Multi-stage linkage agricultural flood warning information automatic broadcasting system

Abstract

The application relates to the technical field of agricultural early warning, in particular to a multi-stage linkage agricultural flood early warning information automatic broadcasting system, and the following scheme is provided: a sensing terminal collects farmland images, combines environmental data, analyzes waterlogging features by using a graph neural network, and thus precise flood disaster risk assessment is realized. After disaster identification, the system dynamically adjusts the broadcasting range and the communication protocol according to the risk level, so as to ensure that the early warning information is rapidly transmitted to the receiving terminal. In addition, an iterative operation mechanism is adopted, a cluster head terminal is selected according to the consistency of the waterlogging features, information redundancy transmission is avoided, and network resource use is optimized. The application improves the real-time performance, accuracy and resource utilization efficiency of the flood disaster early warning system, and has obvious innovation and application value.

Description

Technical Field

[0001] This invention relates to the field of agricultural early warning technology, and in particular to a multi-level linkage automatic broadcasting system for agricultural flood early warning information. Background Technology

[0002] Traditional flood warning systems typically rely on single sensor data or centralized processing methods, which often exhibit low accuracy and real-time performance when faced with complex farmland environments. For example, existing systems are generally unable to effectively handle water accumulation in different areas of farmland, and due to limitations in network bandwidth and computing power, information transmission is prone to delays or misjudgments, leading to reduced reliability of the warning system. Furthermore, most existing flood warning methods depend on centralized servers or host computers, resulting in high system costs and difficulty in adapting to the needs of large-scale farmland monitoring. Therefore, how to improve the accuracy, real-time performance, and adaptability of flood disaster warning systems through more efficient and flexible technological means has become a pressing technical challenge in the agricultural sector.

[0003] For example, Chinese patent application CN116434479A discloses a method and system for predicting and warning of regional flood disaster levels. The method includes: establishing an LSTM neural network prediction model to predict the cumulative future rainfall in the region; acquiring weather data for the next 24 hours and real-time water level data for the regional watershed; predicting the flood disaster level; and issuing an early warning signal when the predicted flood disaster level exceeds a preset value. The system includes a data acquisition module, a prediction model construction module, a rainfall prediction module, a flood disaster level prediction module, and an early warning module. This invention's method is not limited to specific flood disaster locations but is universally applicable to all types of flood disasters and can classify urban flooding levels.

[0004] The above-mentioned existing technologies all suffer from the problems mentioned in the background: due to the limitations of network bandwidth and computing power, delays or misjudgments are prone to occur during information transmission. To solve the above problems, this application designs a multi-level linkage automatic broadcasting system for agricultural flood early warning information. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a multi-level, interconnected automatic broadcasting system for agricultural flood warnings. This system collects farmland images through sensor terminals and combines them with environmental data, using graph neural networks to analyze water accumulation characteristics, thereby achieving accurate flood disaster risk assessment. After disaster identification, the system dynamically adjusts the broadcast range and communication protocol according to the risk level, ensuring rapid transmission of warning information to receiving terminals. Furthermore, an iterative operation mechanism is employed, selecting cluster head terminals based on the consistency of water accumulation characteristics to avoid redundant information transmission and optimize network resource utilization. This application improves the real-time performance, accuracy, and resource utilization efficiency of flood disaster warning systems, demonstrating significant innovation and application value.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A multi-level, interconnected automatic agricultural flood warning information broadcasting system is applied to distributed farmland. Each independent farmland within this system is equipped with interconnected sensor terminals. These sensor terminals integrate image sensors and processing modules. The system includes a terminal communication module, a flood identification module, and a warning notification module.

[0008] The terminal communication module is used to respond to the water level alarm information sent by the water level gauge, wake up the sensor terminal, build a sensor terminal communication network, and select the cluster head terminal. The water level alarm information is the information triggered by the water level gauge after the water level is higher than a set threshold.

[0009] The flood identification module is used to broadcast farmland images collected by the cluster head terminal to each sensor terminal, and each sensor terminal processes the farmland images to generate flood disaster risk, and returns the flood disaster risk to the cluster head terminal, selecting the judgment result with the most votes as the flood disaster risk identification result;

[0010] The early warning notification module is used to broadcast the flood disaster risk identification results to other terminals in the sensor terminal communication network.

[0011] The terminal communication module includes:

[0012] The terminal wake-up unit is used to activate the corresponding sensor terminal and start its working mode after receiving the water level alarm information from the water level gauge. The terminal wake-up unit interacts with the water level gauge through a wireless communication protocol to obtain water level data and determine whether the wake-up conditions are met.

[0013] The network construction unit is used to establish communication links between the sensing terminals and form a sensing terminal communication network after the sensing terminals are woken up.

[0014] The network construction unit includes:

[0015] The awakened sensor terminal initiates a self-organizing network mechanism, establishes a communication link with neighboring terminals through a communication protocol, and forms a sensor terminal communication network.

[0016] In the sensor terminal communication network, cluster head terminals are selected based on the geographical location, terrain changes, and water flow direction of the sensor terminals.

[0017] The flood identification module includes:

[0018] An image preprocessing unit is used to preprocess the farmland image, wherein the preprocessing includes illumination compensation and image dehazing;

[0019] The feature extraction unit is used to input the preprocessed farmland image into a preset feature extraction model and output water accumulation features through the feature extraction model.

[0020] The feature fusion unit is used to deeply fuse the water accumulation features with the environmental data collected by the sensing terminal itself to generate global graph features;

[0021] The disaster identification unit is used to map the fused data to the corresponding flood disaster risk based on the global map features and a preset flood level classification standard.

[0022] The image preprocessing unit includes:

[0023] Calculate the global brightness distribution of the farmland image, and calculate the illumination compensation parameters based on the global brightness distribution using histogram equalization;

[0024] Farmland images are dehazed using a generative adversarial network (GAN), which is trained on both foggy and fog-free farmland images during the training phase.

[0025] The feature extraction unit includes:

[0026] The preprocessed farmland image is converted to the spectral domain, and local features of the spectral curve are extracted in the spectral domain using a one-dimensional convolution operation.

[0027] Global features of the spectral curve are extracted using global pooling, and spectral indices are calculated based on the local and global features.

[0028] Spectral indicators are fused using a feature pyramid to generate a feature map. The feature map is then processed using an attention mechanism to extract feature vectors.

[0029] The feature vector is input into a fully connected layer for classification and regression operations, and the water accumulation feature is output.

[0030] The feature fusion unit includes:

[0031] The water accumulation features and the environmental data are used as nodes in the graph data, where the water accumulation features and the environmental data are respectively used as different types of nodes in the graph, and the relationships between the data are used as edges in the graph.

[0032] The graph data is subjected to graph convolution operations through a graph neural network. The graph convolution operation uses a combination of multiple convolutional layers and nonlinear activation functions to achieve the propagation and updating of features of different nodes, capture the spatiotemporal dynamic information between nodes, and form graph node features.

[0033] When multiple flood disaster risk assessments result in the same number of votes, the flood identification module further includes:

[0034] Multiple flood disaster risks are sorted according to preset priority rules to generate a candidate sequence, wherein the priority rules are determined based on the current rate of expansion of waterlogged areas in farmland;

[0035] Acquire environmental data from adjacent sensor terminals and calculate the applicability scores of multiple flood disaster risks in adjacent areas;

[0036] The candidate sequences are reordered based on the applicability score, and the flood disaster risk ranked first is taken as the flood disaster risk identification result.

[0037] The early warning notification module includes:

[0038] A broadcast generation unit is used to adjust the broadcast range and communication protocol according to the flood disaster risk identification results in order to transmit early warning information to the receiving terminal;

[0039] The iterative reporting unit is used to compare the farmland images collected by the receiving terminal with the early warning information and broadcast the early warning information level by level.

[0040] The iterative notification unit includes:

[0041] The receiving terminal collects current farmland images. If the water accumulation characteristics of the current farmland image are consistent with those of the received farmland image, the receiving terminal will continue to transmit the warning information as the cluster head terminal until the water accumulation characteristics of the current farmland image and the received farmland image are inconsistent, thus completing the broadcast of the warning information.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. This invention combines image processing and environmental data, employing graph neural networks to extract water accumulation features, enabling accurate identification of waterlogged areas and flood risks in farmland. This improves the early warning system's ability to identify flood disasters under different environmental conditions, reducing the risk of false alarms and missed alarms.

[0044] 2. This invention employs an iterative mechanism based on the consistency of water accumulation characteristics to transmit early warning information level by level among terminals, avoiding unnecessary duplicate transmissions. This makes information transmission more flexible and efficient, avoids network overload, and optimizes bandwidth and battery usage. Attached Figure Description

[0045] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0046] Figure 1 This is a module diagram of the multi-level linkage automatic broadcasting system for agricultural flood early warning information according to Embodiment 1 of the present invention;

[0047] Figure 2 This is a flowchart illustrating the multi-level linkage automatic broadcasting method for agricultural flood early warning information according to Embodiment 2 of the present invention.

[0048] Figure 3 This is a schematic diagram of farmland image broadcasting in Embodiment 2 of the present invention;

[0049] Figure 4 This is a schematic diagram of the return of flood disaster identification results in Embodiment 2 of the present invention. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Example 1:

[0052] Please see Figure 1 This invention provides an embodiment of a multi-level linkage agricultural flood early warning information automatic broadcasting system, applied to distributed farmland. Each independent farmland in the distributed farmland is equipped with interconnected sensor terminals. Each sensor terminal integrates an image sensor and a processing module. The system includes a terminal communication module, a flood identification module, and an early warning notification module.

[0053] The terminal communication module is used to respond to the water level alarm information sent by the water level gauge, wake up the sensor terminal, build a sensor terminal communication network, and select the cluster head terminal. The water level alarm information is the information triggered by the water level gauge after the water level is higher than a set threshold.

[0054] The flood identification module is used to broadcast farmland images collected by the cluster head terminal to each sensor terminal, and each sensor terminal processes the farmland images to generate flood disaster risk, and returns the flood disaster risk to the cluster head terminal, selecting the judgment result with the most votes as the flood disaster risk identification result;

[0055] The early warning notification module is used to broadcast the flood disaster risk identification results to other terminals in the sensor terminal communication network.

[0056] The terminal communication module includes:

[0057] The terminal wake-up unit is used to activate the corresponding sensor terminal and start its working mode after receiving the water level alarm information from the water level gauge. The terminal wake-up unit interacts with the water level gauge through a wireless communication protocol to obtain water level data and determine whether the wake-up conditions are met.

[0058] The network construction unit is used to establish communication links between the sensing terminals and form a sensing terminal communication network after the sensing terminals are woken up.

[0059] The network construction unit includes:

[0060] The awakened sensor terminal initiates a self-organizing network mechanism, establishes a communication link with neighboring terminals through a communication protocol, and forms a sensor terminal communication network.

[0061] In the sensor terminal communication network, cluster head terminals are selected based on the geographical location, terrain changes, and water flow direction of the sensor terminals.

[0062] The flood identification module includes:

[0063] An image preprocessing unit is used to preprocess the farmland image, wherein the preprocessing includes illumination compensation and image dehazing;

[0064] The feature extraction unit is used to input the preprocessed farmland image into a preset feature extraction model and output water accumulation features through the feature extraction model.

[0065] The feature fusion unit is used to deeply fuse the water accumulation features with the environmental data collected by the sensing terminal itself to generate global graph features;

[0066] The disaster identification unit is used to map the fused data to the corresponding flood disaster risk based on the global map features and a preset flood level classification standard.

[0067] The image preprocessing unit includes:

[0068] Calculate the global brightness distribution of the farmland image, and calculate the illumination compensation parameters based on the global brightness distribution using histogram equalization;

[0069] Farmland images are dehazed using a generative adversarial network (GAN), which is trained on both foggy and fog-free farmland images during the training phase.

[0070] The feature extraction unit includes:

[0071] The preprocessed farmland image is converted to the spectral domain, and local features of the spectral curve are extracted in the spectral domain using a one-dimensional convolution operation.

[0072] Global features of the spectral curve are extracted using global pooling, and spectral indices are calculated based on the local and global features.

[0073] Spectral indicators are fused using a feature pyramid to generate a feature map. The feature map is then processed using an attention mechanism to extract feature vectors.

[0074] The feature vector is input into a fully connected layer for classification and regression operations, and the water accumulation feature is output.

[0075] The feature fusion unit includes:

[0076] The water accumulation features and the environmental data are used as nodes in the graph data, where the water accumulation features and the environmental data are respectively used as different types of nodes in the graph, and the relationships between the data are used as edges in the graph.

[0077] The graph data is subjected to graph convolution operations through a graph neural network. The graph convolution operation uses a combination of multiple convolutional layers and nonlinear activation functions to achieve the propagation and updating of features of different nodes, capture the spatiotemporal dynamic information between nodes, and form graph node features.

[0078] When multiple flood disaster risk assessments result in the same number of votes, the flood identification module further includes:

[0079] Multiple flood disaster risks are sorted according to preset priority rules to generate a candidate sequence, wherein the priority rules are determined based on the current rate of expansion of waterlogged areas in farmland;

[0080] Acquire environmental data from adjacent sensor terminals and calculate the applicability scores of multiple flood disaster risks in adjacent areas;

[0081] The candidate sequences are reordered based on the applicability score, and the flood disaster risk ranked first is taken as the flood disaster risk identification result.

[0082] The early warning notification module includes:

[0083] A broadcast generation unit is used to adjust the broadcast range and communication protocol according to the flood disaster risk identification results in order to transmit early warning information to the receiving terminal;

[0084] The iterative reporting unit is used to compare the farmland images collected by the receiving terminal with the early warning information and broadcast the early warning information level by level.

[0085] The iterative notification unit includes:

[0086] The receiving terminal collects current farmland images. If the water accumulation characteristics of the current farmland image are consistent with those of the received farmland image, the receiving terminal will continue to transmit the warning information as the cluster head terminal until the water accumulation characteristics of the current farmland image and the received farmland image are inconsistent, thus completing the broadcast of the warning information.

[0087] Example 2:

[0088] Please see Figure 2 This invention provides an embodiment of a multi-level linkage method for automatically broadcasting agricultural flood early warning information, applied to distributed farmland. Each independent farmland in the distributed farmland is equipped with interconnected sensor terminals, each integrating an image sensor and a processing module. The specific steps of the method are as follows:

[0089] S1: In response to the water level alarm information sent by the water level gauge, wake up the sensing terminal;

[0090] In this embodiment, when the water level gauge detects that the water level exceeds a set threshold, it sends a water level alarm message to the corresponding sensor terminal via a wireless communication protocol. The sensor terminal is designed with a low-power standby mode, instantly waking up upon receiving the water level alarm message to begin data acquisition and processing. By employing a low-power standby mechanism, the operating time of the sensor terminal can be effectively extended, ensuring a rapid response in the event of flooding.

[0091] S2: Construct a sensor terminal communication network and select cluster head terminals;

[0092] In this embodiment, once the sensor terminal is activated, it establishes communication links with neighboring terminals via a self-organizing network mechanism, automatically constructing a multi-layered sensor terminal communication network. Through the self-organizing network protocol, the sensor terminal can dynamically select a cluster head terminal based on factors such as the node's geographical location and signal strength. The cluster head terminal is responsible not only for data aggregation and preliminary processing but also for coordinating and transmitting data to other terminals. By selecting a cluster head terminal, the efficiency of data transmission and network stability are effectively improved, ensuring that information can be transmitted quickly and accurately across large areas of farmland.

[0093] As will be readily understood by those skilled in the art, a neighboring terminal can be understood as a terminal that can communicate directly.

[0094] S3: Broadcast the farmland images collected by the cluster head terminal to each sensor terminal, and have each sensor terminal process the farmland images to generate flood disaster risk;

[0095] In this embodiment, the cluster terminal broadcasts the farmland images it collects to all interconnected sensor terminals via wireless communication. Each terminal processes the image information received along with its local environmental data (such as soil moisture and temperature), and uses image recognition technology to generate a flood risk assessment. The various sensor terminals can work collaboratively, processing images locally and performing risk assessments based on inputs from different sensors. This reduces computational burden and improves image processing accuracy, especially in complex farmland environments, enabling rapid detection of potential flood risks.

[0096] S4: Return the flood disaster risk to the cluster head terminal and select the judgment result with the most votes as the flood disaster risk identification result;

[0097] In this embodiment, each sensor terminal returns its processed flood risk information to the cluster head terminal. The cluster head terminal selects the most frequent judgment result as the final flood risk identification result through a voting mechanism. By integrating the judgment results of multiple nodes, the accuracy of flood risk identification is enhanced, and false alarms or missed alarms caused by errors in a single node are avoided. The voting mechanism effectively improves the robustness of the early warning system, ensuring more reliable disaster early warnings even in complex farmland environments.

[0098] S5: Broadcast the flood disaster risk identification results to other terminals on the network;

[0099] In this embodiment, the cluster head terminal broadcasts the final flood disaster risk identification result to other terminals, ensuring that sensor terminals in each area can obtain the latest flood disaster early warning information in a timely manner. Through a multi-level linkage broadcast mechanism, the flood disaster risk identification result is transmitted between terminals, ensuring that information is quickly and accurately delivered to every terminal in the entire network, achieving multi-level linkage intelligent early warning. This reduces the redundancy of information transmission in traditional broadcasting methods, ensures the effective use of network resources, and reduces unnecessary information broadcasting in low-risk areas.

[0100] Traditional flood disaster early warning methods typically rely on unified monitoring data transmission and centralized processing. However, due to the complexity and geographical variations of farmland environments, this approach struggles to achieve both accuracy and efficiency. Especially in farmland with significant topographic relief, low-lying areas are most prone to waterlogging, yet traditional systems do not specifically consider risk assessment for these particular areas. This application addresses this by placing cluster-head terminals in the lowest-lying, most flood-prone locations, combined with a hierarchical disaster assessment mechanism based on these terminals, to accurately identify and assess potential disaster risks.

[0101] Specifically, this application prioritizes low-lying areas as monitoring zones by dynamically adjusting the disaster assessment process. Since these areas are most vulnerable to flooding, cluster-head terminals play a crucial role. By collecting specific data from these areas, the cluster-head terminals can initially assess the disaster risk of the region and then progressively transmit the assessment results to other terminals. This hierarchical assessment and information transmission method avoids the information processing lag problems of centralized systems, enabling disaster information to be transmitted to a wider area in the shortest possible time. Furthermore, because the cluster-head terminals are located in areas most likely to experience disasters, the real-time data they collect is more representative and provides more accurate guidance for generating early warning information.

[0102] Furthermore, the mechanism proposed in this application, which uses cluster-head terminals for tiered disaster assessment and multi-level linkage, solves the problems of insufficient information coverage and slow response speed in traditional systems when dealing with farmland flooding. By flexibly adjusting the roles and responsibilities of each sensor terminal in the network, it ensures that the areas most prone to flooding receive priority assessment and attention. The cluster-head terminal design is not limited to data collection and transmission; it also makes intelligent decisions based on geographic information and can adjust assessment strategies according to the needs of different regions, thereby ensuring that sensors at different levels perform the most appropriate response actions based on local environmental differences. This improves the accuracy of early warnings and enhances the overall effectiveness of the disaster early warning system through tiered linkage.

[0103] The specific steps of S2 are as follows:

[0104] S2.1: The awakened sensor terminal initiates the self-organizing network mechanism, establishes a communication link with neighboring terminals through the communication protocol, and forms a sensor terminal communication network;

[0105] In this embodiment, low-power wireless communication protocols (such as Zigbee, LoRa, NB-IoT, etc.) are used to achieve the connection between the sensing terminals. These protocols can maintain low power consumption during long-term operation and have strong anti-interference capabilities, making them suitable for use in agricultural environments.

[0106] Specifically, when a sensor terminal is activated, it first scans the surrounding wireless signal environment and selects the nearest terminal with the strongest and most stable signal to establish a communication connection. This process dynamically adjusts the network topology based on the terminal density in the actual environment, avoiding network congestion or signal interference. Through this self-organizing method, the network can flexibly adapt to environmental changes, ensuring system stability and efficiency.

[0107] S2.2: In the sensor terminal communication network, the cluster head terminal is selected according to the geographical location, terrain changes and water flow direction of the sensor terminal;

[0108] Specifically, in agricultural environments, the terrain varies greatly, and the direction of water flow in certain areas can affect the spread of flood disasters. By collecting terrain data in real time, the system can analyze the direction of water flow in different areas and select sensor terminals located in water confluence areas as cluster heads to ensure that disaster early warning information in these areas can be processed and transmitted with priority.

[0109] Furthermore, the selection of cluster head terminals is performed using an optimization algorithm. This algorithm considers the computing power, communication capabilities, and location priority of each node to calculate the most suitable cluster head terminal. By assigning weights, the algorithm gives higher priority to sensor terminals in low-lying areas and those near the direction of water flow, while also considering the remaining battery power of the sensor terminals and network load, selecting the most advantageous sensor terminal as the cluster head. The accuracy of cluster head terminal selection ensures the timeliness and accuracy of flood disaster early warning. Especially in agricultural areas with complex terrain, the system can prioritize vulnerable areas, reducing the risk of misjudgment and missed judgment. In addition, the cluster head selection strategy based on terrain and water flow direction also improves the system's adaptability, enabling it to intelligently adjust according to environmental changes in different areas, ensuring the stability and efficient operation of the entire network.

[0110] The specific steps for S3 are as follows:

[0111] S3.1: Preprocess the farmland image, wherein the preprocessing includes illumination compensation and image dehazing;

[0112] Specifically, during image acquisition, environmental factors (such as weather and shooting angle) may cause uneven brightness distribution in the image. Especially in rainy weather with poor lighting conditions, image details may be obscured, leading to a decrease in the accuracy of subsequent processing. Therefore, illumination compensation technology can dynamically adjust the global brightness of the image, and histogram equalization algorithms can be used to enhance the details in the image, ensuring that the structural information of the image is clearly presented.

[0113] Furthermore, even after illumination compensation, images may still be affected by haze, smoke, or other obstacles, leading to reduced visibility and clarity. This paper introduces an image dehazing technique based on a Generative Adversarial Network (GAN) model. By training with a large number of paired clear and hazy images, the dehazing model can effectively identify and remove haze interference from images, restoring the original details. This technique can quickly and effectively improve image quality, ensuring the accuracy of subsequent feature extraction. Especially in agricultural environments, dehazing can reduce the impact of meteorological factors and improve the identification of waterlogged areas.

[0114] S3.2: Input the preprocessed farmland image into the preset feature extraction model, and output the water accumulation features through the feature extraction model;

[0115] In this embodiment, the feature extraction model is trained on a deep convolutional neural network architecture and is specifically designed to extract flood-related features from farmland images. By learning from a large number of farmland image samples, the model can automatically identify important features in the images, such as waterlogged areas, ground reflections, color changes, and water textures.

[0116] Specifically, the core of the feature extraction model is to continuously extract local features from the image through convolutional layers, while simultaneously compressing the dimensionality of the feature map through pooling layers to extract global features, ultimately forming descriptive features for waterlogged areas. These features include not only the shape, size, and distribution of the water, but also information such as the water depth and flow state. In this way, the model can accurately detect waterlogged areas in complex farmland images, providing a data foundation for subsequent flood risk assessment.

[0117] S3.3: Deeply fuse the water accumulation features with the environmental data collected by the sensing terminal itself to generate a global graph feature;

[0118] In this embodiment, multi-dimensional data fusion overcomes the limitations that may arise from a single data source, thereby enabling a more comprehensive and accurate assessment of flood disaster risks.

[0119] Specifically, the sensing terminal simultaneously collects real-time data from the environment, which typically reflects physical quantities related to waterlogging, such as climatic conditions and soil moisture in farmland. By combining this data with image features and using graph neural networks for fusion, a global graph feature that incorporates environmental factors and waterlogging conditions is generated. These global features not only reflect the physical properties of the waterlogged area but also take into account the impact of environmental changes in different areas on waterlogging, thereby improving the accuracy of disaster risk assessment.

[0120] S3.4: Based on the global graph features, according to the preset flood level classification standard, the fused data is mapped to the corresponding flood disaster risk;

[0121] This embodiment defines a preset flood risk classification standard. Based on historical data and experience, this standard classifies flood risk into multiple levels (such as low risk, medium risk, and high risk) according to different factors such as water depth, area, and duration. This standard is trained using machine learning methods to map different features to the corresponding risk levels.

[0122] Specifically, based on the fused global graph features, classification algorithms such as Support Vector Machines (SVM), decision trees, or neural networks are used to classify and predict risks. This process learns the correlation between various features and actual flood disasters through a training dataset, enabling the system to make accurate risk assessments based on new environmental data and image information. Finally, the system outputs a flood disaster risk warning and takes different response measures according to the warning level.

[0123] The specific steps of S3.1 are as follows:

[0124] S3.1.1: Calculate the global brightness distribution of the farmland image, and calculate the illumination compensation parameters based on the global brightness distribution through histogram equalization;

[0125] Specifically, due to the unevenness of lighting conditions in the natural environment, especially on rainy days, farmland images may have different areas with excessively low or high brightness, making it difficult to observe and analyze certain details.

[0126] In this embodiment, firstly, the brightness histogram of the farmland image is calculated to obtain the frequency distribution of pixel brightness in the image. Next, based on the brightness histogram, the illumination compensation parameters of the image are calculated. By adjusting the grayscale distribution of the image, the brightness of the image is made more uniform, reducing the impact of locally overly bright or dark areas. The histogram equalization process transforms the pixel values ​​of the image so that the pixel value distribution covers the entire brightness range, thereby enhancing the image contrast and making details more clearly visible.

[0127] S3.1.2: Dehazing of farmland images is performed using a generative adversarial network, wherein the generative adversarial network is trained on farmland images with and without fog during the training phase;

[0128] Specifically, due to rainy weather, farmland images may be affected by haze or fog due to weather conditions or shooting environment, resulting in loss of image clarity and affecting subsequent extraction of water accumulation features and flood risk assessment.

[0129] In this embodiment, a generative adversarial network (GAN) is a deep learning model consisting of two parts: a generator and a discriminator. The generator aims to generate a clear image from a foggy image, while the discriminator judges the difference between the generated image and the real fog-free image. Through continuous training and optimization, the generator learns how to effectively remove fog and restore image details.

[0130] Furthermore, during the training phase, a large number of paired images of foggy and fog-free farmland were used as training data. The generative adversarial network (GAN) optimized the parameters of its generator to make the generated fog-free images as close as possible to the real, clear images. During the testing phase, the trained generator was able to automatically convert foggy farmland images into clear, dehazed images. This dehazing process eliminates image blurring caused by weather and other factors, making features such as puddles and farmland topography more clearly visible.

[0131] The specific steps of S3.2 are as follows:

[0132] S3.2.1: Convert the preprocessed farmland image to the spectral domain, and extract local features of the spectral curve using a one-dimensional convolution operation in the spectral domain;

[0133] Specifically, the main purpose of the transformation is to extract information-rich spectral features from the original image, not just visible light information. The spectral domain provides more detailed information about soil, vegetation, and water bodies, helping to extract features closely related to flooding from images. In farmland environments, spectral information is more effective than simple color information because different soils, vegetation, and water bodies exhibit significant differences in their spectral representation.

[0134] In this embodiment, the purpose of the convolution operation is to extract relevant information between different wavelength bands. By performing sliding window convolution on the spectral curve, the system can capture local features within different wavelength regions, such as the reflectance characteristics of water bodies and the spectral absorption characteristics of vegetation. These local features can help the system identify the waterlogging conditions in different areas of farmland. Features can be extracted layer by layer for each wavelength band without increasing computational load, thus laying the foundation for subsequent feature fusion and risk assessment.

[0135] S3.2.2: Extract global features of the spectral curve through global pooling operation, and calculate spectral indices based on the local and global features;

[0136] Specifically, in image processing, pooling operations are typically used to reduce the size of feature maps, thereby reducing computational cost and preventing overfitting. Traditional pooling operations usually use max pooling or average pooling to extract local features, while global pooling, on the other hand, processes the entire feature map to extract global, representative information, thus enabling the identification of waterlogged features in farmland over a wide area.

[0137] In this embodiment, the spectral domain enters a global pooling layer for convergence and compression, yielding global features for each band. The global pooling operation first calculates the features of each spectral band (e.g., visible light, near-infrared, etc.). The feature maps of these bands contain feature information from all pixels within that band. Global pooling calculates the maximum or average value for each band. This operation extracts the most representative feature values ​​from each band, reflecting the main features and distribution of the entire band. By extracting the maximum value for each band, the most salient features in the image can be preserved, especially useful for identifying prominent water accumulation areas or unique reflective properties in the image. By calculating the average value for each band, the overall spectral information of the band can be preserved, avoiding overemphasis on local features and resulting in more balanced and comprehensive global features.

[0138] Furthermore, the global features extracted by global pooling need to be combined with the previously extracted local features to comprehensively assess the water accumulation in the image. Local features represent the subtle changes in the image, while global features reflect the overall structure of the image. To effectively combine this information, spectral indices are calculated to quantify relevant features such as water accumulation areas, soil reflectance, and vegetation cover in the image.

[0139] Furthermore, by weighted averaging of local features extracted from local convolutional operations with global features output from the global pooling layer, local details are combined with the overall global structure. Based on the fused local and global features, a series of spectral indices, such as the vegetation index (NDVI) and the water body index (NDWI), are calculated to measure water features in the image. Specifically, NDVI is used to assess the health of vegetation, while NDWI is primarily used to detect water bodies in the image. By combining these spectral indices, the system can accurately identify the extent and depth of water accumulation in farmland, thereby assessing the risk of flooding.

[0140] S3.2.3: The spectral indicators are fused through the feature pyramid to generate a feature map, and the feature map is processed according to the attention mechanism to extract the feature vector;

[0141] Specifically, Feature Pyramid Network (FPN) is a multi-scale feature extraction method that processes image features at different scales, thereby enhancing the ability to recognize detailed features. Through feature pyramids, image features at different resolutions can be fused and optimized at multiple levels, improving adaptability to complex farmland environments.

[0142] In this embodiment, the feature pyramid technique processes and fuses spectral indicators at multiple scales to form multi-level feature maps. These feature maps encompass a wide range of feature information, from details to the overall picture, providing rich contextual information for subsequent feature extraction. Through this multi-level feature fusion, the system can better identify water accumulation features at different scales, improving detection accuracy in complex scenes. An attention mechanism is then applied to the feature map processing, aiming to focus on the most important feature regions. Through the attention mechanism, the network can automatically learn and select the most relevant features, ignoring noise or irrelevant information. Specifically, the attention mechanism assigns a weight to each feature, allowing the model to focus more on features of water accumulation areas and other disaster-related areas. This approach not only improves the efficiency of feature extraction but also enhances the system's automation capabilities, enabling adaptive adjustments based on key regions in the image.

[0143] S3.2.4: Input the feature vector into the fully connected layer for classification and regression operations, and output the water accumulation features;

[0144] Specifically, the feature vectors processed through the attention mechanism are input into a fully connected layer for classification and regression operations. The fully connected layer integrates the input feature information to perform the final classification task, namely, determining the risk of flooding in the image. Through regression operations, the system can calculate the size, depth, and distribution of flooded areas. The output of the fully connected layer is the final flood feature, which will guide subsequent disaster early warning and response.

[0145] In this embodiment, the system utilizes fully connected layers in a deep neural network (DNN) to transform complex spectral features into simple, interpretable results. By combining classification and regression, the system can not only identify waterlogged areas but also quantitatively assess the risk level of those areas, providing data support for early warning of agricultural flooding disasters.

[0146] In this embodiment, step S3.3 involves processing water accumulation features and environmental data using a graph neural network (GNN) to establish a dynamic spatiotemporal graph data model and further capture the relationships between nodes. This process not only effectively integrates multi-source data but also enables efficient feature propagation and updating based on graph convolution operations, thereby generating global graph features and providing a more accurate flood disaster risk assessment. The specific steps of S3.3 are as follows:

[0147] S3.3.1: The water accumulation features and the environmental data are used as nodes in the graph data, wherein the water accumulation features and the environmental data are used as different types of nodes in the graph, and the relationships between the data are used as edges in the graph;

[0148] Specifically, the first step is to transform the water accumulation features extracted from farmland images and the environmental data (such as soil moisture, temperature, and precipitation) collected by sensor terminals into a graph data structure. In a graph neural network, each data item can be considered a node in the graph. Here, water accumulation features and environmental data are treated as different types of nodes in the graph, meaning their roles and information flows will differ. Specifically, water accumulation feature nodes represent the water accumulation area features identified in the image, while environmental data nodes contain environmental variables related to flood risk, such as temperature, humidity, and soil condition.

[0149] In this embodiment, to establish relationships between nodes, these different types of nodes are connected by graph edges. The edges are established based on the actual correlation between these nodes; for example, a water accumulation feature node may be associated with a specific environmental data node (such as soil moisture or precipitation), reflecting how environmental conditions affect the distribution of water accumulation. In this way, the system can capture the complex dependencies between water accumulation features and environmental data, thereby providing rich semantic information for subsequent graph convolution operations.

[0150] Furthermore, by mapping water accumulation features and environmental data into a graph data structure and using edges to connect these nodes, the system can structure complex spatiotemporal information, providing a natural framework for subsequent graph neural network learning. This design allows the entire information processing process to go beyond traditional pixel-level processing, taking into account multi-dimensional data interaction relationships and improving the model's robustness in flood disaster identification.

[0151] S3.3.2: The graph data is subjected to graph convolution operation through graph neural network, wherein the graph convolution operation achieves the propagation and updating of features of different nodes through the combination of multiple convolutional layers and nonlinear activation functions, captures the spatiotemporal dynamic information between nodes, and forms graph node features;

[0152] Specifically, the graph-structured data (flood features and environmental data) enters a graph neural network for processing via graph convolution operations. A graph neural network is a deep learning model that propagates and updates features through neighboring nodes. In graph convolution operations, each node in the graph not only processes its own features but also receives information from neighboring nodes, thus achieving information propagation and updating. This operation helps capture the spatiotemporal dependencies between nodes, especially in dynamic farmland environments where flood features and environmental data are often influenced by surrounding nodes. Therefore, this propagation mechanism is well-suited for flood risk assessment.

[0153] In this embodiment, graph convolution operations extract node features layer by layer through multiple convolutional layers. Each convolutional layer updates the node's features, enabling them to better reflect the relationship with neighboring nodes. For example, a waterlogged feature node will fuse information with its neighboring environmental data nodes (such as soil moisture, precipitation, etc.), making the features of each node more comprehensive and able to reflect the interactions between nodes. Through the stacking of multiple convolutional layers, the system can progressively extract higher-order features, which can better capture complex spatiotemporal dynamic changes, especially in flood risk assessment, where changes in waterlogged areas are often closely related to changes in the surrounding environment.

[0154] Furthermore, after each convolutional layer, node features are processed using a non-linear activation function (such as ReLU). The role of the activation function is to introduce non-linear relationships, enabling the model to better capture complex data patterns. The use of activation functions enhances the learning ability of graph neural networks, effectively avoiding overly simplistic linear fitting during information transmission, thereby improving the model's expressive power.

[0155] Furthermore, through this graph convolution operation, the system can accurately capture the complex relationships between different nodes, especially in farmland environments where water accumulation characteristics are closely related to surrounding environmental conditions (such as soil moisture and temperature). This process can fully utilize spatial information and temporal variation data to generate graph node features with high accuracy.

[0156] S3.3.3: Perform a convergence operation on the graph node features to generate global graph features;

[0157] In S4, when multiple flood disaster risks receive the same number of votes, the method further includes:

[0158] S4.1: Sort multiple flood disaster risks according to preset priority rules to generate a candidate sequence, wherein the priority rules are determined based on the current rate of expansion of the waterlogged area in the farmland;

[0159] S4.2: Acquire environmental data from adjacent sensor terminals and calculate the applicability scores of multiple flood disaster risks in adjacent areas;

[0160] S4.3: The candidate sequences are reordered according to the applicability score, and the flood disaster risk ranked first is taken as the flood disaster risk identification result.

[0161] The specific steps for S5 are as follows:

[0162] S5.1: Adjust the broadcast range and communication protocol according to the flood disaster risk identification results to transmit early warning information to the receiving terminal, wherein the early warning information includes the flood disaster risk identification results and the corresponding farmland images;

[0163] Specifically, different regions have different levels of flood risk, so it is necessary to adjust the scope and method of information dissemination according to the level of risk.

[0164] In this embodiment, the system first determines which areas are at high, low, or medium risk based on the flood disaster risk identification results, and then adjusts the broadcast range and communication protocol according to the risk level. Specifically, for high-risk areas, the system increases the broadcast frequency and expands the broadcast range to ensure that the early warning information can be quickly transmitted to every terminal in the disaster area; for low-risk areas, the system reduces the broadcast frequency to avoid unnecessary waste of resources.

[0165] Furthermore, the choice of communication protocol will also change according to the adjustment of the broadcast range. In high-risk areas, the system adopts low-latency, high-bandwidth communication protocols (such as 4G / 5G communication, Wi-Fi, or LPWAN) to ensure that disaster information can be transmitted to each receiving terminal in the shortest possible time; while in low-risk areas, the system can choose low-power, low-bandwidth communication protocols (such as LoRa, Zigbee, etc.) to save network resources and battery consumption.

[0166] S5.2: Perform iterative operation. The receiving terminal collects the current farmland image. If the water accumulation characteristics of the current farmland image are consistent with the water accumulation characteristics of the received farmland image, the current receiving terminal is used as the cluster head terminal to continue transmitting the warning information until the water accumulation characteristics of the current farmland image are inconsistent with the water accumulation characteristics of the received farmland image, and the broadcast of the warning information is completed.

[0167] Specifically, traditional flood warning systems often rely on fixed broadcast mechanisms for information transmission, lacking flexibility. This embodiment, however, uses an iterative approach to make information transmission more dynamic and accurate. Each time a warning is broadcast, the receiving terminal first acquires the latest farmland image, analyzes the water accumulation characteristics in the image, and compares them with the received water accumulation characteristics to determine if they match the previous information.

[0168] In this embodiment, each receiving terminal analyzes the water accumulation features in the current farmland image, extracting information such as the extent, depth, and distribution of the water accumulation. The receiving terminal first compares this information with the water accumulation features of previous farmland images received from the cluster head terminal. If the water accumulation features match, it indicates that the water accumulation status in the area has not changed, and the receiving terminal will continue to act as the cluster head terminal, transmitting the warning information to the next terminal. If the water accumulation features do not match, it indicates that the water accumulation status in the current farmland area has changed, and the system will stop the current iteration process and reassess the broadcast range before transmitting new information. This iterative process, through dynamic terminal selection and real-time status updates, ensures that information transmission is error-free. In this way, the system can maintain high flexibility and real-time performance in multi-level linkage, while avoiding information lag or misbroadcasting problems that may occur in traditional methods. Each terminal can decide whether to continue participating in information propagation based on the current water accumulation features, thereby avoiding redundant information transmission and effectively saving network bandwidth and battery resources.

[0169] Compared to traditional static broadcasting, iterative operations achieve precise early warning of farmland flooding by incorporating dynamic changes in water accumulation characteristics. It can flexibly adjust information dissemination strategies based on changes in regional water accumulation, ensuring the efficiency and accuracy of the flood warning system when dealing with complex agricultural environments.

[0170] Please see Figure 3 , Figure 3 This is a schematic diagram of farmland image broadcasting according to an embodiment of this application. Figure 3 The diagram illustrates that in the flood identification process, after the cluster terminal acquires images of farmland, it broadcasts the images to multiple connected sensor terminals via a wireless communication protocol. Figure 3 The broadcast path between the cluster head terminal and multiple sensor terminals is shown. After receiving the image, the sensor terminals combine it with local environmental data to perform image processing and flood risk assessment.

[0171] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating the feedback of flood disaster identification results in an embodiment of this application. Figure 4 This demonstrates that after each sensor terminal completes the risk assessment of the broadcast farmland image, it transmits the risk level results it identifies back to the cluster head terminal via a communication link. Figure 4 The document demonstrates the data transmission path of the identification results from each terminal and the process by which the cluster head terminal votes on and tallies the results for each risk level.

[0172] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-level linkage automatic broadcasting system for agricultural flood early warning information, applied to distributed farmland, wherein each independent farmland in the distributed farmland is equipped with interconnected sensor terminals, and the sensor terminals integrate image sensors and processing modules, characterized in that, The system includes a terminal communication module, a flood identification module, and an early warning notification module, wherein: The terminal communication module is used to respond to the water level alarm information sent by the water level gauge, wake up the sensor terminal, build a sensor terminal communication network, and select the cluster head terminal. The water level alarm information is the information triggered by the water level gauge after the water level is higher than a set threshold. The flood identification module is used to broadcast farmland images collected by the cluster head terminal to each sensor terminal, and each sensor terminal processes the farmland images to generate flood disaster risk, and returns the flood disaster risk to the cluster head terminal, selecting the judgment result with the most votes as the flood disaster risk identification result; The early warning notification module is used to broadcast the flood disaster risk identification results to other terminals in the sensor terminal communication network; The flood identification module includes: An image preprocessing unit is used to preprocess the farmland image, wherein the preprocessing includes illumination compensation and image dehazing; The feature extraction unit is used to input the preprocessed farmland image into a preset feature extraction model and output water accumulation features through the feature extraction model. The feature fusion unit is used to deeply fuse the water accumulation features with the environmental data collected by the sensing terminal itself to generate global graph features; The disaster identification unit is used to map the fused data to the corresponding flood disaster risk based on the global map features and a preset flood level classification standard; The feature fusion unit includes: The water accumulation features and the environmental data are used as nodes in the graph data, where the water accumulation features and the environmental data are respectively used as different types of nodes in the graph, and the relationships between the data are used as edges in the graph. The graph data is subjected to graph convolution operation through graph neural network. The graph convolution operation uses a combination of multiple convolutional layers and nonlinear activation functions to realize the propagation and updating of features of different nodes, capture the spatiotemporal dynamic information between nodes, and form graph node features. The early warning notification module includes: A broadcast generation unit is used to adjust the broadcast range and communication protocol according to the flood disaster risk identification results in order to transmit early warning information to the receiving terminal; The iterative reporting unit is used to compare the farmland images collected by the receiving terminal with the early warning information and broadcast the early warning information level by level. The iterative notification unit includes: The receiving terminal collects current farmland images. If the water accumulation characteristics of the current farmland image are consistent with those of the received farmland image, the receiving terminal will continue to transmit the warning information as the cluster head terminal until the water accumulation characteristics of the current farmland image and the received farmland image are inconsistent, thus completing the broadcast of the warning information.

2. The multi-level linkage automatic broadcasting system for agricultural flood early warning information according to claim 1, characterized in that, The terminal communication module includes: The terminal wake-up unit is used to activate the corresponding sensor terminal and start its working mode after receiving the water level alarm information from the water level gauge. The terminal wake-up unit interacts with the water level gauge through a wireless communication protocol to obtain water level data and determine whether the wake-up conditions are met. The network construction unit is used to establish communication links between the sensing terminals and form a sensing terminal communication network after the sensing terminals are woken up.

3. The multi-level linkage automatic broadcasting system for agricultural flood early warning information according to claim 2, characterized in that, The network construction unit includes: The awakened sensor terminal initiates a self-organizing network mechanism, establishes a communication link with neighboring terminals through a communication protocol, and forms a sensor terminal communication network. In the sensor terminal communication network, cluster head terminals are selected based on the geographical location, terrain changes, and water flow direction of the sensor terminals.

4. The multi-level linkage automatic broadcasting system for agricultural flood early warning information according to claim 1, characterized in that, The image preprocessing unit includes: Calculate the global brightness distribution of the farmland image, and calculate the illumination compensation parameters based on the global brightness distribution using histogram equalization; Farmland images are dehazed using a generative adversarial network (GAN), which is trained on both foggy and fog-free farmland images during the training phase.

5. The multi-level linkage automatic broadcasting system for agricultural flood early warning information according to claim 1, characterized in that, The feature extraction unit includes: The preprocessed farmland image is converted to the spectral domain, and local features of the spectral curve are extracted in the spectral domain using a one-dimensional convolution operation. Global features of the spectral curve are extracted using global pooling, and spectral indices are calculated based on the local and global features. Spectral indicators are fused using a feature pyramid to generate a feature map. The feature map is then processed using an attention mechanism to extract feature vectors. The feature vector is input into a fully connected layer for classification and regression operations, and the water accumulation feature is output.

6. The multi-level linkage automatic broadcasting system for agricultural flood early warning information according to claim 1, characterized in that, When multiple flood disaster risk assessments result in the same number of votes, the flood identification module further includes: Multiple flood disaster risks are sorted according to preset priority rules to generate a candidate sequence, wherein the priority rules are determined based on the current rate of expansion of waterlogged areas in farmland; Acquire environmental data from adjacent sensor terminals and calculate the applicability scores of multiple flood disaster risks in adjacent areas; The candidate sequences are reordered based on the applicability score, and the flood disaster risk ranked first is taken as the flood disaster risk identification result.

Images