A weather recognition method, system, device and medium based on a radar echo map

Abstract

The application discloses a weather recognition method, system and device based on a radar echo map and a medium, relates to the technical field of image recognition and weather warning, and comprises the following steps: obtaining a radar echo map of a target area; performing image segmentation on the radar echo map of the target area to obtain a plurality of regions of interest; performing abnormal target detection on each region of interest by using a target detection algorithm to obtain one or more abnormal targets and the positions of each abnormal target; the abnormal targets are thunderstorm, squall line, bow echo, hook echo and hailstorm; and a long short-term memory network model algorithm is used to predict weather changes in a future period in a position area of each abnormal target. The application can intelligently recognize the degree of thunderstorm and abnormal weather.

Description

Technical Field

[0001] This invention relates to the field of image recognition and weather warning technology, and in particular to a weather identification method, system, electronic device and storage medium based on image recognition technology and radar echo maps. Background Technology

[0002] Image recognition technology is the technology by which computers process, analyze, and understand images to identify targets and objects of various patterns. It forms the basis for practical technologies such as stereo vision, motion analysis, and data fusion, and has significant applications in fields such as navigation, map and terrain registration, natural resource analysis, weather forecasting, environmental monitoring, and physiological disease research.

[0003] Radar echo maps are a powerful tool for meteorological stations to create short-term forecasts of severe convective weather, heavy rain, and general precipitation. They consist of a local map and irregularly shaped color blocks, ranging from blue to green, yellow, orange, red, and purple. On the radar map, the color represents the intensity of the radar echo; a gradual change from blue to purple indicates increasing echo intensity and thus, gradually increasing rainfall intensity. dBZ, or reflectivity factor, indicates higher precipitation intensity. Generally, brighter areas on the radar echo map are more likely to experience severe convective weather, especially the yellow, orange, and red areas, which are more prone to severe thunderstorms. The shape of the basic reflectivity can be used to classify the severity of thunderstorms. Currently, the functionality of intelligently identifying thunderstorm severity and abnormal weather based on radar echo maps is still relatively weak. Summary of the Invention

[0004] The purpose of this invention is to provide a weather identification method, system, device, and medium based on radar echo maps, which can intelligently identify the severity of thunderstorms and abnormal weather.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] In a first aspect, the present invention provides a weather identification method based on radar echo maps, comprising:

[0007] Acquire radar echo images of the target area;

[0008] Image segmentation is performed on the radar echo image of the target area to obtain multiple regions of interest;

[0009] Anomaly detection algorithms are used to detect anomalies in each region of interest, resulting in one or more anomalies and their locations. The anomalies are thunderstorms, squall lines, bow echoes, hook echoes, and hailstorms.

[0010] The Long Short-Term Memory (LSTM) network model algorithm is used to predict weather changes in the area where each anomalous target is located in the future.

[0011] Optionally, the step of performing image segmentation on the radar echo image of the target area to obtain multiple regions of interest specifically includes:

[0012] A feature-enhanced segmentation method is used to segment the radar echo image of the target area to obtain multiple regions of interest.

[0013] Optionally, the step of using a feature-enhanced segmentation method to segment the radar echo image of the target region to obtain multiple regions of interest specifically includes:

[0014] The radar echo image of the target area is input into an image segmentation convolutional neural network model to obtain multiple regions of interest;

[0015] The training process of the image segmentation convolutional neural network model is as follows:

[0016] Construct a convolutional neural network;

[0017] Construct sample data; the sample data includes multiple radar echo images of sample areas and multiple regions of interest corresponding to each radar echo image of a sample area; the multiple regions of interest corresponding to the radar echo images of sample areas are obtained by superpixel segmentation of the radar echo images of sample areas using the SLIC algorithm.

[0018] The radar echo map of the sample area is input into the convolutional neural network to obtain the prediction result map. The cross-entropy loss function is used to calculate the error between the prediction result map and the real result map, and the error is backpropagated to update the parameters of the convolutional neural network model, thereby obtaining the image segmentation convolutional neural network model.

[0019] The predicted result image is a series of regions of interest obtained by inputting the radar echo image of the sample region into a convolutional neural network; the actual result image is a series of regions of interest obtained by superpixel segmentation of the radar echo image of the sample region using the SLIC algorithm.

[0020] Optionally, the step of using a target detection algorithm to detect abnormal targets in the region of interest, obtaining one or more abnormal targets and the location region of each abnormal target, specifically includes:

[0021] The APE-YOLO network model algorithm is used to detect abnormal targets in the region of interest, resulting in one or more abnormal targets and the location region of each abnormal target.

[0022] The APE-YOLO network model algorithm is obtained by first reducing the structural depth of the YOLONano network using the ape-walking edge computing model algorithm to obtain the APE-YOLO backbone network, and then adding dense connection blocks to the APE-YOLO backbone network.

[0023] Optionally, the step of employing a Long Short-Term Memory (LSTM) network model algorithm to predict future weather changes within the area where each anomalous target is located specifically includes:

[0024] A long short-term memory network model algorithm is used to predict the radar echo map and radar reflectivity of the area where each abnormal target is located in the future time period.

[0025] Based on radar echo maps and radar reflectivity for future periods, predict weather changes in the area where each anomalous target is located for future periods.

[0026] Secondly, the present invention provides a weather identification system based on radar echo maps, comprising:

[0027] Radar echo image acquisition module, used to acquire radar echo images of the target area;

[0028] The region of interest determination module is used to perform image segmentation on the radar echo image of the target area to obtain multiple regions of interest.

[0029] The abnormal target location calculation module is used to perform abnormal target detection on each region of interest using a target detection algorithm, and obtain one or more abnormal targets and the location of each abnormal target; the abnormal targets are thunderstorms, squall lines, bow echoes, hook echoes and hailstorms;

[0030] The target area weather change prediction module is used to predict the weather changes in the area where each abnormal target is located in the future using a long short-term memory network model algorithm.

[0031] Thirdly, the present invention provides an electronic device including a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the weather identification method based on radar echo maps according to the first aspect.

[0032] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the weather identification method based on radar echo maps described in the first aspect.

[0033] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0034] This invention provides a weather identification method, system, electronic device, and storage medium based on radar echo images. This invention achieves intelligent identification of thunderstorm intensity and abnormal weather by reading radar echo images and performing image preprocessing, image segmentation, feature extraction, and judgment matching. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of a single thunderstorm according to the present invention;

[0037] Figure 2 This is a schematic diagram of a multi-cell thunderstorm according to the present invention;

[0038] Figure 3 This is a schematic diagram of the squall line of the present invention;

[0039] Figure 4 This is a schematic diagram of the bow-shaped echo of the present invention;

[0040] Figure 5 This is a schematic diagram of the hook-shaped echo of the present invention;

[0041] Figure 6 This is a schematic diagram of a hailstorm according to the present invention;

[0042] Figure 7 This is a flowchart of the weather identification method based on radar echo maps according to the present invention;

[0043] Figure 8 This is the original radar echo image of the present invention;

[0044] Figure 9 This is the radar echo image after the color space has been converted from RGB to CIE-Lab color space according to the present invention;

[0045] Figure 10 This is a diagram showing the separation result of the radar echo map in this invention;

[0046] Figure 11 This is a structural diagram of the densely connected network of the present invention;

[0047] Figure 12 This is a structural diagram of the dense convolutional block of the present invention;

[0048] Figure 13 This is a structural diagram of the YOLONano network of the present invention;

[0049] Figure 14 This is a structural diagram of the APE-YOLO backbone network of the present invention;

[0050] Figure 15 This is a structural diagram of the APE-YOLO network model of the present invention;

[0051] Figure 16 This is a schematic diagram of the LSTM structure of the present invention;

[0052] Figure 17 This is a schematic diagram of the hidden layer cell structure of the present invention;

[0053] Figure 18 This is a structural diagram of the weather identification system based on radar echo maps according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Weather radar is a type of active microwave atmospheric remote sensing device specifically designed for atmospheric sounding. It is a crucial instrument for meteorological observation, playing a vital role in the monitoring, forecasting, and warning of sudden and severe weather events. It is one of the primary tools for detecting small-scale weather systems (such as typhoons and rainstorm cloud systems).

[0057] Weather radar can detect basic meteorological elements such as reflectivity factor, Doppler radial velocity, and spectral width, thus providing data support for short-term weather forecasting and warnings. Radar reflectivity data, in particular, is most frequently used because it is directly related to severe convective weather systems. Radar data has numerous applications in daily operational and scientific research. For example, radar data can be used in numerical model assimilation to provide a more accurate initial field for numerical models; radar short-term forecasting systems based on radar reflectivity data can forecast severe convective weather within the radar's detection range for the next two hours.

[0058] Generally, brighter areas on a radar image are more likely to experience severe convective weather, with yellow, yellow, and red areas being particularly prone to severe thunderstorms. Thunderstorms can be easily identified based on radar colors. Furthermore, the shape of the basic reflectivity can be used to classify the severity of thunderstorms.

[0059] A thunderstorm cell describes a single thunderstorm event, more precisely, those thunderstorms that occur only once in their life cycle. Most thunderstorm cells are not strong thunderstorms, but if they possess sufficient unstable energy, they can still produce brief periods of intense convective weather. These thunderstorms are called pulse thunderstorms, which, like a pulse, suddenly spike after a period of calm. Specifically... Figure 1 As shown.

[0060] Multicell thunderstorms typically consist of at least 2-4 thunderstorm cells lining up and then moving together in the same direction. They are usually formed by the merging of pulse thunderstorms (because the preceding thunderstorm was strong and hadn't dissipated before the subsequent one began to develop), and are the most common type of thunderstorm. Specific details are as follows... Figure 2 As shown.

[0061] When a thunderstorm develops into a line, multi-cell thunderstorms develop into a squall line. Squall lines typically extend for hundreds of kilometers and are characterized by sudden changes in wind direction, rapid increases in wind speed, a sharp rise in air pressure, and a sudden drop in temperature. Near squall lines, in addition to the drastic changes in wind, air pressure, and temperature, severe weather events such as lightning, torrential rain, hail, and tornadoes are often observed. Figure 3 As shown.

[0062] A bow echo is a fast-moving, forward-convective echo shaped like a bow. It is often accompanied by severe weather phenomena such as downbursts, hail, torrential rain, or tornadoes, posing a significant threat to agricultural production. Furthermore, tornadoes are more likely to occur at the end of a bow echo because the air undergoes cyclonic rotation there. Specifically, for example… Figure 4 As shown.

[0063] Hook echoes typically form due to the intense rotation of winds within a supercell, specifically caused by strong wind shear at different heights in the region, leading to wind rotation. Hook echoes symbolize the occurrence of a supercell, which is often closely related to tornadoes, and tornadoes usually occur within the hook-shaped region, as shown in the example below. Figure 5 As shown.

[0064] Because hailstones are large and solid, they reflect radar electromagnetic waves more strongly. This results in very high reflectivity values ​​on the radar echo, typically above 60 dBZ. The presence of red, pink, purple, and a central white area usually indicates hail. A reflectivity line often extends from the radar echo; this is known as TBSS (three-body scattering), indicating the presence of large hailstones. Figure 6 As shown.

[0065] Image recognition technology is data-driven and helps intelligent systems make judgments and corresponding perception processes. Its judgments in specific scenarios can approach or even surpass those of humans.

[0066] Example 1

[0067] like Figure 7 As shown in the figure, this embodiment provides a weather identification method based on radar echo maps, which includes the following steps.

[0068] Step 100: Obtain radar echo map of the target area.

[0069] Step 200: Perform image segmentation on the radar echo map of the target area to obtain multiple regions of interest.

[0070] In this embodiment, image segmentation refers to dividing an image into several non-overlapping regions based on features such as grayscale, color, spatial texture, and geometric shape. This ensures that these features exhibit consistency or similarity within the same region, while showing significant differences between different regions. Simply put, it's about separating the target from the background in an image.

[0071] When performing radar echo image recognition, image segmentation techniques are needed to separate and extract the regions of interest from the background. In this embodiment, a feature-enhanced segmentation method is used to segment the radar echo image to be recognized, obtaining the regions of interest.

[0072] Feature-enhanced segmentation methods include extracting multi-scale features or extracting features from a series of nested regions. In deep networks for image segmentation, CNNs are often applied to small squares of an image, typically referred to as fixed-size convolutional kernels centered on each pixel, which label the classification of each pixel by observing the small regions surrounding it. In the network model, the image is converted from the RGB color space to the CIE-Lab color space, and the (L, a, b) color value and (x, y) coordinates of each pixel form a 5-dimensional vector V[L, a, b, x, y]. The similarity between two pixels can be measured by their vector distance; the greater the distance, the smaller the similarity.

[0073] First, K seed points need to be generated. Then, in the surrounding space of each seed point, search for the nearest pixels and classify them into the same category as the seed point. Repeat this process until all pixels are classified. Next, calculate the average vector value of all pixels in these K superpixels to obtain K new cluster centers. Then, use these K cluster centers to search for the nearest pixels that are most similar to them. After all pixels are classified, obtain K new superpixels, update the cluster centers, and iterate again. Repeat this process until convergence, completing the segmentation model training. Inputting the binarized radar echo image into the image segmentation model will result in K superpixels, which are the separated regions of interest. The process is as follows: Figure 8 , Figure 9 and Figure 10 As shown.

[0074] In this embodiment, step 200 specifically includes:

[0075] Step 201: Input the radar echo image of the target area into the image segmentation convolutional neural network model to obtain multiple regions of interest.

[0076] The training process of the image segmentation convolutional neural network model is as follows:

[0077] (1) Construct a convolutional neural network;

[0078] (2) Construct sample data; the sample data includes multiple sample area radar echo images and multiple regions of interest corresponding to each sample area radar echo image; the multiple regions of interest corresponding to the sample area radar echo images are obtained by superpixel segmentation of the sample area radar echo images using the SLIC algorithm.

[0079] (3) Input the radar echo map of the sample area into the convolutional neural network to obtain the prediction result map. Use the cross-entropy loss function to calculate the error between the prediction result map and the real result map, and backpropagate the error to update the parameters of the convolutional neural network model, thereby obtaining the image segmentation convolutional neural network model.

[0080] The predicted result image is a series of regions of interest obtained by inputting the radar echo image of the sample region into a convolutional neural network; the actual result image is a series of regions of interest obtained by superpixel segmentation of the radar echo image of the sample region using the SLIC algorithm.

[0081] Step 300: Anomaly detection is performed on each region of interest using a target detection algorithm to obtain one or more anomaly targets and the location of each anomaly target; the anomaly targets are thunderstorms, squall lines, bow echoes, hook echoes, and hailstorms, specifically including:

[0082] The APE-YOLO network model algorithm is used to detect abnormal targets in the region of interest, resulting in one or more abnormal targets and the location region of each abnormal target.

[0083] The APE-YOLO network model algorithm is obtained by first reducing the structural depth of the YOLONano network using the ape-walking edge computing model algorithm to obtain the APE-YOLO backbone network, and then adding dense connection blocks to the APE-YOLO backbone network.

[0084] Image classification is a pattern classification problem. Its goal is to classify different images into different categories with minimal classification error. In this embodiment, a deep learning algorithm will be used to complete the radar echo image shape classification task, and the DenseNet network will be used for implementation.

[0085] like Figure 11 As shown, DenseNet, proposed by Zhuang Liu et al. from Tsinghua University, is a lightweight network architecture. As CNNs become increasingly sophisticated, a new challenge arises: as information about the input or gradual changes passes through more and more layers, significant information loss can occur when the network reaches its end (or beginning), even if accuracy improves, but at the cost of enormous computational resources. Zhuang Liu et al.'s work has shown that if deep convolutional networks have shorter connections between nearby input and output layers, they can be trained more richly, accurately, and effectively. This is a feedforward method that connects each layer to every other layer.

[0086] A traditional convolutional network with n layers has n connections (one connection between each layer and its subsequent layers), while the DenseNet network has n(n+1) / 2 direct connections. For each layer, the feature maps of all previous layers are taken as input, and its own feature map is used as the input to all subsequent layers. Downsampling layers that modify the size of feature maps when they change are an important component of deep convolutional networks. Therefore, DenseNet is divided into multiple tightly connected dense blocks, such as... Figure 12 As shown; the layers between densely connected blocks are called transition layers, and they undergo convolution and pooling.

[0087] Image classification is a fundamental task in computer vision, detecting and identifying the locations of anomalies such as hook echoes and hailstorms. Therefore, this embodiment employs a target detection algorithm to detect and identify anomalous targets such as hook echoes and hailstorms.

[0088] Object detectors are divided into two categories: two-stage detectors and one-stage detectors. Among mainstream object detection and recognition methods, deep learning models are trending towards lightweight edge computing, with the YOLO series of models being the most outstanding. The YOLO series boasts significantly faster running speeds (i.e., real-time performance) and higher accuracy than other algorithms, along with the dense connectivity of DenseNet. Therefore, the edge computing algorithms developed for "Apewalking" will fully consider the advantages of these networks, proposing the APE-YOLO network based on the YOLO series to achieve edge computing-based object detection and recognition tasks.

[0089] To meet the requirements of object detection in edge computing, Alexander et al. explored human-machine collaboration strategies and designed a very concise convolutional neural network, YOLONano. This network is extremely compact, featuring a highly customized module-level macroscopic structure and a microscopic structure specifically designed for object detection on edge computing platforms. The early network structure of YOLONano is shown below. Figure 13 As shown in (a) above, the YOLONano late-stage network structure is as follows: Figure 13 As shown in (b), the later network structure has three efficient modules: PEP module, EP module and FCA module, which together improve the parameter efficiency and computational efficiency of the model.

[0090] The PEP module is a unique residual module. Its main structure consists of: ① a 1×1 convolutional mapping layer that maps the input feature map to a lower tensor; ② a 1×1 convolutional dilation layer that dilates the channels of the feature map to a higher dimension; ③ a depth-wise convolutional layer that performs spatial convolutions on the output channels of different dilation layers using different filters; and ④ a 1×1 convolutional mapping layer that maps the output channels of the previous layer to a lower dimension. The PEP module effectively reduces model and computational complexity while maintaining the model's feature learning ability. The EP module is similar to the PEP module, but it reduces one 1×1 convolutional mapping layer. The FCA module is a lightweight fully connected attention module consisting of two fully connected layers. This allows for more effective utilization of the neural network's learning ability, i.e., learning as much important feature information as possible with a limited number of parameters. This module achieves a better balance between pruning the model architecture, reducing model complexity, and increasing the model's feature learning ability.

[0091] The high heterogeneity of these modules allows the network architecture to achieve a good balance between accuracy, model size, and computational complexity through specific design, making it very suitable for edge and mobile devices.

[0092] To better adapt to the computational resource limitations and detection performance of edge computing, the Ape-Walking edge computing model algorithm reduces the structural depth of YOLONano to obtain a new backbone network, which we name the APE-YOLO backbone network. The APE-YOLO backbone network is as follows: Figure 14 As shown.

[0093] To address the high repetition rate of convolutional network blocks in the APE-YOLO backbone network across three different scales, densely connected blocks are introduced, reducing the overall network parameter count and computational resource consumption. Simultaneously, the number of times densely connected blocks are used for downsampling in the three dimensions of the APE-YOLO backbone network is reduced, decreasing the number of network layers, while the initial residual network blocks remain unchanged, resulting in the APE-YOLO network model. Compared to the previous basic YOLO backbone network, the APE-YOLO network model halve the convolutional operations at scales 1 to 3, leaving room for optimization of the later detection network; the introduction of densely connected blocks also ensures the learning capability of the representation.

[0094] The APE-YOLO network model algorithm, a commonly used method for target detection and recognition in edge computing, is an improvement on YOLONano. It introduces improved PEP and Denseblock modules at the ends of the YOLO backbone network. Both approaches leverage the advantages of feature information fusion, reducing the number of serial channels, improving information utilization, and minimizing computational resource consumption, making them crucial for edge computing applications.

[0095] APE-YOLO is also inspired by the lightweight design of YOLONano. It adds an FCA module to the three convolutional regions of different dimensions in the detection network to learn the non-linear interdependencies between channels, focusing more on information features and thus making better use of the available network capacity. The overall network structure of the APE-YOLO network model is as follows: Figure 15 As shown.

[0096] exist Figure 15 In this model, data first passes through the APE-YOLO backbone network for feature learning. The introduction of a dense block module reduces the reduction in feature learning caused by compression of the base network. The improved PEP module inherits the characteristics of depthwise separable convolution, reducing computational resource consumption while optimizing the network model's feature representation capabilities. Finally, the addition of an FCA module enhances the network's feedforward and feedback propagation capabilities, further ensuring model accuracy. The APE-YOLO network, with its highly compatible modules and lightweight design, is particularly suitable for edge computing-based object detection and recognition scenarios.

[0097] Step 400: Employ a Long Short-Term Memory (LSTM) network model algorithm to predict future weather changes within the area where each anomalous target is located; specifically including:

[0098] A long short-term memory network model algorithm is used to predict the radar echo map and radar reflectivity of the area where each abnormal target is located in the future time period.

[0099] Based on radar echo maps and radar reflectivity for future periods, predict weather changes in the area where each anomalous target is located for future periods.

[0100] In this embodiment, a target detection algorithm is used to detect, identify, and locate the positions of abnormal targets in the radar echo image. When an abnormal target location area is found in the radar echo image, a Long Short-Term Memory (LSTM) network model algorithm is used to predict information such as radar echo image and radar reflectivity in the location area in the near future, thereby completing the abnormal weather analysis and extrapolation.

[0101] Long Short-Term Memory (LSTM) networks are a specific form of Recurrent Neural Networks (RNNs). LSTM addresses the short-term memory problem of RNNs by adding gates, enabling RNNs to effectively utilize long-range temporal information. The LSTM model adds three logical control units—Input Gate, Output Gate, and Forget Gate—to the basic RNN structure, each connected to a multiplication element. By setting the weights at the edges connecting the memory cells to other parts of the neural network, the LSTM controls the input and output of information and the state of the memory cells.

[0102] exist Figure 17 In the LSTM, InputGate controls whether information flows into the Memorycell; ForgetGate controls whether information from the previous Memorycell is accumulated into the current Memorycell; OutputGate controls whether information from the current Memorycell flows into the current hidden state ht; cell: memory unit, representing the memory of the neuron's state, enabling the LSTM unit to save, read, reset, and update long-distance historical information.

[0103] During the training of the LSTM model, the data features at time t are first input into the input layer, and the results are output after passing through the activation function. The output results, the hidden layer output at time t-1, and the information stored in the cell at time t-1 are input into the nodes of the LSTM model. Through the processing of InputGate, OutputGate, ForgetGate, and the cell, the data is output to the next hidden layer or output layer. The results of the LSTM structure nodes are output to the output layer neurons, the backpropagation error is calculated, and the weights are updated.

[0104] When performing abnormal weather analysis and extrapolation, radar echo maps and radar reflectivity of the abnormal areas are used as input parameters for the LSTM model; the data processing method is as follows:

[0105] In the process of quantitative precipitation measurement using radar, there is a statistical relationship between radar reflectivity factor and rainfall intensity: Z = AIb, where I is rainfall intensity in mm / h, and A and b are two statistical constants. The ZI relationship varies not only with location, season, and precipitation type, but also with time and space, even within the same precipitation event. The accuracy of rainfall prediction depends on the accuracy of radar echo extrapolation; that is, the evaluation of precipitation prediction models boils down to the evaluation of the accuracy of radar reflectivity factor extrapolation. In radar meteorology, the radar reflectivity factor of stratiform cloud precipitation is generally between 20 and 30 dBz, while that of convective cloud precipitation is generally not less than 40 dBz. Therefore, referring to the rainfall intensity classification scheme of Shi et al. (2017), different weights are assigned to radar reflectivity factors in different intervals, i.e., weights are allocated according to the magnitude of the radar reflectivity factor of each pixel. To eliminate precipitation imbalance and improve the forecasting ability for strong echoes, stronger echoes are given greater weight to increase their influence in the fitting process, leading to better convergence.

[0106] Pixel analysis was performed on radar echo images of abnormal areas over a recent period. The echo intensity (unit: dBz, effective range [0, 75]) of each pixel in the radar image was divided into 5 levels: Level 1 ranges from [0, 20] (no rain or drizzle); Level 2 ranges from (20, 30] (light rain); Level 3 ranges from (30, 40] (moderate rain); Level 4 ranges from (40, 50] (heavy rain, torrential rain); and Level 5 ranges from (50, 75] (extremely heavy rain). Different weights were assigned to the radar reflectivity factor for each level. Finally, the radar reflectivity information and the pixel level information of the radar echo image were combined into a multi-dimensional vector as the input parameters of the LSTM model to complete the data processing task.

[0107] As a supervised learning model algorithm in machine learning, LSTM requires the processing of historical radar echo maps and radar reflectivity information as described above. The processed historical data is then input into the network model, and the output values ​​are the historical weather type and the expected value. During model training, the LSTM network model gradually approximates the expected value. When the error between the predicted and expected values ​​is sufficiently small, the model converges, completing the LSTM model training.

[0108] Example 2

[0109] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a weather identification system based on radar echo maps is provided below.

[0110] like Figure 18 As shown, this embodiment provides a weather identification system based on radar echo maps, including:

[0111] Radar echo image acquisition module 1 is used to acquire radar echo images of the target area.

[0112] The region of interest determination module 2 is used to perform image segmentation on the radar echo map of the target area to obtain multiple regions of interest.

[0113] The abnormal target location calculation module 3 is used to perform abnormal target detection on each region of interest using a target detection algorithm to obtain one or more abnormal targets and the location of each abnormal target; the abnormal targets are thunderstorms, squall lines, bow echoes, hook echoes and hailstorms.

[0114] The target area weather change prediction module 4 is used to predict the weather changes in the area where each abnormal target is located in the future using a long short-term memory network model algorithm.

[0115] Example 3

[0116] This invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to execute a weather identification method based on radar echo maps according to Embodiment 1.

[0117] Alternatively, the aforementioned electronic device may be a server.

[0118] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements a weather identification method based on radar echo maps according to Embodiment 1.

[0119] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0120] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A weather identification method based on radar echo maps, characterized in that, include: Acquire radar echo images of the target area; Image segmentation is performed on the radar echo image of the target area to obtain multiple regions of interest; An anomalous target detection algorithm is used to detect one or more anomalous targets and their locations in each region of interest. The target detection algorithm is the APE-YOLO network model algorithm. This APE-YOLO network model algorithm is obtained by first reducing the structural depth of the YOLONano network using the ape-walking edge computing model algorithm to obtain the APE-YOLO backbone network, and then adding densely connected blocks to the APE-YOLO backbone network. Anomalous targets include thunderstorms, squall lines, bow echoes, hook echoes, and hailstorms. The APE-YOLO network model includes a backbone network and a detection network. The backbone network includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in sequence. The detection network includes a first detection network, a second detection network, and a third detection network. The ends of the third, fourth, and fifth convolutional modules are respectively connected to a first dense connection module, a second dense connection module, and a third dense connection module. The first, second, and third dense connection modules are then respectively connected to the first, second, and third detection networks. The first detection network includes a first PEP module, a fourth dense connection module, and a first FCA module connected in sequence. The second detection network includes a second PEP module, a fifth dense connection module, and a second FCA module connected in sequence. The third detection network includes a third PEP module, a sixth dense connection module, and a third FCA module connected in sequence. Using a long short-term memory network model algorithm, radar echo maps and radar reflectivity are predicted for the location area of ​​each anomalous target in the future. Based on the radar echo maps and radar reflectivity in the future, weather changes in the location area of ​​each anomalous target in the future are predicted.

2. The weather identification method based on radar echo maps according to claim 1, characterized in that, The step of segmenting the radar echo image of the target area to obtain multiple regions of interest specifically includes: A feature-enhanced segmentation method is used to segment the radar echo image of the target area to obtain multiple regions of interest.

3. The weather identification method based on radar echo maps according to claim 2, characterized in that, The segmentation method based on feature enhancement is used to segment the radar echo image of the target area to obtain multiple regions of interest, specifically including: The radar echo image of the target area is input into an image segmentation convolutional neural network model to obtain multiple regions of interest; The training process of the image segmentation convolutional neural network model is as follows: Constructing a convolutional neural network; Construct sample data; the sample data includes multiple radar echo images of sample areas and multiple regions of interest corresponding to each radar echo image of a sample area; the multiple regions of interest corresponding to the radar echo images of sample areas are obtained by superpixel segmentation of the radar echo images of sample areas using the SLIC algorithm. The radar echo map of the sample area is input into the convolutional neural network to obtain the prediction result map. The cross-entropy loss function is used to calculate the error between the prediction result map and the real result map, and the error is backpropagated to update the parameters of the convolutional neural network model, thereby obtaining the image segmentation convolutional neural network model. The predicted result image is a series of regions of interest obtained by inputting the radar echo image of the sample region into a convolutional neural network; the actual result image is a series of regions of interest obtained by superpixel segmentation of the radar echo image of the sample region using the SLIC algorithm.

4. A weather identification system based on radar echo maps, characterized in that, include: Radar echo image acquisition module, used to acquire radar echo images of the target area; The region of interest determination module is used to perform image segmentation on the radar echo image of the target area to obtain multiple regions of interest. The abnormal target location calculation module is used to perform abnormal target detection on each region of interest using a target detection algorithm to obtain one or more abnormal targets and the location of each abnormal target. The target detection algorithm is the APE-YOLO network model algorithm. The APE-YOLO network model algorithm is obtained by first using the ape-walking edge calculation model algorithm to reduce the structural depth of the YOLONano network to obtain the APE-YOLO backbone network, and then adding dense connection blocks to the APE-YOLO backbone network. The abnormal targets are thunderstorms, squall lines, bow echoes, hook echoes, and hailstorms. The APE-YOLO network model includes a backbone network and a detection network. The backbone network includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in sequence. The detection network includes a first detection network, a second detection network, and a third detection network. The ends of the third, fourth, and fifth convolutional modules are respectively connected to a first dense connection module, a second dense connection module, and a third dense connection module. The first, second, and third dense connection modules are then respectively connected to the first, second, and third detection networks. The first detection network includes a first PEP module, a fourth dense connection module, and a first FCA module connected in sequence. The second detection network includes a second PEP module, a fifth dense connection module, and a second FCA module connected in sequence. The third detection network includes a third PEP module, a sixth dense connection module, and a third FCA module connected in sequence. The target area weather change prediction module uses a long short-term memory network model algorithm to predict the radar echo map and radar reflectivity of the area where each abnormal target is located in the future time period. Based on the radar echo map and radar reflectivity of the future time period, it predicts the weather changes in the area where each abnormal target is located in the future time period.

5. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform a weather identification method based on radar echo maps according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a weather identification method based on radar echo maps as described in any one of claims 1 to 3.

Images