A heavy rain terminal early warning method and system based on deep learning

Abstract

The present application relates to the technical field of heavy rain early warning, in particular to a heavy rain terminal early warning method and system based on deep learning, which realizes consistent determination of image features such as the overturning state of a rain gauge and rain curtain, water marks and the like by introducing an image perception and semantic recognition mechanism, effectively excludes non-natural triggering factors (such as artificial water pouring), and improves the reliability of rainfall perception data. A locally deployable double-branch deep learning model is adopted to fuse multi-source data such as rainfall, temperature and humidity, and terrain factors, cross-attention mechanism is used for feature extraction and return period prediction, and a threshold correction coefficient is dynamically generated in combination with historical events to realize adaptive adjustment of the early warning standard and effectively reduce the risk of false positives and false negatives. The system supports low-power local operation and executes early warning decisions, effectively improving the timeliness and accuracy of local early warning response.

Description

Technical Field

[0001] This invention relates to the field of rainstorm early warning technology, and more particularly to a rainstorm terminal early warning method and system based on deep learning. Background Technology

[0002] Heavy rains can easily trigger disasters such as flash floods, landslides, and urban flooding, which are particularly destructive in areas with complex terrain and weak infrastructure. Current heavy rain warning methods mainly rely on large-scale meteorological models, radar echo monitoring, and satellite remote sensing simulations. Most existing warning systems depend on central meteorological platforms and cloud servers for model calculations and decision-making logic execution. This results in their inability to operate normally in scenarios with network outages, weak networks, or poor power conditions, and also leads to high response delays.

[0003] Existing technologies, through local deployment of rain gauges in warning areas, achieve a certain degree of localized data collection and early warning. However, the following problems remain: rain gauges are susceptible to unnatural rainfall factors such as human-caused water spills during the data collection process, leading to a high false alarm rate and affecting the reliability of warnings and the accuracy of public response. Current local early warnings use fixed rainfall threshold judgment logic, without dynamically adjusting warning standards based on regional differences, historical event statistical characteristics, and real-time meteorological conditions, which easily results in false or missed warnings.

[0004] Therefore, there is a need for a method and system for early warning of rainstorms that is suitable for local deployment and has high accuracy. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for early warning of rainstorm terminals based on deep learning.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning-based method for early warning of rainstorms via terminals, comprising:

[0008] S1. Collect rainfall sensing data and image data within the target area, wherein the rainfall sensing data includes rainfall data and temperature and humidity data;

[0009] S2. Based on the image data, identify the rainfall status, output the rainfall scene verification result, and filter the valid rainfall perception data according to the rainfall scene verification result;

[0010] S3. Input the effective rainfall sensing data into a dual-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period and dynamic threshold correction coefficient corresponding to the target area.

[0011] S4. Based on the dynamic threshold correction coefficient, adjust the engineering threshold parameters of the current area to obtain the dynamic threshold, compare the real-time rainfall data with the dynamic threshold, generate rainstorm warning information based on the comparison result, and control the terminal display device to complete the warning action output based on the rainstorm warning information.

[0012] Furthermore, the rainfall data is acquired by setting up tipping bucket rain gauges within the target area, and the temperature and humidity data is acquired by distributing temperature and humidity sensors within the target area.

[0013] Furthermore, the image data is acquired by setting up a camera within the target area, and the image data includes an image of the area where the tipping bucket rain gauge is set.

[0014] Furthermore, the rainfall status identification based on the image data includes the following steps:

[0015] Structured detection is performed on the original image data to identify and extract the spatial location of the tipping bucket rain gauge in the image, thus obtaining the corresponding tipping bucket image region;

[0016] Based on the tipping bucket image region, an image sequence is constructed. Through inter-frame pixel difference and edge contour extraction algorithms, tipping bucket state recognition is performed on continuous image frames to extract the tipping bucket flipping start and end frames, flipping amplitude and flipping frequency, and generate structural motion feature data representing the tipping bucket action state.

[0017] Based on the background image region excluding the tipping bucket area in the original image data, the rainfall semantic feature vectors of rain curtain texture, water droplet trajectory and ground water spread are extracted by a convolutional neural network model;

[0018] The consistency between the structural motion feature data and the rainfall semantic feature vector is determined. If there is a semantic feature distribution in the image within the time period corresponding to the tipping bucket flipping behavior that is consistent with natural rainfall, then the rainfall scene verification result is generated as triggered by natural rainfall; if there is a deviation between the tipping bucket structural motion feature and the semantic feature distribution, then the rainfall scene verification result is generated as triggered by non-natural rainfall.

[0019] Furthermore, the dual-branch deep learning model based on cross-attention is constructed through the following steps:

[0020] Based on valid rainfall sensing data verified by rainfall scenarios in historical rainstorm events, a coefficient scan is performed on the rainfall sequence and corresponding engineering threshold for each event. The false alarm rate and false alarm rate of each candidate coefficient are calculated to obtain the optimal threshold correction coefficient that minimizes the false alarm rate and does not exceed the preset upper limit. The optimal threshold correction coefficient and the actual rainstorm recurrence period level of the corresponding rainstorm event are used together to form a supervision label set.

[0021] Topographic factors are constructed based on the terrain type, altitude, and environmental category of the rainstorm event location, and rainfall and temperature / humidity sequences are constructed based on the effective rainfall sensing data of the rainstorm event location.

[0022] A multi-source training sample set is constructed using topographic factors, rainfall sequences, and temperature and humidity sequences, and the multi-source training sample set is associated with a supervision label set.

[0023] Based on the multi-source training sample set, temporal location embedding and cross-channel attention fusion are performed on the rainfall sequence, temperature and humidity sequence and topographic factor encoding to obtain the spatiotemporal joint feature sequence;

[0024] Based on the spatiotemporal joint feature sequence, the backbone network containing residual connections is input to obtain the prediction vector for return period prediction and the correction vector for dynamic threshold regression, and the return period error and threshold error are calculated according to the supervision label.

[0025] By performing gradient optimization on the network parameters using the loss function, a two-branch deep learning model is obtained.

[0026] Furthermore, the loss function is as follows:

[0027] ;

[0028] in, This is the loss value; This represents the number of training samples; For the first Predicted return period rank for each sample; For the first True recurrence period gradation labels for each sample; For the first The prediction threshold correction coefficient for each sample; For the first The optimal threshold correction coefficient label for each sample; For the first The historical false alarm rate weights corresponding to each sample; This is the weighting factor.

[0029] Furthermore, the generation of rainstorm warning information based on the comparison results includes:

[0030] Real-time rainfall data is acquired and compared with the dynamic threshold. If the real-time rainfall data is greater than or equal to the dynamic threshold, a rainstorm warning is generated; if the rainfall data in the previous time period is less than the dynamic threshold, no rainstorm warning is generated.

[0031] Furthermore, the rainstorm warning information includes: target area identifier, warning trigger time, real-time rainfall data, dynamic threshold, predicted rainstorm recurrence period level, and warning level identifier.

[0032] Furthermore, the warning actions include: changing the color of the visual interface, pop-up prompts, voice broadcasts, and buzzer sounds.

[0033] A deep learning-based rainstorm terminal early warning system, applied to any of the aforementioned deep learning-based rainstorm terminal early warning methods, includes: a data perception layer, an operation and maintenance layer, a core computing layer, and an early warning decision layer connected in sequence;

[0034] The data sensing layer is used to collect rainfall sensing data and image data within the target area. The rainfall sensing data includes rainfall data and temperature and humidity data.

[0035] The operation and maintenance layer is used to identify rainfall status based on the image data, output rainfall scene verification results, and filter valid rainfall sensing data based on the rainfall scene verification results;

[0036] The core computing layer is used to input the effective rainfall sensing data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period corresponding to the target area and the dynamic threshold correction coefficient used to correct the engineering threshold.

[0037] The early warning decision layer is used to adjust the engineering threshold parameters of the current area based on the dynamic threshold correction coefficient to obtain the dynamic threshold, compare the real-time rainfall data with the dynamic threshold, generate rainstorm early warning information based on the comparison result, and control the terminal display device to complete the early warning action output based on the early warning information.

[0038] The beneficial effects of this invention are as follows: By introducing image perception and semantic recognition mechanisms, this invention visually verifies the actual rainfall triggering status of rain gauges within the designated area. It extracts image features such as rain curtains, watermarks, and tipping bucket rotation, and performs semantic consistency determination with rain gauge data, significantly reducing false alarms caused by non-natural factors (such as human spillage or accidental triggering), thus improving the reliability of the perceived data. Through a branch deep learning model deployable on local terminals, a dual-branch deep learning model is constructed based on a cross-attention structure. On one hand, it performs feature fusion and pattern learning on the input multi-source rainfall perception data (such as rainfall, temperature and humidity, and terrain factors), and generates dynamic threshold correction coefficients based on historical event samples. Combined with regional characteristics and recurrence interval levels, it achieves adaptive adjustment of the warning threshold, thereby avoiding the false or missed warnings caused by traditional static threshold methods. This system can autonomously operate in extreme environments such as network outages, weak network conditions, and power outages using low-power LoRa, ensuring timely generation and on-site output of warning information, greatly enhancing the response speed and deployment flexibility at the initial stage of a disaster. Attached Figure Description

[0039] Figure 1 This is a flowchart of the steps of a rainstorm terminal early warning method based on deep learning in this invention.

[0040] Figure 2 This is a flowchart illustrating the construction steps of the dual-branch deep learning model based on cross-attention in this invention. Detailed Implementation

[0041] Please see Figures 1-2 As shown, this invention relates to a deep learning-based method for early warning of rainstorms via a terminal, comprising:

[0042] S1. Collect rainfall sensing data and image data within the target area, wherein the rainfall sensing data includes rainfall data and temperature and humidity data;

[0043] S2. Based on the image data, identify the rainfall status, output the rainfall scene verification result, and filter the valid rainfall perception data according to the rainfall scene verification result;

[0044] S3. Input the effective rainfall sensing data into a dual-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period and dynamic threshold correction coefficient corresponding to the target area.

[0045] S4. Based on the dynamic threshold correction coefficient, adjust the engineering threshold parameters of the current area to obtain the dynamic threshold, compare the real-time rainfall data with the dynamic threshold, generate rainstorm warning information based on the comparison result, and control the terminal display device to complete the warning action output based on the rainstorm warning information.

[0046] In some embodiments, a tipping bucket rain gauge and a temperature and humidity sensor are deployed inside the target area to obtain the original rainfall and temperature and humidity parameters. At the same time, a continuous image sequence covering the area of the rain gauge and its surrounding environment is collected by setting up a camera. In step S2, through image structuring processing, the area of the rain gauge in the image is automatically identified and its tipping action features (such as the start and end frames of tipping, tipping amplitude, frequency, etc.) are extracted. At the same time, a convolutional neural network is used to semantically encode the rain curtain, water trace diffusion texture, etc. in the background image area to form a visual semantic feature vector. Subsequently, the physical tipping structure features are matched with the visual semantic features. When they are consistent, it is determined that natural rainfall is triggered; otherwise, it is identified as non-natural interference (such as artificial pouring of water), which can significantly reduce the false alarms caused by the mis-triggering of the rain gauge and improve the local perception credibility. In step S3, the system inputs the effective rainfall perception data screened by semantic verification into a pre-trained dual-branch deep learning model based on cross-attention. This model consists of a recurrence period branch and a threshold correction branch, which respectively predict the heavy rain recurrence period level in the future time period and the adjustment coefficient required to correct the static engineering threshold of the current area. Among them, the model input not only includes the current rainfall and temperature and humidity sequences, but also includes regional background factors such as terrain type, altitude, and environmental category. Through the temporal position embedding and inter-channel attention fusion mechanism, a deep feature expression with spatio-temporal consistency is constructed, significantly improving the adaptability and generalization ability of the model to complex regional differences. In step S4, the engineering preset threshold is adjusted using the dynamically predicted threshold correction coefficient obtained above, and the rainfall data of the current time period is compared in real time. If the current rainfall intensity reaches or exceeds the dynamically corrected threshold, a heavy rain warning message including the target area, real-time rainfall value, recurrence period level, and warning level is generated locally, and the local linkage response of the terminal device is triggered, such as screen highlighting, voice broadcast, beeping reminder, etc. The local decision-making ability significantly enhances the independent operation ability of the system in a network-disconnected or extreme disaster environment, avoiding the shortcoming of traditional cloud computing architecture response delay and network dependence.

[0047] In a preferred embodiment, the rainfall sensing data is not limited to rainfall amount and temperature / humidity data, but may also include air pressure, wind speed and direction, and water level. The equipment specifically includes: a rain gauge for acquiring instantaneous and cumulative ground rainfall; a digital temperature and humidity sensor for collecting atmospheric temperature and relative humidity; a micro-pressure sensor for recording real-time air pressure change trends; an ultrasonic anemometer for collecting wind speed and direction changes; and an ultrasonic or radar water level sensor for measuring surface or river water level changes. The data collected by these sensors, after initial cleaning and timestamp alignment by edge computing nodes, constitutes a multi-dimensional rainfall sensing dataset. For this multi-dimensional rainfall sensing data, in model construction, all dimensions of rainfall sensing variables are normalized using a unified encoder structure and constructed into a unified multi-source input tensor according to time alignment, which is then input into a dual-branch deep learning model. This not only provides richer contextual information for determining the recurrence interval of heavy rainfall but also provides more comprehensive environmental parameter support for dynamic threshold correction, improving the system's ability to distinguish extreme weather events in their early stages.

[0048] Furthermore, the rainfall data is acquired by setting up tipping bucket rain gauges within the target area, and the temperature and humidity data is acquired by distributing temperature and humidity sensors within the target area.

[0049] Specifically, the rainfall data is collected using a tipping bucket rain gauge, which boasts advantages such as simple structure, high response, and low power consumption, making it suitable for long-term deployment in remote areas or regions with weak infrastructure. The tipping bucket rain gauge quantifies and records rainfall by driving a tipping mechanism to rotate the collected rainwater. Each rotation represents a fixed volume of rainfall, thus achieving accurate rainfall data collection. To further enhance the multi-dimensional coverage of rainfall perception, a distributed array of temperature and humidity sensors is deployed within the target area to monitor local air temperature and humidity changes in real time, compensating for potential blind spots in single rain gauges. By fusing temperature and humidity data with rainfall information, a more comprehensive understanding of the meteorological dynamics within the area is achieved, providing a richer data foundation for subsequent model temporal feature extraction, spatial attribution, and semantic consistency judgment.

[0050] Furthermore, the image data is acquired by setting up a camera within the target area, and the image data includes an image of the area where the tipping bucket rain gauge is set.

[0051] It should be noted that the image data is acquired by deploying camera equipment in the target area. The camera installation positions are strategically positioned to ensure that their field of view covers the installation area of ​​the tipping bucket rain gauge, thereby enabling real-time acquisition of image sequences containing the rain gauge and its surrounding environment. This setup not only captures changes in the physical state of the rain gauge but also captures visual information features during rainfall, including rain curtain textures, water droplet trajectories, and ground water diffusion. By processing the image data acquired by the camera, a continuous image sequence can be constructed to analyze the semantic consistency between the rain gauge's tipping behavior and external rainfall characteristics. This image fusion method effectively compensates for the shortcoming of traditional rain gauges, which cannot self-identify when subjected to human interference (such as intentional water injection or accidental impact). The image data provided by the camera provides the necessary input foundation for subsequent image recognition algorithms and is the core information source for constructing an image-verified rainfall authenticity recognition mechanism in this scheme.

[0052] Furthermore, the rainfall status identification based on the image data includes the following steps:

[0053] Structured detection is performed on the original image data to identify and extract the spatial location of the tipping bucket rain gauge in the image, thus obtaining the corresponding tipping bucket image region;

[0054] Based on the tipping bucket image region, an image sequence is constructed. Through inter-frame pixel difference and edge contour extraction algorithms, tipping bucket state recognition is performed on continuous image frames to extract the tipping bucket flipping start and end frames, flipping amplitude and flipping frequency, and generate structural motion feature data representing the tipping bucket action state.

[0055] Based on the background image region excluding the tipping bucket area in the original image data, the rainfall semantic feature vectors of rain curtain texture, water droplet trajectory and ground water spread are extracted by a convolutional neural network model;

[0056] The consistency between the structural motion feature data and the rainfall semantic feature vector is determined. If there is a semantic feature distribution in the image within the time period corresponding to the tipping bucket flipping behavior that is consistent with natural rainfall, then the rainfall scene verification result is generated as triggered by natural rainfall; if there is a deviation between the tipping bucket structural motion feature and the semantic feature distribution, then the rainfall scene verification result is generated as triggered by non-natural rainfall.

[0057] In some embodiments, the original image data is first subjected to structured detection, specifically using a target detection model based on YOLOv5 or Faster R-CNN to locate the tipping bucket rain gauge in the image frame. This detection process obtains the bounding box coordinates of the rain gauge in the image, thereby cropping out a stable tipping bucket image region containing the target device, providing a focused area for subsequent action recognition processing. After spatial localization, the system constructs an image sequence based on the tipping bucket image region. For this image sequence, an inter-frame pixel difference algorithm is applied to extract grayscale variation regions in consecutive frames to capture the dynamic changes of the tipping bucket device. This is combined with an edge contour extraction algorithm based on the Canny or Sobel operator to further enhance the extraction effect of the tipping bucket edge structure. By setting a threshold for the flipping angle change and a continuous frame count condition, the start and end frames of the tipping bucket flip can be identified. Simultaneously, the flipping amplitude and the number of flips per unit time are statistically analyzed, ultimately generating a structural motion feature data sequence to quantify the behavioral characteristics of the tipping bucket action within that time period. To enhance the robustness of scene recognition, the system further performs semantic feature extraction on the background image region other than the tipping bucket region in the original image data. This step introduces a convolutional neural network as the basic feature encoder, combining multi-scale receptive fields to extract key semantic features related to natural rainfall from the image, such as the flow texture of the rain curtain, the tracks formed by water droplets on the lens surface, and the direction and density of water spread on the ground. The high-dimensional feature vector output by the network is mapped through a fully connected layer to form a standardized semantic representation. Finally, the structural motion feature data and the rainfall semantic feature vector are input into a similarity matching module, and consistency is determined using metrics such as cosine similarity or Mahalanobis distance. If the flipping behavior of the bucket within the specified time period is highly consistent with the semantic image features, it is inferred to be caused by natural rainfall, and the output rainfall scene verification result is "triggered by natural rainfall"; if the two features differ significantly, it is determined that there are abnormal triggering factors (such as human interference), and the output verification result is "triggered by non-natural rainfall".

[0058] Furthermore, the dual-branch deep learning model based on cross-attention is constructed through the following steps:

[0059] Based on valid rainfall sensing data verified by rainfall scenarios in historical rainstorm events, a coefficient scan is performed on the rainfall sequence and corresponding engineering threshold for each event. The false alarm rate and false alarm rate of each candidate coefficient are calculated to obtain the optimal threshold correction coefficient that minimizes the false alarm rate and does not exceed the preset upper limit. The optimal threshold correction coefficient and the actual rainstorm recurrence period level of the corresponding rainstorm event are used together to form a supervision label set.

[0060] Topographic factors are constructed based on the terrain type, altitude, and environmental category of the rainstorm event location, and rainfall and temperature / humidity sequences are constructed based on the effective rainfall sensing data of the rainstorm event location.

[0061] A multi-source training sample set is constructed using topographic factors, rainfall sequences, and temperature and humidity sequences, and the multi-source training sample set is associated with a supervision label set.

[0062] Based on the multi-source training sample set, temporal location embedding and cross-channel attention fusion are performed on the rainfall sequence, temperature and humidity sequence and topographic factor encoding to obtain the spatiotemporal joint feature sequence;

[0063] Based on the spatiotemporal joint feature sequence, the backbone network containing residual connections is input to obtain the prediction vector for return period prediction and the correction vector for dynamic threshold regression, and the return period error and threshold error are calculated according to the supervision label.

[0064] By performing gradient optimization on the network parameters using the loss function, a two-branch deep learning model is obtained.

[0065] In some embodiments, firstly, a supervisory label is constructed based on valid rainfall sensing data from historical rainstorm events, verified as natural rainfall through image verification. For each event's rainfall change curve, a set of candidate correction coefficients is set, and the original engineering warning threshold is adjusted weighted accordingly. The false alarm rate and false negative rate under this correction condition are calculated. A heuristic search strategy is adopted, selecting the correction coefficient with the lowest false alarm rate as the optimal threshold correction coefficient for that event, provided the false negative rate does not exceed a preset tolerance. Simultaneously, a corresponding dual-label supervisory structure is formed by combining the actual rainstorm recurrence period level marked in the meteorological data for subsequent model training. During feature construction, the system constructs terrain factors based on the terrain type (e.g., hills, mountains, plains), altitude, and environmental category (e.g., urban-rural distribution) of the rainstorm event area. Combined with rainfall and temperature / humidity sequences collected by sensing terminals within the same area, three types of heterogeneous input data are generated. To ensure the model's ability to perceive temporal evolution information, positional encoding is performed on each type of sequence, introducing time step identifiers to enhance the ability to capture phased feature changes. Subsequently, structured topographic factors are constructed by combining the terrain type, altitude, and environmental category of the event location. These, along with corresponding rainfall and temperature / humidity sequences, constitute multi-source input features, forming a unified multimodal training sample set. In the feature encoding stage, the model first embeds the temporal positions of the time-series data from different modalities to preserve their temporal evolution information during the rainstorm formation process. Furthermore, a cross-channel attention mechanism is introduced to achieve information flow and dependency modeling between modalities. Specifically, the cross-channel attention module generates statistical descriptions by globally pooling the three input channels (rainfall, temperature / humidity, and topographic factors), then calculates channel attention weights via a shared weight network. This adaptively adjusts the importance of each channel feature during fusion, enhancing the model's ability to focus on key influencing factors. Residual connections preserve the basic perception layer features, ultimately generating a structurally stable spatiotemporal joint feature representation. At the output, the model has two branches: one for predicting the rainstorm recurrence interval level, and the other for regressing the dynamic threshold correction coefficient. During the training stage, a joint loss function is defined with the constructed supervision label set as the target. Finally, the parameters are iteratively updated using the Adam optimizer to achieve training convergence of the two-branch model.

[0066] Furthermore, the loss function is as follows:

[0067] ;

[0068] in, This is the loss value; This represents the number of training samples; For the first Predicted return period rank for each sample; For the first True recurrence period gradation labels for each sample; For the first The prediction threshold correction coefficient for each sample; For the first The optimal threshold correction coefficient label for each sample; For the first The historical false alarm rate weights corresponding to each sample; This is the weighting factor.

[0069] It should be noted that the loss function consists of two sub-terms, corresponding to the prediction error of the return period level and the threshold correction error, respectively. The first term measures the regression accuracy of the model for the return period level of the rainstorm, and its value is the squared error between the model's predicted level and the actual level; this part ensures that the model can effectively distinguish different levels of rainfall events and provide graded risk warnings. The second term weights the prediction error of the dynamic threshold correction coefficient, introducing the historical false alarm rate as a weighting factor, giving higher training penalty intensity to samples prone to false alarms, significantly reducing the probability of false alarms in practical applications. At the same time, to balance the relative importance of the two prediction objectives in the training process, an adjustable weight coefficient is set in the loss function to control the contribution ratio between the return period error term and the threshold correction error term. Through this structural design, the loss function not only improves the model's multi-objective learning ability, but also enhances its robustness and adaptability to abnormal rainfall events in edge scenarios, effectively making up for the problems of traditional early warning methods, such as the inability to dynamically adjust the threshold and the high false alarm rate.

[0070] Furthermore, the generation of rainstorm warning information based on the comparison results includes:

[0071] Real-time rainfall data is acquired and compared with the dynamic threshold. If the real-time rainfall data is greater than or equal to the dynamic threshold, a rainstorm warning is generated; if the rainfall data in the previous time period is less than the dynamic threshold, no rainstorm warning is generated.

[0072] Specifically, the process of generating rainstorm warning information based on the comparison results relies on a precise comparison logic between real-time rainfall and a dynamic threshold. The system first continuously receives real-time rainfall data collected by tipping bucket rain gauges within the target area and compares it against a threshold dynamically calculated using a deep learning model, time-by-time. This dynamic threshold comprehensively considers regional terrain features, the recurrence interval of historical rainfall events, and false alarm control factors, allowing for flexible adjustment of warning standards for different scenarios. During the comparison process, it determines whether the actual rainfall in the current time period has reached or exceeded the dynamic threshold. If the comparison result "meets" the condition, meaning the real-time rainfall intensity has reached the dynamically set critical level, a corresponding rainstorm warning is immediately generated as a trigger signal for disaster response. Simultaneously, this warning information carries structured fields including regional identifier, trigger time, current rainfall and threshold, and warning level for subsequent decision-making and terminal display. Conversely, if the rainfall in the current time period has not yet reached the dynamic threshold, it is determined that a rainstorm trigger situation does not exist. In this case, the system does not generate a warning to avoid false alarms due to minor rainfall, thereby improving the overall reliability of the warning and the accuracy of public response. This mechanism differs significantly from the traditional static threshold method, possessing stronger environmental adaptability and error control capabilities, and is suitable for local autonomous early warning applications under complex meteorological backgrounds.

[0073] Furthermore, the rainstorm warning information includes: target area identifier, warning trigger time, real-time rainfall data, dynamic threshold, predicted rainstorm recurrence period level, and warning level identifier.

[0074] Furthermore, the warning actions include: changing the color of the visual interface, pop-up prompts, voice broadcasts, and buzzer sounds.

[0075] In some embodiments, visual interface color changes refer to the automatic switching of background color or border highlight color of the terminal interface under different warning levels, such as yellow, orange, red, etc., corresponding to the national meteorological warning level, to enhance risk perception; pop-up prompts are warning windows that pop up on the terminal screen to display key information such as current rainfall, dynamic threshold, return period level, and warning level, so that on-site managers can quickly understand the current risk status; voice broadcasting is the real-time broadcasting of the current warning information through built-in or external speakers, such as "The current area has triggered a ten-year return period rainstorm warning, please take precautions immediately", to ensure effective reception even when vision is limited or the operator is not in front of the terminal; buzzer sounding is a high-intensity sound alarm device that starts to sound continuously or intermittently when a preset level is reached, to strengthen the warning signal in noisy or visually and aurally limited scenarios.

[0076] The present invention also includes a rainstorm terminal early warning system based on deep learning, applied to the rainstorm terminal early warning method based on deep learning described in any of the preceding claims, comprising: a data perception layer, an operation and maintenance layer, a core computing layer and an early warning decision layer connected in sequence;

[0077] The data sensing layer is used to collect rainfall sensing data and image data within the target area. The rainfall sensing data includes rainfall data and temperature and humidity data.

[0078] The operation and maintenance layer is used to identify rainfall status based on the image data, output rainfall scene verification results, and filter valid rainfall sensing data based on the rainfall scene verification results;

[0079] The core computing layer is used to input the effective rainfall sensing data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period corresponding to the target area and the dynamic threshold correction coefficient used to correct the engineering threshold.

[0080] The early warning decision layer is used to adjust the engineering threshold parameters of the current area based on the dynamic threshold correction coefficient to obtain the dynamic threshold, compare the real-time rainfall data with the dynamic threshold, generate rainstorm early warning information based on the comparison result, and control the terminal display device to complete the early warning action output based on the early warning information.

[0081] Specifically, the deep learning-based rainstorm terminal early warning system adopts a modular integrated design, integrating the operation and maintenance layer, core computing layer, and early warning decision layer into a screen host device with embedded computing capabilities and graphical visualization output capabilities. This screen host can be an all-in-one device with an edge computing chipset, supporting Linux or lightweight operating system environments. It integrates a local neural network inference engine and a graphical user interface display module, enabling it to perform deep learning model inference calculations locally and present early warning results and data visualization interfaces in real time, meeting the requirements for independent operation and information prompts in environments with network outages or weak network conditions. The data sensing layer is connected to the screen host via a low-power wireless communication protocol or a wired connection, enabling data access from multiple sensing devices. Rainfall data is collected by connecting to a tipping bucket rain gauge, temperature and humidity data are collected by distributed temperature and humidity sensors deployed in the target area, and image data is acquired through fixed cameras. All sensing devices support communication with the screen host via LoRa, and the data is uniformly scheduled and cached by the host, forming a standardized input format for deep learning processing. This integrated deployment approach significantly reduces reliance on cloud platforms and remote networks, improves the system's accessibility and response time in remote and underdeveloped areas, and provides stronger autonomous early warning capabilities and on-site disaster prevention support.

[0082] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based heavy rain terminal warning method, characterized by, include: S1. Collect rainfall sensing data and image data within the target area, wherein the rainfall sensing data includes rainfall data and temperature and humidity data; S2. Based on the image data, identify the rainfall status, output the rainfall scene verification result, and filter the valid rainfall sensing data according to the rainfall scene verification result; S3. Input the effective rainfall sensing data into a dual-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period and dynamic threshold correction coefficient corresponding to the target area. S4. Based on the dynamic threshold correction coefficient, the engineering threshold of the current area is adjusted to obtain the dynamic threshold. The real-time rainfall data is compared with the dynamic threshold. Based on the comparison result, rainstorm warning information is generated, and the terminal display device is controlled to complete the warning action output based on the rainstorm warning information. The rainfall status identification based on the image data includes the following steps: Structured detection is performed on the original image data to identify and extract the spatial location of the tipping bucket rain gauge in the image, thus obtaining the corresponding tipping bucket image region; Based on the tipping bucket image region, an image sequence is constructed. Through inter-frame pixel difference and edge contour extraction algorithms, tipping bucket state recognition is performed on continuous image frames to extract the tipping bucket flipping start and end frames, flipping amplitude and flipping frequency, and generate structural motion feature data representing the tipping bucket action state. Based on the background image region excluding the tipping bucket area in the original image data, the rainfall semantic feature vectors of rain curtain texture, water droplet trajectory and ground water spread are extracted by a convolutional neural network model; The consistency between the structural motion feature data and the rainfall semantic feature vector is determined. If there is a semantic feature distribution in the image within the time period corresponding to the tipping bucket flipping behavior that is consistent with natural rainfall, then the rainfall scene verification result is generated as triggered by natural rainfall; if there is a deviation between the tipping bucket structural motion feature and the semantic feature distribution, then the rainfall scene verification result is generated as triggered by non-natural rainfall. The cross-attention-based dual-branch deep learning model is constructed through the following steps: Based on valid rainfall sensing data verified by rainfall scenarios in historical rainstorm events, a coefficient scan is performed on the rainfall sequence and corresponding engineering threshold for each event. The false alarm rate and false alarm rate of each candidate coefficient are calculated to obtain the optimal threshold correction coefficient that minimizes the false alarm rate and does not exceed the preset upper limit. The optimal threshold correction coefficient and the actual rainstorm recurrence period level of the corresponding rainstorm event are used together to form a supervision label set. Topographic factors are constructed based on the terrain type, altitude, and environmental category of the rainstorm event location, and rainfall and temperature / humidity sequences are constructed based on the effective rainfall sensing data of the rainstorm event location. A multi-source training sample set is constructed using topographic factors, rainfall sequences, and temperature and humidity sequences, and the multi-source training sample set is associated with a supervision label set. Based on the multi-source training sample set, temporal location embedding and cross-channel attention fusion are performed on the rainfall sequence, temperature and humidity sequence and topographic factor encoding to obtain the spatiotemporal joint feature sequence; Based on the spatiotemporal joint feature sequence, the backbone network containing residual connections is input to obtain the prediction vector for return period prediction and the correction vector for dynamic threshold regression, and the return period error and threshold error are calculated according to the supervision label. By performing gradient optimization on the network parameters using the loss function, a two-branch deep learning model is obtained. 2.The heavy rain terminal warning method based on deep learning according to claim 1, wherein, The rainfall data is obtained by setting up tipping bucket rain gauges in the target area, and the temperature and humidity data is obtained by setting up temperature and humidity sensors in a distributed manner in the target area. 3.The heavy rain terminal warning method based on deep learning according to claim 2, characterized in that, The image data is acquired by setting up a camera in the target area, and the image data includes images of the area where the tipping bucket rain gauge is set. 4.The heavy rain terminal warning method based on deep learning according to claim 1, characterized in that, The loss function is as follows: ； in, This is the loss value; This represents the number of training samples; For the first Predicted return period rank for each sample; For the first True recurrence period gradation labels for each sample; For the first The prediction threshold correction coefficient for each sample; For the first The optimal threshold correction coefficient label for each sample; For the first The historical false alarm rate weights corresponding to each sample; This is the weighting factor.

5. The method for early warning of rainstorms based on deep learning according to claim 1, characterized in that, The process of generating rainstorm warning information based on the comparison results includes: Real-time rainfall data is acquired and compared with the dynamic threshold. If the real-time rainfall data is greater than or equal to the dynamic threshold, a rainstorm warning is generated; if the rainfall data in the previous time period is less than the dynamic threshold, no rainstorm warning is generated.

6. The method for early warning of rainstorms based on deep learning according to claim 5, characterized in that, The rainstorm warning information includes: target area identifier, warning trigger time, real-time rainfall data, dynamic threshold, predicted rainstorm recurrence period level, and warning level identifier.

7. A method for early warning of rainstorms based on deep learning according to claim 6, characterized in that, The warning actions include: changing the color of the visual interface, pop-up prompts, voice broadcasts, and buzzer sounds.

8. A deep learning-based rainstorm terminal early warning system, applied to the deep learning-based rainstorm terminal early warning method according to any one of claims 1-7, characterized in that, include: The data perception layer, operation and maintenance layer, core computing layer, and early warning decision layer are connected sequentially. The data sensing layer is used to collect rainfall sensing data and image data within the target area. The rainfall sensing data includes rainfall data and temperature and humidity data. The operation and maintenance layer is used to identify rainfall status based on the image data, output rainfall scene verification results, and filter valid rainfall sensing data based on the rainfall scene verification results; The core computing layer is used to input the effective rainfall sensing data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm recurrence period corresponding to the target area and the dynamic threshold correction coefficient used to correct the engineering threshold. The early warning decision layer is used to adjust the engineering threshold parameters of the current area based on the dynamic threshold correction coefficient to obtain the dynamic threshold, compare the real-time rainfall data with the dynamic threshold, generate rainstorm early warning information based on the comparison result, and control the terminal display device to complete the early warning action output based on the early warning information.

Images