparameter in deep learning, it controls the pace of model updates during training. It determines the magnitude of parameter adjustments each time the model parameters are updated, and is typically expressed as a percentage. α To represent this. Specifically, given a loss function... ,in It is the set of parameters of the model, and the gradient descent algorithm updates these parameters using the following formula: in, This is the gradient of the loss function with respect to the model parameters. Here, the learning rate... α Directly decided The update range.

[0031] Patience: 10 refers to the level of patience required to handle performance changes during training. When a model's performance does not show significant improvement over a certain number of iterations, we need to be patient enough to wait for performance changes rather than prematurely stopping training to avoid missing further improvements.

[0032] `use_early_stopping: false` indicates whether to use early stopping. Early stopping calculates the model's performance on the validation set during training. When the model's performance on the validation set begins to decline, training is stopped to avoid overfitting caused by continued training.

[0033] weight_decay: 0.0001 (Weight decay, keeping the weights within a small range to avoid overfitting) seed: 42 (random seed, ensuring reproducibility of results).

[0034] S34. After setting the training parameters, run the `train.py` file. The model will first output the model and module information used in this training, the relationship between the dataset categories and the number of files, and the training configuration information. Training will begin. The model will perform convolution to extract low-level features, batch normalization to accelerate training, apply activation functions, max pooling to reduce feature map size, and residual processing on each received image file. Finally, the data will be input into the attention mechanism module to extract important features. During training, the model will output the training and validation loss and accuracy settings for each EPOCH training iteration. At the end of training, the model weights and training data will be saved in .pth and .json formats in the user-defined directory.

[0035] S4. On the cloud server, use the constructed deep learning model to perform drought analysis on the uploaded images and obtain prediction results.

[0036] S41. The main steps are the same as in S32. The dependencies that need to be installed during prediction are specified in the requirements_deploy.txt file.

[0037] S42. After installing PyTorch, NumPy, Watchdog, and other related dependency libraries and setting up the interpreter for the Python program (the Python environment configured in step S31), run the start_service.py file in the scripts folder. An interactive interface will appear in the terminal area below. Execute commands in sequence to install / check Python dependencies, output the project root directory, model path, and create related directories. Then, the user can choose the startup mode, which can be: ① Start full mode (FlaskAPI + folder monitoring); ② WebAPI only; ③ Start folder monitoring only; ④ Development and debug mode.

[0038] S43. When executing the FlaskAPI+ folder monitoring service, the `floder_monitor.py` file first checks if there are any image files uploaded by cameras in the `upload` folder. If new files are found, they are moved to the `processed` folder for processing. The `ModelManager.py` file loads the trained model and configuration files from the `model_path` and `config_path` paths. After identifying the model type (EfficientNet or ResNet), it reconstructs the corresponding model architecture and loads the model training weights to the device (CPU or GPU) running the model. The preprocessing of the input image mainly includes loading the image, resizing (to the required 224×224 pixel resolution), converting to PyTorch tensor format, and normalization. Afterward, the program inputs the processed image tensor into the model for forward propagation and computation.

[0039] S5. Output and store the prediction results.

[0040] The program ultimately outputs the prediction results by applying the softmax function to transform the original output from step S43 into probability distributions for each category, and then using the index of the category with the highest probability as the prediction result. The program encapsulates information such as image name, original path, prediction time, prediction result, and probability of each category in the same JSON file, stored in the result folder under data. If multiple images are predicted in this operation, all prediction results will be aggregated into the same JSON file.

[0041] S6. Based on the prediction results, launch a visualization webpage to achieve real-time monitoring of corn drought conditions through a user interface.

[0042] Step S6 mainly involves running the recognition program. The API link provided on the terminal below can be clicked to access and launch the corn drought monitoring system's visualization webpage. Here, users can view the model information and loading status recognized by the program. Alternatively, users can use the APP provided by the camera manufacturer to manually control the camera to capture images. In the automatic monitoring system, users can select single-image prediction or batch prediction, import the images to be predicted, and the interface will display a preview. Clicking the button below will initiate the prediction, and the bottom window will display the prediction results and the probabilities of each category. The flowchart of this embodiment is as follows: Figure 1 As shown.

[0043] Example 2 This embodiment mainly describes the main model ① in S3 of Embodiment 1.

[0044] Taking ResNet50 as an example, the module insertion position is after layer 4 and before the Global Avg pool; the complete ResNet50 network consists of three parts: the head, the body, and the tail, including: Head: Contains a 7×7 convolutional layer (stride 2) and a 3×3 max pooling layer (stride 2) for initial feature extraction and downsampling, reducing the spatial size of the input image to 1 / 4 of its original size.

[0045] Body: Consists of 4 groups of residual blocks, each group containing multiple Bottleneck residual blocks. Group 1: 3 residual blocks, 256 output channels, no spatial downsampling; The second group consists of 4 residual blocks with 512 output channels and a halved space size. The third group consists of 6 residual blocks, with 1024 output channels and a halved space size. Group 4: 3 residual blocks, 2048 output channels, space size halved; Tail: Contains a global average pooling layer, flattening operation, and fully connected layer, which maps features to several categories (adjusted according to the dataset).

[0046] Furthermore, this embodiment also uses EfficientNet-B0 as an example, such as... Figure 2 As shown. The structure obtained through Neural Architecture Search (B1-B7 are modifications of B0 in Resolution, Channels, and Layers) consists of a series of Stages. The convolutional layers in Table 1 are all followed by Batch Normalization (BN) and the Swish activation function by default. The module insertion position is after the last MBConv module and before the Conv Head. The characters in Table 1 are explained as follows: Resolution: The height and width of the input feature matrix for each stage; Channels: Channels that output the feature matrix after passing through this Stage; Layers: refers to how many times the Operator operation is repeated; stride: refers to the stride of the convolution kernel in the first MBConv of each stage; the default stride for the rest is 1.

[0047] Table 1 The network is divided into 9 stages. The first stage is a regular convolutional layer with a kernel size of 3x3 and a stride of 2 (containing BN and the activation function Swish). Stages 2 to 8 are repeated stacks of MBConv structures. Stage 9 consists of a regular 1x1 convolutional layer (containing BN and the activation function Swish), an average pooling layer, and a fully connected layer.

[0048] Each MBConv in the table is followed by a number 1 or 6. Here, 1 or 6 is the scaling factor n, which means that the first 1x1 convolutional layer in MBConv will expand the channels of the input feature matrix to n times. Here, k3x3 or k5x5 represents the size of the convolutional kernel used by the Depthwise Conv in MBConv.

[0049] like Figure 3 As shown, the MBConv structure mainly consists of a 1x1 ordinary convolution (for dimensionality increase, including BN and Swish), a kxk Depthwise Conv convolution (including BN and Swish). The specific value of k can be found in the EfficientNet-B0 network framework, which mainly has two cases: 3x3 and 5x5. It consists of an SE module, a 1x1 ordinary convolution (for dimensionality reduction, including BN), and a Dropout layer.

[0050] Example 3 This embodiment mainly describes the ② additional attention mechanism module in S3 of Embodiment 1.

[0051] This invention can add SEAttention, ECAAAttention, CBAMAttention, or not add an attention module.

[0052] ①SEAttention module.

[0053] like Figure 4 As shown, weights for each channel are generated using global average pooling, and the importance of each channel is learned through two fully connected layers. These weights are then applied to the feature map to enhance the features of important channels and suppress unimportant features. The specific structure is as follows: Residual: Residual connection; Global Avg Pool (Global Average Pooling, GAP): Global average pooling; Fully Connected: Fully connected layer; Non-linear: Non-linear activation functions (such as ReLU); Sigmoid: Sigmoid activation function; Re-weighting; Squeeze: Performs global average pooling on the input features, compressing the features of each channel into a scalar, capturing global spatial information. The output dimension is C×1×1, where C is the number of channels.

[0054] Excitation: This module uses two fully connected layers and an activation function (ReLU and Sigmoid) to generate weights for each channel. Through these two fully connected layers, the SE module learns the importance of each channel, and the output dimension is (C).

[0055] Recalibration: The generated channel weights are multiplied channel by channel with the original features to recalibrate the channel weights of the feature map.

[0056] Key parameters: Reduction ratio: Channel dimensionality reduction factor (usually set to 16), used to control the number of parameters in the fully connected layer. In the configuration file, this parameter is called reduction_ratio, and the default value is 16.

[0057] ②ECAA attention module.

[0058] The ECAA attention module computes the attention between channels by performing convolution operations on local regions, eliminating the need for fully connected layers, thus reducing computational complexity and improving efficiency. Figure 5 As shown, its structure includes: Input Feature: Input feature (feature map); C, H, W: Number of channels (C), height (H), width (W); GAP (Global Average Pooling): Global average pooling is used to extract global features; Kernel: Convolution kernel (here, a 1D convolution kernel). 1×1×C: Represents the shape of the feature, with both width and height being 1, retaining only the channel dimension; σ(Sigmoid Activation): The Sigmoid activation function is used to generate channel attention weights; Output Feature: The weighted feature map.

[0059] After channel-wide average pooling without dimensionality reduction, ECANet uses one-dimensional convolutions to achieve cross-channel information interaction. The kernel size is adaptively determined by a function, where kernel size k represents the coverage of local cross-channel interaction, i.e., how many neighbors participate in the attention prediction of a channel. This method has been proven to guarantee efficiency and effectiveness. Given the channel dimension C, the kernel size k can be adaptively determined as follows: in γ The usual value is 2. b Take 1 In the configuration file, the kernel size parameter k_size defaults to "auto", meaning it is automatically adjusted by the function.

[0060] ③CBAMAttention module The CBAMAttention module combines channel attention (CA) and spatial attention (SA), dynamically adjusting the importance of different parts of the feature map by modeling channel and spatial attention separately. For example... Figure 6 As shown, the structural composition (channel priority order) is as follows: GAP+GMP (Global Avg Pool + Global Max Pool): Global average pooling + global max pooling; Conv+ReLU: Convolutional layer + ReLU activation function; 1×1Conv: 1×1 convolution; Sigmoid: Sigmoid activation function; Channel Pool: Channel pooling (max pooling + average pooling). 7×7Conv: 7×7 convolution; BN+Sigmoid (Batch Normalization+Sigmoid): Batch Normalization + Sigmoid; Re-weighting; Kernel size k: The size of the convolution kernel used for feature fusion in CBAM spatial attention. The default value is 7. In the configuration file, this parameter is "kernel_size".

[0061] Example 3 This embodiment also provides an automatic drought monitoring system for corn based on a large sprinkler irrigation machine, including: The system comprises several modules: a data acquisition module (using cameras deployed on large sprinkler irrigation machines to capture images as the machines move along a preset path); an upload module (automatically uploading captured images to a specified directory on a cloud server via FTP); a data construction module (organizing the dataset according to a specified path and format, building and training a deep learning model); a prediction module (analyzing the uploaded images based on the constructed deep learning model on the cloud server to obtain prediction results); an output module (outputting and storing the prediction results); and an interaction module (launching a visual webpage based on the prediction results, enabling real-time monitoring of corn drought conditions through a user interface).

[0062] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine, characterized in that, Includes the following steps: S1. Deploy a camera on a large sprinkler irrigation machine to control the sprinkler irrigation machine to move along a preset path and collect images; S2. Automatically upload the acquired images to the specified directory on the cloud server via FTP protocol; S3. Organize the dataset according to the specified path and format, and build and train the deep learning model. S4. In the cloud server, drought analysis is performed on the uploaded images based on the constructed deep learning model to obtain prediction results; S5. Output and store the prediction results; S6. Based on the prediction results, launch a visual webpage to realize real-time monitoring of corn drought through a user interface.

2. The automatic monitoring method for corn drought based on a large sprinkler irrigation machine according to claim 1, characterized in that, S1 includes: Install a 4G camera with FTP transfer capability on the truss of a large sprinkler machine and configure it with timed image capture or motion detection functions. Based on monitoring needs, control the sprinkler machine to move and trigger the camera to collect images of corn.

3. The automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine according to claim 1, characterized in that, S2 includes: Enable the FTP service on the cloud server and configure the relevant FTP parameters; The camera establishes an FTP connection with the cloud server via a 4G network and automatically uploads the captured images to a designated directory.

4. The automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine according to claim 1, characterized in that, S3 includes: Deep learning models are built based on EfficientNet or ResNet model architectures, combined with attention mechanism modules. The trained model is used to perform forward propagation calculations on the uploaded images to obtain the probability distribution of drought category corresponding to the images.

5. The automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine according to claim 4, characterized in that, The steps for building a deep learning model based on the EfficientNet or ResNet model architecture, combined with an attention mechanism module, include: Set the training parameters for the model, including the main model type, attention module type, dataset path, training batch size, and training cycle. Run the training program to train the model and output the trained model weights and configuration file.

6. The automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine according to claim 1, characterized in that, S5 includes: The softmax function is used to convert the raw data output by the model into probability distributions for each category. The image name, original path, prediction time, prediction result, and probability of each category are encapsulated into a JSON file and stored in the specified directory.

7. The automatic monitoring method for corn drought conditions based on a large sprinkler irrigation machine according to claim 1, characterized in that, S6 includes: Run the identification program and provide an API link to access the visualization webpage of the corn drought monitoring system; Import the image to be predicted into the user interface, display the preview image and prediction results, and support single or batch prediction operations.

8. An automatic drought monitoring system for maize based on a large sprinkler irrigation machine, the system being used to implement the method described in any one of claims 1-7, characterized in that, include: The data acquisition module consists of a camera deployed on a large sprinkler irrigation machine, which acquires images as the sprinkler irrigation machine travels along a preset path. The upload module is used to automatically upload the captured images to a specified directory on the cloud server via the FTP protocol; The build module is used to organize datasets according to specified paths and formats, and to build and train deep learning models. The prediction module, located on a cloud server, performs drought analysis on uploaded images based on a constructed deep learning model to obtain prediction results. The output module is used to output and store the prediction results; The interactive module is used to launch a visual webpage based on the prediction results, enabling real-time monitoring of corn drought conditions through a user interface.

Images