Construction method of real-time inversion model of sea fog based on deep convolutional neural network

Abstract

The invention discloses a method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network, comprising the steps of: S1: obtaining meteorological record information in a preset inversion sea area according to a preset record data acquisition rule, and establishing a data sample set; S2: constructing a real-time inversion model of sea fog according to a deep residual convolutional network algorithm; S3: inputting the current meteorological grid record information of the preset inversion sea area into the real-time inversion model of sea fog, and outputting the sea fog state in the current preset sea area through the real-time inversion model of sea fog, wherein the output sea fog state includes fog and no fog; S4: obtaining visibility data information of the preset inversion sea area, and forming a real-time visibility inversion grid data product corresponding to the preset inversion sea area according to a preset fusion analysis algorithm by using the obtained visibility data information of the preset inversion sea area and the output sea fog state in the current preset inversion sea area. The method can realize effective observation of sea fog with high precision and high frequency.

Description

Technical Field

[0001] The present invention relates to the field of meteorological model construction, and in particular to a method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network. Background Art

[0002] At present, compared with weather phenomena on land, direct observation data on sea fog is scarce, with only observation data from coastal meteorological stations, a few island stations, buoy stations and ships. It is impossible to achieve large-scale and long-term continuous monitoring of sea fog. Moreover, sea fog has obvious boundary characteristics along the coastline in many cases, and has the characteristics of generating and expanding at night and retreating to the sea during the day.

[0003] Therefore, the fog observed by coastal stations is less than the actual fog in the sea area. With the development and application of satellite remote sensing technology, which has the advantages of relatively fast speed, wide coverage and continuous observation, it is possible to observe the occurrence, development and disappearance of sea fog in real time. Especially with the emergence of satellite lidar, the detection accuracy of fog has been greatly improved. However, sea fog identification and detection based on passive satellite remote sensors, whether it is threshold method, texture recognition or dual-channel difference method, there are difficulties in distinguishing sea fog from low clouds; satellite-borne lidar can obtain a large number of accurate sea fog and low cloud samples, but it is limited to single-point detection of sub-satellite points and has long revisit cycles, which cannot achieve high temporal resolution and continuous detection of sea fog in a large spatial range. At present, it has become an important research direction to develop low cloud and sea fog identification algorithms based on deep learning by comprehensively utilizing passive and active remote sensing observation data. At the same time, due to the complex terrain around the port, it is easy to form small-scale weather systems with strong local characteristics such as cluster fog. Relying on existing hardware equipment, it is difficult to monitor and has a great impact on shipping.

[0004] As the most widely used and influential network model in deep learning technology, the research on deep convolutional neural network has been increasingly in-depth in recent years. Generally speaking, the deeper the DCNN layer, the better the detection effect, but the large amount of calculation and the difficulty of network training also arise. Traditional convolutional networks or fully connected networks will have more or less information loss and loss when transmitting information, and will also cause gradient vanishing or gradient explosion, making it impossible to train very deep networks. Summary of the invention

[0005] The technical problem to be solved by the embodiments of the present invention is to make up for the current lack of effective means of observing sea fog with high precision and high frequency at sea, and provide a method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network, thereby achieving the purpose of constructing high-frequency and refined sea fog recognition.

[0006] The embodiment of the present invention provides a method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network, comprising the steps of:

[0007] S1: Acquire meteorological record information in a preset inversion sea area according to preset record data acquisition rules;

[0008] The meteorological record information includes temperature record information, relative humidity record information, visibility record information, sea-air temperature difference record information and preset port area control record information in the preset inversion sea area;

[0009] The temperature record information, relative humidity record information, visibility record information, and sea temperature difference record information are interpolated according to a preset interpolation algorithm to form meteorological grid data;

[0010] Marking the preset port area control record information on the grid data of the corresponding time according to the preset corresponding time marking algorithm to form a sea fog label;

[0011] The meteorological grid data and sea fog labels are filtered according to the preset rules to form a data sample set;

[0012] The data sample set includes a training sample set and an independent test sample set;

[0013] S2: Construct a real-time inversion model of sea fog based on the deep residual convolutional network algorithm;

[0014] S3: inputting the current meteorological grid record information of the preset inversion sea area into the sea fog real-time inversion model constructed in step S2, and outputting the sea fog state in the current preset sea area through the sea fog real-time inversion model, wherein the output sea fog state includes fog and no fog;

[0015] S4: Obtain visibility data information of a preset inversion sea area, and form a real-time visibility inversion grid data product corresponding to the preset inversion sea area according to a preset fusion analysis algorithm by combining the obtained visibility data information of the preset inversion sea area and the sea fog state in the current preset inversion sea area output in step S3.

[0016] The present invention has at least the following beneficial effects:

[0017] The present invention adopts a deep residual convolutional network algorithm to construct a model. The deep residual convolutional network algorithm adds a direct channel in the network. On the basis of retaining a certain proportion of the output of the previous network layer, it allows the original input information to be directly transmitted to the subsequent layer, so as to protect the integrity of the information. The entire network only needs to learn the part of the input and output difference, simplifying the learning objectives and difficulty. Not only can it learn deeper features, but it can also solve the shortcomings of gradient diffusion and large amount of calculation in deep neural networks. The real-time inversion model of sea fog based on the deep residual convolutional network algorithm has a high hit rate and high time and space resolution for the identification of sea fog in the port area, especially the identification of small-scale fog clusters and visibility inversion products. It can also make full use of the existing observation system, and the construction and operation and maintenance costs are low. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the background technology, the drawings required for use in the embodiments of the present invention or the background technology will be described below.

[0019] Figure 1 This is the process in the embodiment of the method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network Figure 1 ;

[0020] Figure 2 This is the process in the embodiment of the method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network. Figure 2 ;

[0021] Figure 3 It is a schematic diagram of the fusion of visibility site interpolation and model inversion in the embodiment. DETAILED DESCRIPTION

[0022] The embodiments of the present invention are described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0023] Example

[0024] This embodiment provides a method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network

[0025] See also Figures 1 to 3 ;

[0026] The observation data of temperature record information, relative humidity record information, visibility record information, and sea-temperature temperature difference record information of the automatic meteorological monitoring stations and buoy stations in the preset inversion sea area (islands around the Xiazhi waterway) from January to July and November to December every year from 2016 to 2021 are sorted out, and then interpolated to form grid data.

[0027] Calculation of sea temperature difference, T st =T s -Tq .in,

[0028] T st : Sea temperature difference; T s : sea temperature; T q : Air temperature. In view of the small number of sea temperature observation points and the small variation range, the sea temperature near the station is replaced by the sea temperature measured by the nearest buoy station.

[0029] The preset interpolation algorithm is the inverse distance weighted interpolation method; the interpolation function of the visibility record information is a third-order function; the interpolation function of the relative humidity record information is a first-order function; and the interpolation function of the sea temperature difference record information is a second-order function.

[0030] The station data is interpolated to 0.01×0.01° grid data, and the inverse distance weighted interpolation method is used. The interpolation parameters of each element are shown in the following table (Table 1).

[0031] Table 1 Inverse distance weighted interpolation parameters of each element

[0032] elements visibility Sea temperature difference Relative humidity Function order Third level Second order First level Impact distance 0.15° 0.15° 0.25°

[0033] Arrange the preset port area control record information of the same period; the preset port area control record information is the port area control record caused by poor visibility, which serves as a label for the occurrence of sea fog in the Xiazhi channel at the corresponding time. And form a data sample set according to the preset screening rules. In view of the low probability of sea fog and the large amount of data, the first screening according to the preset screening rules can preliminarily screen out more accurate data. The preset screening rules are: the maximum relative humidity of the grid data within the preset inversion sea area is >95%; after screening the sample set, a data sample set is formed; according to the preset time classification algorithm, the data sample set is divided into a training sample set from 2016 to 2020, and an independent test sample set from January to July 2021.

[0034] According to the deep residual convolutional network algorithm, a real-time inversion model of sea fog is constructed;

[0035] The sea fog real-time inversion model includes a data input layer, a network model construction intermediate layer, and a result output layer;

[0036] The data input layer is used to input the meteorological grid data in the preset inversion sea area;

[0037] The network model intermediate layer is constructed to construct an intermediate layer according to a deep residual convolutional network algorithm for the input meteorological grid point data in the preset inversion sea area;

[0038] The result output layer is used to output a binary classification result of whether there is fog in the current preset inversion sea area, and the result includes fog and no fog.

[0039] Specifically: The deep convolutional neural network model structure adopts a 10-layer model (convolutional layer or fully connected layer), with 4 residual convolution modules, and the residual module adopts a 2-layer convolution layer structure with a convolution step size of 3×3. The input factors of the input layer are the simultaneous temperature record information, relative humidity record information, visibility record information, sea temperature difference record information, and relative humidity record information 3 hours ago obtained by interpolating the observation data of the surrounding stations. The spatial range of the input data is 0.4×0.4° in longitude and latitude, that is, 41×41 grid points. First, the input data is convolved to extract preliminary features, 64 convolution kernels are configured, the kernel size is 5×5, and the convolution step is set to 2; then after the Batch Normalization operation, the ReLU activation function is used; then it enters the pooling layer, using the Max pooling method with a size of 3×3 and a step size of 2; then it passes through two residual convolution modules with 64 convolution kernels; then it passes through two residual convolution modules with 128 convolution kernels through downsampling transformation; the calculation results are then passed through an average pooling layer, and finally output to the fully connected layer and the SoftMax layer. The output layer is the binary classification result of whether there is fog or not, that is, foggy and fog-free.

[0040] Before training the sea fog real-time inversion model, the training sample set is first normalized.

[0041] The normalization formula for the training sample set is:

[0042]

[0043] x min is the minimum value in the training sample set;

[0044] x max is the maximum value in the training sample set;

[0045] y is the normalized value;

[0046] x is the value before normalization.

[0047] According to the normalized y value, the model is trained and finally a real-time inversion model of sea fog is established.

[0048] The current meteorological grid point record information of the preset inversion sea area is input into the constructed sea fog real-time inversion model to determine whether the preset start-up conditions for starting the sea fog real-time inversion model are met; if the maximum relative humidity record information of the current meteorological grid point record information of the current preset inversion sea area is greater than the preset start-up humidity record information and the relative humidity is greater than 95%. If the start-up conditions are not met, it is considered that there is no fog. The sea fog state in the current preset sea area is output by the sea fog real-time inversion model, and the output sea fog state includes fog and no fog.

[0049] Obtain visibility data information of the preset inversion sea area, and form a real-time visibility inversion grid data product corresponding to the preset inversion sea area according to the preset fusion analysis algorithm by combining the obtained visibility data information of the preset inversion sea area and the output sea fog state in the current preset inversion sea area. The preset fusion analysis algorithm includes the following steps: obtaining shape data information of the current preset inversion sea area; calibrating the preset representative points in the preset inversion sea area according to the preset punctuation algorithm; obtaining the sea fog state corresponding to the representative points through the sea fog real-time inversion model; assigning corresponding visibility values ​​according to the sea fog state corresponding to the representative points; interpolating according to the visibility values ​​of the representative points to form a real-time visibility inversion grid data product. Specifically, according to the real-time observation data of the visibility stations around the Xiazhi waterway and the model recognition results, after fusion analysis, a 0.01×0.01° refined visibility real-time inversion grid data product of the sea area and its surroundings can be formed. The fusion analysis method is as follows: First, according to the long strip shape of the Xiazhi waterway, three fixed positions of the upper, middle and lower positions are artificially set as representative points. Secondly, according to the model recognition results, if there is fog, the visibility of all representative points is set to 300m; if there is no fog, the visibility interpolation data is used as the final visibility of the representative point. Finally, the surrounding visibility measured data and representative point data are interpolated together to form a grid product.

[0050] The real-time inversion model of sea fog based on deep convolutional neural network is used to construct and test the hit rate and TS score of the model recognition of sea fog in the preset inversion sea area (Xiazhi Channel), which are better than the surrounding station representative method and the direct application of visibility interpolation method. The spatial distribution accuracy of visibility of the model inversion product in 2021 is better than the direct interpolation result; for example Figure 3 Shown is the visibility station interpolation (left) and model inversion fusion (right) of the Xiazhi Channel at 10:00 on May 9, 2021. At 10:00 on May 9, 2021, navigation control was imposed in the preset inversion sea area (Xiazhi Channel) due to poor visibility. The inversion model was able to identify the sea fog map, but the visibility observation stations at the surrounding sites could not.

[0051] The present invention has the characteristics of high hit rate and high time and space resolution for sea fog recognition in port areas, especially recognition of small-scale cluster fog, and visibility inversion products; and can make full use of existing observation systems and has the advantages of low construction and operation and maintenance costs.

[0052] Those skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. When the programs are executed, they can include the processes of the above-mentioned method embodiments. The aforementioned storage media include: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.

Claims

1. A method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network, characterized in that: Includes steps: S1: Acquire meteorological record information in a preset inversion sea area according to preset record data acquisition rules; The meteorological record information includes temperature record information, relative humidity record information, visibility record information, sea-air temperature difference record information and preset port area control record information in the preset inversion sea area; The temperature record information, relative humidity record information, visibility record information, and sea temperature difference record information are interpolated according to a preset interpolation algorithm to form meteorological grid data; Marking the preset port area control record information on the grid data of the corresponding time according to the preset corresponding time marking algorithm to form a sea fog label; The meteorological grid data and sea fog labels are filtered according to the preset rules to form a data sample set; The data sample set includes a training sample set and an independent test sample set; S2: Construct a real-time inversion model of sea fog based on the deep residual convolutional network algorithm; S3: inputting the current meteorological grid record information of the preset inversion sea area into the sea fog real-time inversion model constructed in step S2, and outputting the sea fog state in the current preset sea area through the sea fog real-time inversion model, wherein the output sea fog state includes fog and no fog; S4: Obtain visibility data information of a preset inversion sea area, and form a real-time visibility inversion grid data product corresponding to the preset inversion sea area according to a preset fusion analysis algorithm by combining the obtained visibility data information of the preset inversion sea area and the sea fog state in the current preset inversion sea area output in step S3.

2. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 1, characterized in that: The preset interpolation algorithm is the inverse distance weighted interpolation method; The interpolation function of the visibility record information is a third-order function; The interpolation function of the relative humidity record information is a first-order function; The interpolation function of the sea-temperature difference record information is a second-order function.

3. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 1, characterized in that: The preset screening rule is to put the grid point data with relative humidity record information greater than the preset humidity in the preset inversion sea area into the data sample set.

4. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 3, characterized in that: According to a preset time classification algorithm, the data sample set is divided into a training sample set and an independent test sample set.

5. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 1, characterized in that: The sea fog real-time inversion model in step S2 includes a data input layer, a network model construction intermediate layer, and a result output layer; The data input layer is used to input the meteorological grid data in the preset inversion sea area; The network model intermediate layer is constructed to construct an intermediate layer according to a deep residual convolutional network algorithm for the input meteorological grid point data in the preset inversion sea area; The result output layer is used to output a binary classification result of whether there is fog in the current preset inversion sea area, and the result includes fog and no fog.

6. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 5, characterized in that: Before step S2, the training sample set is normalized.

7. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 6, characterized in that: The normalization formula for the training sample set is: x min is the minimum value in the training sample set; x max is the maximum value in the training sample set; y is the normalized value; x is the value before normalization.

8. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 6, characterized in that: Before outputting the sea fog state in the current preset sea area through the sea fog real-time inversion model in step S3, the method further includes the steps of: judging whether the preset start condition for starting the sea fog real-time inversion model is met; The preset start condition is: the maximum relative humidity record information of the current meteorological grid record information of the current preset inversion sea area is greater than the preset start humidity record information.

9. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 1, characterized in that: The preset fusion analysis algorithm in step S4 includes the following steps: Get the shape data information of the current preset inversion sea area; Calibrate the preset representative points in the preset inversion sea area according to the preset punctuation algorithm; The sea fog state corresponding to the representative point is obtained through the sea fog real-time inversion model; Assign corresponding visibility values ​​according to the sea fog state corresponding to the representative point; Interpolation is performed based on the visibility values ​​of representative points to form real-time visibility inversion grid data products.

10. The method for constructing a real-time inversion model of sea fog based on a deep convolutional neural network according to claim 1, characterized in that: The calculation formula for the sea temperature difference is: Tst=Ts-Tq Tst is the sea temperature difference; Ts is the sea temperature; Tq is the air temperature.

Images