Cotton drought identification management system based on image identification

Abstract

The invention discloses a cotton drought identification management system based on image identification, and relates to the technical field of cotton drought management systems. The system comprises an image acquisition module, a data transmission module and an image recognition and processing module, an improved convolutional neural network model is adopted to process and analyze acquired images, then a trained convolutional neural network model is utilized to perform feature extraction and classification on the preprocessed images, and whether cotton is in a drought state or not is judged; the drought evaluation and decision-making module is used for establishing a drought evaluation model by adopting a fuzzy comprehensive evaluation method according to the result of the image recognition and processing module in combination with the growth stage of the cotton and the soil humidity information and evaluating the drought degree of the cotton, and the user interaction module is used for providing a visual operation interface for a user. The system can efficiently and accurately identify and manage the drought of the cotton.

Description

Technical Field

[0001] The present invention relates to the technical field of cotton drought management systems, and in particular, to a cotton drought recognition and management system based on image recognition. Background Art

[0002] As an important cash crop, the growth and development of cotton have relatively strict requirements for water conditions. During the cotton planting process, drought seriously affects the yield and quality of cotton. Traditional drought monitoring methods mostly rely on manual on-site inspections in the fields, and the degree of cotton drought is judged based on experience. This method not only consumes a large amount of manpower and time, but also the monitoring range is limited to the areas accessible to personnel, with strong subjectivity, and it is impossible to timely and accurately grasp the drought situation changes in large-scale cotton planting areas. With the development of technology, although there are monitoring means such as satellite remote sensing and unmanned aerial vehicle (UAV) remote sensing, the resolution of satellite remote sensing is difficult to meet the requirements of fine farmland monitoring, and UAV remote sensing is easily restricted by weather conditions, resulting in inaccurate drought recognition and monitoring. Summary of the Invention

[0003] The technical problem to be solved by the present invention is how to provide a cotton drought recognition and management system based on image recognition that can be efficient, accurate and less affected by the environment.

[0004] To solve the above technical problem, the technical solution adopted by the present invention is: a cotton drought recognition and management system based on image recognition, the system includes: Image acquisition module: A high-definition camera and a multispectral camera carried by a UAV are used to acquire images of the cotton planting area at different time periods and different heights; the high-definition camera is installed at a fixed position in the cotton planting area to obtain the cotton growth images near the ground and the soil surface conditions; the multispectral camera carried by the UAV is used to obtain the reflection information of the cotton plants and the surrounding environment in different bands; Data transmission module: Transmit the image data obtained by the image acquisition module to the image recognition and processing module through a wireless transmission network for processing; Image recognition and processing module: An improved convolutional neural network model is used to process and analyze the acquired images, and identify the growth state, leaf color and texture features of cotton; first, the images are preprocessed, and then the trained convolutional neural network model is used to extract features and classify the preprocessed images to determine whether the cotton is in a drought state; Drought assessment and decision-making module: According to the results of the image recognition and processing module, combined with the growth stage of cotton and the soil humidity information, a drought assessment model is established by using the fuzzy comprehensive evaluation method to evaluate the degree of cotton drought, and the drought is divided into three levels: mild drought, moderate drought and severe drought, and according to the evaluation results, corresponding irrigation decision-making suggestions are provided for users; User interaction module: used to provide users with a visual operation interface, through which users can view the growth status of cotton, drought assessment results and irrigation decision suggestions; at the same time, users can set system parameters on the interface to manage the system.

[0005] The beneficial effect of adopting the above technical solution is that the system can realize real-time image acquisition and data transmission, and with efficient algorithm processing, it can complete drought identification and assessment in a short time, buying time for growers to fight drought. The image acquisition method that combines ground and air not only covers a large area of ​​cotton planting areas, but also takes into account the details of local plants, and can realize comprehensive monitoring of cotton fields of different sizes. A large number of image features are learned and analyzed using deep learning algorithms, and the drought assessment model combined with multi-source data fusion can accurately judge the degree of cotton drought. Scientific irrigation decision recommendations are automatically generated based on the drought assessment results, without the need for growers to have professional agricultural knowledge, reducing the difficulty and cost of decision-making, and improving the level of intelligence in agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0007] Figure 1 is a principle block diagram of the system according to an embodiment of the present invention; Figure 2 is a principle block diagram of an image acquisition module in the system according to an embodiment of the present invention; Figure 3 is a principle block diagram of a data transmission module in the system according to an embodiment of the present invention; Figure 4 is a flow chart for constructing a convolutional neural network model in the system according to an embodiment of the present invention; Figure 5 It is a flow chart for implementing the drought assessment and decision-making module in the system described in the embodiment of the present invention. DETAILED DESCRIPTION

[0008] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0009] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0010] Generally, as Figure 1 shown, the embodiment of the present invention discloses a cotton drought recognition and management system based on image recognition. The system includes: Image acquisition module 101: A high-definition camera and a multispectral camera carried by a drone are used to acquire images of the cotton planting area at different time periods and different heights. The high-definition camera is installed at a fixed position in the cotton planting area to obtain the cotton growth images near the ground and the soil surface conditions. The multispectral camera carried by the drone is used to obtain the reflection information of the cotton plants and the surrounding environment in different bands; Data transmission module 102: Transmits the image data acquired by the image acquisition module to the image recognition and processing module for processing through a wireless transmission network (WiFi, 4G, and 5G); Image recognition and processing module 103: An improved convolutional neural network model is used to process and analyze the acquired images, and to recognize the growth state, leaf color, and texture features of the cotton. First, the images are preprocessed, and then the trained convolutional neural network model is used to extract features and classify the preprocessed images to determine whether the cotton is in a drought state; Drought assessment and decision-making module 104: According to the results of the image recognition and processing module, combined with the growth stage of the cotton and the soil humidity information, a drought assessment model is established using the fuzzy comprehensive evaluation method to evaluate the degree of drought of the cotton. The drought is divided into three levels: mild drought, moderate drought, and severe drought. And according to the evaluation results, corresponding irrigation decision-making suggestions are provided for users; User interaction module 105: Used to provide a visual operation interface for users. Users can view the growth state of the cotton, the drought assessment results, and the irrigation decision-making suggestions through this interface. At the same time, users can set system parameters on the interface to manage the system.

[0011] Furthermore, as Figure 2 shown, the image acquisition module 101 includes: Camera Image Acquisition Unit 1011: An HD camera is adopted. According to the terrain, area, and layout of the cotton planting area, the installation positions of the HD cameras are determined and distributed at the boundaries and central positions of the planting area to cover the entire area without obvious monitoring dead angles. The installation height is between 1 - 2 meters to enable clear shooting of cotton plants and the soil surface. An HD camera with high resolution (such as 4K and above), wide dynamic range, and good low-light performance is selected, which can clearly capture the images of cotton growth near the ground and the soil surface conditions under different lighting conditions; according to the growth cycle and lighting characteristics of cotton, different image acquisition time periods are set; for the images acquired by the HD camera, median filtering or Gaussian filtering is used for filtering, and then an image enhancement algorithm is used to enhance the acquired images; according to the growth cycle and lighting characteristics of cotton, different image acquisition time periods are set. During the vigorous growth period of cotton, image acquisition is carried out at 10 - 11 am and 3 - 4 pm every day. At this time, the lighting conditions are better, and the morphological characteristics of cotton can be clearly displayed. At the same time, acquisition can also be set under different weather conditions, such as sunny days and cloudy days, to obtain more comprehensive image data.

[0012] Multispectral Camera Image Acquisition Unit 1012: A drone with stable flight performance, long endurance, and high-precision positioning function is selected. The multispectral camera has the detection ability of multiple specific bands (such as blue light, green light, red light, near-infrared, etc.) to obtain the reflection information of cotton plants and the surrounding environment under different bands. The multispectral camera is installed on the stable gimbal system of the drone to ensure the stability of the camera during flight; using drone route planning software, according to the shape and size of the cotton planting area, the flight route is planned, and the flight height is set between 30 - 100 meters. The route planning has an overlap degree of 60% - 80% between adjacent images for subsequent image stitching and processing; select a time period with clear weather and uniform lighting for flight acquisition, and avoid flying under cloudy occlusion or when the light is too strong or too weak. The flight frequency is determined according to the growth stage of cotton and the demand for drought monitoring. During the critical growth period of cotton, fly 1 - 2 times a week.

[0013] Furthermore, as Figure 3 shown, the data transmission module 102 includes: Data Format Conversion Unit 1021: It is used to uniformly convert the image data in different formats collected by the image acquisition module into a standard format. Since the image data obtained by the image acquisition module may have multiple original formats, for example, the images collected by a high-definition camera may be common formats such as JPEG and PNG, while the data collected by a multi-spectral camera carried by a drone may be a specific scientific data format. To facilitate subsequent transmission and processing, it is necessary to uniformly convert these different formats of image data into a standard format.

[0014] Data Compression Unit 1022: Since image data usually occupies a large storage space and bandwidth, in order to improve the transmission efficiency, it is necessary to compress the image data. Lossless compression or lossy compression methods can be used. Lossless compression is suitable for situations where high image quality is required, such as compressing using the ZIP format; lossy compression can sacrifice image quality to a certain extent in exchange for a higher compression ratio, such as the compression of the JPEG format. The corresponding compression method can be selected according to the specific form of the image. Wireless Transmission Network Selection Unit 1023: Selects a wireless transmission network according to the location and environmental conditions of the image acquisition module. Common wireless transmission networks include Wi-Fi, 4G / 5G networks, etc. If the image acquisition device is in a fixed position and there is an available Wi-Fi network nearby, the Wi-Fi network is preferentially selected for transmission. If the acquisition device is in a remote area or requires mobile acquisition, the 4G or 5G network is used for data transmission. Network Connection Configuration Unit 1024: For the Wi-Fi network, configures the Wi-Fi connection parameters of the image acquisition device, including the SSID and password; for the 4G or 5G network, inserts the corresponding SIM card and configures the network connection parameters of the device. After the device is powered on, it automatically searches for and connects to the available network. Data Encapsulation Unit 1025: It is used to encapsulate the compressed image data into data packets suitable for the wireless transmission network. The data packet includes two parts: a data header and a data payload. The data header includes the source address, destination address, data packet length, and check information of the data packet; the data payload includes the actual image data. Data Transmission Unit 1026: It is used to send the encapsulated data packets to the server where the image recognition and processing module is located through the wireless transmission network. In order to ensure the integrity of the data during the transmission process, a retransmission mechanism and an error checking mechanism can be adopted.

[0015] Data Reception Unit 1027: The server where the image recognition and processing module is located listens on a specified port and waits to receive data packets from the image acquisition module. When a data packet is received, it is stored in a temporary buffer. Data verification unit 1028: It is used to verify the received data packets, check the integrity and accuracy of the data. By calculating the checksum of the data packets and comparing it with the check information in the data header, if the checksum does not match, it is considered that an error has occurred during data transmission, and the image acquisition module is requested to resend the data; Data decompression unit 1029: It decompresses the compressed data packets at the receiving end and uses corresponding decompression methods according to different compression algorithms; Data delivery unit 10210: It is used to deliver the decompressed image data to the image recognition and processing module for subsequent processing.

[0016] After being processed by the above units, the system described in this application can transmit the image data obtained by the image acquisition module to the image recognition and processing module for processing safely and efficiently through the wireless transmission network.

[0017] Further, as Figure 4 shown, the method for introducing residual connections and attention mechanisms into the image recognition and processing module to construct an improved convolutional neural network model includes the following steps: Stacking convolutional layers and residual blocks: After the input layer of the convolutional neural network, convolutional layers and residual blocks are alternately used for feature extraction. Among them, the convolutional layer is used to initially extract the local features of the image, and the residual block is used to deepen the network depth through residual connections and learn more complex features; Inserting an attention module: After the residual block or convolutional layer, a channel attention SE module is inserted to perform channel attention adjustment on the extracted features and enhance the expression of important features; Setting up a pooling layer and a fully connected layer: After the convolutional layer and the residual block, a pooling layer is used for downsampling to reduce the size of the feature map; finally, through the fully connected layer, the features are mapped to the final classification result for judging whether the cotton is in a drought state.

[0018] Further, the stacking of the convolutional layer and the residual block includes the following steps: 1) Defining a residual block: Let the input feature map be , and after a series of convolutional operations, the output is obtained; The output of the residual block is calculated by the following formula: The input is directly added to the output of the convolutional layer; If the number of channels of the input and the output is different, then through a convolutional layer, the input feature map is dimensionally adjusted to make it the same as The number of channels is the same; Let be the convolution kernel, and the adjusted skip connection is , then the output formula of the residual block becomes ; 2) Integrate the residual block into the CNN structure: In the convolutional layer part of the CNN neural network, insert the residual block between the convolutional layers to replace part of the convolutional layer combination.

[0019] The steps included in the inserted attention module are as follows: 1) Construct the channel attention SE module: Squeeze operation: Perform global average pooling on the input feature map to obtain the global features of each channel, where C is the number of channels, and H and W are the height and width of the feature map respectively; the global average pooling formula is: where is the global feature of the c-th channel, is the feature value of the c-th channel at position ; Excitation operation: Perform non-linear transformation on the global feature z through two fully connected layers to learn the correlation between channels; the first fully connected layer reduces the number of channels from C to , r is the reduction ratio, and the activation function uses ReLU; the second fully connected layer restores the number of channels to C, and the activation function uses Sigmoid; Let and be the weight matrices of the two fully connected layers respectively, then , where is the ReLU activation function, is the Sigmoid activation function, is the weight of each channel; Feature recalibration: Multiply the weight s with the input feature map channel by channel to obtain the feature map after attention adjustment, where represents element-wise multiplication; 2) Integrate into the CNN structure: Insert the channel attention SE module after the residual block or convolutional layer, and screen and strengthen the features through the attention mechanism.

[0020] Training the constructed model includes the following steps: 1) Divide the dataset The preprocessed image data is divided into a training set, a validation set, and a test set, generally in the ratio of (7:1:2) or (8:1:1).

[0021] 2) Define the loss function and the optimizer Loss function: Use the cross-entropy loss function L. For a dataset with N samples, the formula is as follows: where, is the true label, is the probability predicted by the model; Optimizer: Select the Adam optimizer. The formula for updating its parameters is: where, and are the first-order and second-order moment estimates, and are the decay rates (usually set to 0.9 and 0.999 respectively), is the learning rate, is a small constant (to prevent division-by-zero errors), is the model parameter at the t-th iteration, is the gradient at the t-th iteration.

[0022] Model training process During the training process, the image data in the training set is input into the model. The output result is calculated through forward propagation, and then the loss value is calculated according to the loss function. Then, the optimizer is used for backpropagation to update the parameters of the model. Repeat this process until the performance of the model on the validation set reaches stability or meets the preset stopping conditions.

[0023] Model prediction: Perform the same preprocessing operations on the cotton image to be judged as the training data, including scaling and normalization. Input the preprocessed image into the trained improved convolutional neural network model. After calculations in the convolutional layer, residual blocks, SE module, pooling layer, and fully connected layer, the output of the model is obtained. Assume the output of the model is , and after being processed by the Softmax function, the probability of each class is obtained. The formula of the Softmax function is: K is the number of categories (K = 2, representing drought and non-drought respectively), is the k-th value output by the model; Judgment result: The judgment is made according to the probability values output by the Softmax function. If the probability of the drought category is greater than the set threshold (0.5), it is judged that the cotton is in a drought state; otherwise, it is judged that the cotton is in a non-drought state.

[0024] Furthermore, as Figure 5 shown, the implementation method of the drought assessment and decision-making module includes the following steps: 1) Data collection and preprocessing: Quantify the growth status, leaf color, and texture features of cotton processed by the image recognition and processing module through an improved convolutional neural network model; determine the current growth stage of the cotton according to agricultural knowledge and actual monitoring, obtain the soil humidity data of the cotton planting area in real time, divide the soil humidity data according to the corresponding standards, and convert it into corresponding numerical values; standardize the image recognition results, cotton growth stage numerical values, and soil humidity numerical values; 2) Determine the evaluation factor set and evaluation grade set: Determine the factors affecting the cotton drought situation and construct an evaluation factor set , where is the image recognition result (used to comprehensively reflect the growth status, leaf color, and texture features of cotton), is the cotton growth stage, is the soil humidity; Divide the cotton drought situation into three grades and construct an evaluation grade set , where is mild drought, is moderate drought, is severe drought; 3) Determine the factor weight vector A: Use the Analytic Hierarchy Process (AHP) to determine the weights of each evaluation factor: Construct a judgment matrix : Invite agricultural experts to compare the relative importance of each evaluation factor and construct a judgment matrix , where represents the degree of importance of factor relative to factor , and usually takes values of 1, 3, 5, 7, 9, representing equally important, slightly important, significantly important, strongly important, and extremely important respectively; Calculate the weight vector A: For the judgment matrix Perform consistency check. If the check passes, calculate its eigenvector, and obtain the factor weight vector after normalization processing. , where represents the weight of factor , and ; 4) Establish the fuzzy relation matrix R: Determine the membership function: For each evaluation factor , determine its membership function for different evaluation levels ; For the image recognition result , establish the corresponding membership function according to the relationship between the withered and yellow degree of the leaves, the curling degree and the drought level ; For the cotton growth stage , establish the membership function according to the sensitivity degree of different growth stages to drought ); For the soil humidity , establish the membership function according to the relationship between the soil humidity and the drought level ; Calculate the membership degree: Substitute the standardized data of each evaluation factor into the corresponding membership function, and calculate the membership degree of each factor for different evaluation levels: For the standardized data of the image recognition result , calculate its membership degree for mild drought as , for moderate drought as , and for severe drought as . Similarly, calculate the membership degrees of the cotton growth stage and the soil humidity for each evaluation level; Construct the fuzzy relation matrix R: Combine the calculated membership degrees into the fuzzy relation matrix , where represents the membership degree of the factor evaluation factor for the evaluation level ; 5) Conduct fuzzy comprehensive evaluation: Fuzzy composition operation Adopt the fuzzy composition operator to perform the composition operation on the factor weight vector A and the fuzzy relation matrix R to obtain the fuzzy comprehensive evaluation result vector : ; Among them, According to the principle of maximum membership degree, select the fuzzy comprehensive evaluation result vector The evaluation level corresponding to the element with the largest membership degree in the set is taken as the drought level of cotton. For example, if is the largest, it is determined that the cotton is in a moderately drought state.

[0025] 6) Provide irrigation decision-making suggestions: Based on the drought level obtained from the assessment, corresponding irrigation decision-making suggestions are provided for users: Mild drought: At this time, the drought impact on cotton is relatively small, but appropriate water supplementation is still required to maintain growth. Adopt the irrigation method of small amounts but multiple times, with each irrigation amount being 30% - 50% of the normal irrigation amount, and the irrigation interval being 3 - 5 days; Moderate drought: The growth of cotton is inhibited to a certain extent, and timely irrigation is required to relieve the drought situation. Increase the irrigation amount, with each irrigation amount being 60% - 80% of the normal irrigation amount, and the irrigation interval being 2 - 3 days; Severe drought: The growth of cotton is severely hindered. Immediately conduct a large amount of irrigation, with each irrigation amount being 80% - 100% of the normal irrigation amount, and continuously irrigate for 2 - 3 days to ensure that the soil humidity returns to an appropriate level.

[0026] Through the above steps, a drought assessment model is established using the fuzzy comprehensive evaluation method, which can comprehensively consider the image recognition results, cotton growth stage, and soil humidity information, accurately evaluate the drought degree of cotton, and provide scientific and reasonable irrigation decision-making suggestions for users.

[0027] In summary, the system can realize real-time image acquisition and data transmission, cooperate with efficient algorithm processing, complete drought recognition and assessment in a short time, use deep learning algorithms to learn and analyze a large number of image features, combine with a drought assessment model of multi-source data fusion, and can accurately judge the drought degree of cotton. Automatically generate scientific irrigation decision-making suggestions according to the drought assessment results, without the need for growers to have professional agricultural knowledge, reduce the decision-making difficulty and cost, and improve the intelligent level of agricultural production.

Claims

1. A cotton drought recognition and management system based on image recognition, characterized in that The system comprises: Image acquisition module: HD cameras and multispectral cameras carried by drones are used to collect images of cotton planting areas at different time periods and heights. HD cameras are installed at fixed locations in cotton planting areas to obtain cotton growth images near the ground and soil surface conditions. Multispectral cameras carried by drones are used to obtain reflection information of cotton plants and surrounding environments in different bands. Data transmission module: transmits the image data acquired by the image acquisition module to the image recognition and processing module through a wireless transmission network for processing; Image recognition and processing module: The improved convolutional neural network model is used to process and analyze the collected images to identify the growth status, leaf color and texture characteristics of cotton. First, the image is preprocessed, and then the trained convolutional neural network model is used to extract and classify the preprocessed image to determine whether the cotton is in a drought state. Drought assessment and decision-making module: Based on the results of the image recognition and processing module, combined with the growth stage of cotton and soil moisture information, a drought assessment model is established using the fuzzy comprehensive evaluation method to assess the drought severity of cotton. The drought is divided into three levels: mild drought, moderate drought and severe drought. Based on the assessment results, corresponding irrigation decision-making suggestions are provided to users. User interaction module: used to provide users with a visual operation interface, through which users can view the growth status of cotton, drought assessment results and irrigation decision suggestions; at the same time, users can set system parameters on the interface to manage the system.

2. The cotton drought condition identification and management system based on image recognition as claimed in claim 1, characterized in that: The image acquisition module comprises: Camera image acquisition unit: including high-definition cameras, which are arranged according to the topography, area and layout of the cotton planting area to cover the entire planting area without obvious monitoring blind spots. The installation height is between 1-2 meters, so that the cotton plants and soil surface can be clearly photographed; different image acquisition time periods are set according to the growth cycle and lighting characteristics of cotton; the images collected by the high-definition camera are filtered using median filtering or Gaussian filtering, and then the collected images are enhanced using image enhancement algorithms; Multispectral camera image acquisition unit: A multispectral camera is used in conjunction with a drone. The multispectral camera has the ability to detect multiple specific bands and is used to obtain the reflection information of cotton plants and the surrounding environment in different bands. The multispectral camera is installed on the gimbal system of the drone to keep the multispectral camera stable during flight. The drone route planning software is used to plan the flight route according to the terrain, area and layout of the cotton-growing area. The flight altitude is set between 30-50 meters. The route planning allows adjacent images to have a 60%-80% overlap for subsequent image stitching and processing.

3. The cotton drought condition identification and management system based on image recognition as claimed in claim 1, characterized in that: The method for constructing an improved convolutional neural network model in the image recognition and processing module comprises the following steps: Convolutional layer and residual block stacking: After the input layer of the convolutional neural network, convolutional layers and residual blocks are used alternately for feature extraction. The convolutional layer is used to preliminarily extract local features of the image, and the residual block is used to deepen the network depth through residual connections and learn more complex features. Insert attention module: insert the channel attention SE module after the residual block or convolution layer, and adjust the channel attention of the extracted features through the channel attention SE module to enhance the expression of the selected features; Set up pooling layer and fully connected layer: After the convolution layer and residual block, use the pooling layer for downsampling to reduce the size of the feature map. Finally, use the fully connected layer to map the features to the final classification result to determine whether the cotton is in a drought state.

4. The cotton drought condition identification and management system based on image recognition as claimed in claim 3, characterized in that: The convolutional layer and residual block stacking includes the following steps: 1) Define the residual block: Assume the input feature map is , after a series of convolution operations, the output is ; Output of residual block Calculated by the following formula: , adding the input directly to the output of the convolutional layer; If the number of channels of the input and output is different, Convolutional layer input feature map Adjust the dimensions to match The number of channels is consistent; set up for The convolution kernel and the adjusted skip connection are , then the residual block output formula becomes ; 2) Integrate the residual block into the CNN structure: Residual blocks are inserted between convolutional layers to replace some convolutional layer combinations.

5. The cotton drought condition identification and management system based on image recognition as claimed in claim 3, characterized in that: The inserting attention module comprises the following steps: 1) Construct channel attention SE module: The input feature map Perform global average pooling to obtain the global features of each channel, where C is the number of channels, H and W are the height and width of the feature map, respectively. The global average pooling formula is: ； in, is the global feature of the cth channel, is the cth channel at position The characteristic value of Global Features Through two fully connected layers, nonlinear transformation is performed to learn the correlation between channels; the first fully connected layer reduces the number of channels from C to , r is the reduction ratio, and ReLU is the activation function; the second fully connected layer restores the number of channels to C, and Sigmoid is the activation function; set up and are the weight matrices of the two fully connected layers respectively, then ,in is the ReLU activation function, is the Sigmoid activation function, is the weight of each channel; The weight s is combined with the input feature map Multiply channel by channel to get the feature map after attention adjustment ,in Represents element-wise multiplication; 2) Integrate into CNN structure: The channel attention SE module is inserted into the residual block or convolutional layer to filter and enhance the features through the attention mechanism.

6. The cotton drought condition identification and management system based on image recognition as claimed in claim 1, characterized in that: The implementation method of the drought assessment and decision-making module comprises the following steps: 1) Data collection and preprocessing: The image recognition and processing module is used to quantify the growth status, leaf color and texture characteristics of cotton after being processed by the improved convolutional neural network model; the current growth stage of cotton is determined based on agricultural knowledge and actual monitoring, and the soil moisture data of the cotton planting area is obtained in real time. The soil moisture data is divided according to the corresponding standards and converted into corresponding values; the image recognition results, cotton growth stage values ​​and soil moisture values ​​are standardized; 2) Determine the evaluation factor set and evaluation level set: Determine the factors affecting cotton drought and construct an evaluation factor set ,in It is the result of image recognition, which is used to comprehensively reflect the growth status, leaf color and texture characteristics of cotton. The cotton growth stage. is soil moisture; The cotton drought condition is divided into three levels and an evaluation level set is constructed. ,in Mild drought, Moderate drought. It is severe drought; 3) Determine the factor weight vector A: The analytic hierarchy process (AHP) is used to determine the weight of each evaluation factor: Constructing a judgment matrix : Invite agricultural experts to compare the relative importance of each evaluation factor and construct a judgment matrix ,in Indication factors Relative to factors The importance level is 1, 3, 5, 7, and 9, which represent equally important, slightly important, obviously important, strongly important, and extremely important, respectively; Calculate the weight vector A: for the judgment matrix Perform a consistency test. If the test is passed, calculate its eigenvector and obtain the factor weight vector after normalization. ,in Indication factors The weight of ; 4) Establish the fuzzy relationship matrix R For each evaluation factor , respectively determine their evaluation levels Membership function; for image recognition results According to the relationship between the degree of leaf yellowing and curling and the drought level, the corresponding membership function is established. ; For cotton growth stage , according to the sensitivity of different growth stages to drought, the membership function is established ); for soil moisture , based on the relationship between soil moisture and drought conditions, the membership function is established ; Substitute the standardized evaluation factor data into the corresponding membership function and calculate the membership degree of each factor to different evaluation levels: Standardized data , calculate its effect on mild drought The membership degree is , for moderate drought The membership degree is , for severe drought The membership degree is , similarly, calculate the cotton growth stage and soil moisture The degree of membership to each evaluation level; Combine the calculated membership degrees into a fuzzy relationship matrix ,in Expression Factor Evaluation Factor Evaluation level The degree of membership; 5) Conduct fuzzy comprehensive evaluation: The fuzzy synthesis operator is used to synthesize the factor weight vector A and the fuzzy relationship matrix R to obtain the fuzzy comprehensive evaluation result vector ; According to the maximum membership principle, select the fuzzy comprehensive evaluation result vector The evaluation level corresponding to the element with the largest membership degree is taken as the drought level of cotton; 6) Provide irrigation decision: Mild drought: cotton is less affected by drought, but still needs to be properly watered to maintain growth. A small amount of irrigation is used multiple times, with each irrigation amount being 30% -50% of the normal amount, and the irrigation interval is 3-5 days. Moderate drought: cotton growth is inhibited to a certain extent, and timely irrigation is needed to alleviate the drought. The irrigation amount should be increased, with each irrigation amount being 60% - 80% of the normal irrigation amount, and the irrigation interval being 2-3 days; Severe drought: Cotton growth is severely hindered. Immediately carry out large-scale irrigation, with each irrigation amount being 80%-100% of the normal irrigation amount. Irrigate continuously for 2-3 days to ensure that soil moisture returns to an appropriate level.

Images