A method for inverting rainfall intensity levels from radar images
By combining the BP neural network and the theory of wave wave number energy proportion, the measurement error and occlusion area limitations of the inversion of the rain intensity level of the ship's navigation conditions are solved, and a high accuracy and reliability of rainfall intensity level estimation is achieved.
Patent Information
- Application Number
- CN202210591637.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-27
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-05-27
AI Technical Summary
The existing inversion method of ocean surface rainfall intensity level has measurement errors and shading area limitations under ship navigation conditions, making it difficult to achieve accurate rainfall intensity estimation.
Using a BP neural network-based method, self-learning and adaptive BP neural networks are combined with the theory of wave wave number energy proportion, and the rainfall intensity level in the radar image is estimated through two-dimensional discrete Fourier transform and BP neural network training.
The high accuracy and reliability of rainfall intensity level inversion under the actual navigation conditions of the ship is achieved, the measurement error and shading area limitations of traditional methods are overcome, and the accuracy of rainfall intensity level estimation is improved.
Smart Images

Figure CN115063675B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of ocean remote sensing, and relates to a radar image rainfall intensity level inversion method, in particular to a radar image rainfall intensity level inversion method based on a BP neural network. Background Art
[0002] The ocean is closely related to human life. In order to rationally utilize marine resources, it is necessary to explore and study the ocean surface environment. Ocean surface rainfall information is an important factor in the marine environment, so rainfall detection and rainfall intensity inversion are important directions for marine environmental remote sensing. Rainfall detection mainly detects whether there is rainfall information in the radar image, while rainfall intensity inversion estimates the rainfall size in the rainfall image.
[0003] Classical methods for measuring rainfall intensity levels on the ocean surface can be divided into two categories: contact rain gauge measurement and radar remote sensing estimation. Since the rain gauge will also move with the ship when the ship is moving, it is impossible to accurately measure the rainfall intensity. Therefore, using marine radar remote sensing to estimate the rainfall intensity level is an important means of modern ocean surface rainfall information research. The rainfall intensity level estimation algorithm in the existing technology can initially realize the inversion of rainfall intensity level, but since the selection of the shielded area as the experimental area has certain limitations, a new algorithm research is needed for a widely applicable detection area.
[0004] Since ships sail in vast sea areas, there are rarely obstructed areas in radar images. Therefore, through actual radar image analysis, the wave monitoring area with better wave texture is used for rainfall detection and rainfall intensity calculation, which can effectively reduce the measurement error. Summary of the invention
[0005] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a radar image rainfall intensity level inversion method based on BP neural network, which combines the self-learning and adaptive BP neural network with the wave number energy ratio theory to realize self-learning and adaptive rainfall intensity level estimation with high reliability and accuracy.
[0006] In order to solve the above technical problems, a method for inverting rainfall intensity levels from radar images is provided in the present invention. The method comprises the following steps: offline reading radar images and rainfall amounts corresponding to the images, randomly selecting radar images containing different rainfall amounts, and determining rainfall intensity levels of the selected radar images according to rainfall intensity level standards; selecting Cartesian frame images from the selected radar images respectively, and converting the polar coordinate images into rectangular coordinate images; respectively obtaining the wave number energy proportion P of each selected radar image. 海浪 ; The wave number energy ratio P 海浪The vector o is taken as input, and the vector d composed of the rainfall intensity level of the selected radar image is o As the corresponding output result expectation, the BP neural network is trained using the vector o, and the trained BP neural network model is recorded; a Cartesian frame image is selected from the radar image to be detected, and the wave number energy proportion P of the radar image to be detected is calculated. 海浪 ; The wave number energy proportion P of the radar image to be detected 海浪 The data is input into the trained BP neural network model to estimate the rainfall intensity level of the radar image to be detected.
[0007] Furthermore, the wave number energy proportion P of the radar image 海浪 The methods for obtaining include:
[0008] The Fourier coefficient matrix corresponding to the echo intensity matrix of the Cartesian frame area selected in the radar image is obtained by two-dimensional discrete Fourier transform:
[0009]
[0010] Where: (x, y) is the coordinate of each pixel in the Cartesian frame image, F(k x ,k y ) is the Fourier coefficient of f(x,y), (k x ,k y ) is the wave number domain coordinate, k x and k y are the component values of the wave number vector, f(x,y) is the echo intensity value in the image, and N is the number of pixel points contained in the side length of the selected Cartesian box;
[0011] Use Fourier coefficients to obtain the energy spectrum:
[0012]
[0013] Find the total energy in the Cartesian box based on the energy spectrum:
[0014]
[0015] In the formula, the wave number resolution on the x-axis and y-axis are and L x , L y are the spatial scales of the x-axis and y-axis respectively; n is the total number of pixel points in the selected Cartesian frame, that is, N×N, En i is the value of the energy spectrum matrix element corresponding to each pixel point in the Cartesian frame;
[0016] The wave number range, i.e. the range of the maximum and minimum wave numbers, is:
[0017]
[0018]
[0019] Where: |k| max is the maximum value of the wave modulus length, |k| min is the minimum value of the ocean wave modulus length, λ max is the maximum wavelength of the ocean wave, λ min is the minimum value of the wave wavelength, then the range of the wave number band is |k| min ~|k| max ;
[0020] Wave number energy ratio P 海浪 The percentage of the energy of the wave number band where the wave is located in the entire wave number band is:
[0021]
[0022] In the formula: En 海浪 En is the sum of all energies in the wave number band where the wave wavelength is located. 总 is the sum of the energies in all wave number bands, i.e., the total image energy; m is the total number of pixel points in the wave number band in the selected Cartesian frame, En j It is the value of the energy spectrum matrix element corresponding to each pixel point in the wave beam segment in the Cartesian frame.
[0023] Furthermore, using vector o to train the BP neural network includes:
[0024] Step a: Preset the number of training times, learning rate and minimum error of training target of BP neural network;
[0025] Step b: vector o is used as the input of the BP neural network, and the rainfall intensity level is obtained as the output through forward propagation;
[0026] Step c: Determine whether the error between the output result and the corresponding output result expectation meets the minimum error requirement or reaches the maximum number of training times. If so, the BP neural network training is completed; otherwise, back propagation is performed to adjust the connection weights between the hidden layer and the output layer and the connection weights between the input layer and the hidden layer, and return to step b.
[0027] Furthermore, the radar is an X-band radar.
[0028] Beneficial effects of the present invention: In view of the theoretical limitations of existing methods and the impact of sea conditions on rainfall intensity level inversion, through the study of traditional rainfall intensity level inversion algorithms and neural network rainfall recognition theory, the present invention discloses a method for rainfall intensity level inversion of marine radar images based on BP neural network, which combines a self-learning and adaptive BP neural network to estimate rainfall intensity levels. It not only has a high accuracy rate, but also realizes self-learning and adaptive rainfall intensity inversion. As the data set increases, the ability to estimate rainfall intensity levels can be continuously improved. The present invention uses X-band marine radar to conduct offshore experiments. The experimental results show that the method is effective in estimating rainfall intensity levels and has high reliability and accuracy. Compared with the prior art, the advantages of the present invention are:
[0029] (1) The present invention utilizes the wave monitoring area in the marine radar image to identify rainfall. Compared with the traditional shielding area that removes the wave information, the present invention can better meet the actual navigation needs of ships.
[0030] (2) Based on the BP neural network, the present invention realizes the estimation of rainfall intensity level by training the wave number energy ratio of the marine radar image. Compared with the traditional method of detecting rainfall by the wave number energy ratio, the present method has the advantages of self-learning and self-adaptation, and can effectively improve the accuracy of rainfall intensity level estimation under the condition of sufficient training data.
[0031] (3) The present invention proposes a method for inverting rainfall intensity levels from marine radar images based on a BP neural network, and determines a BP neural network model by statistically analyzing a large amount of measured marine radar data and combining it with rainfall data measured by a rain gauge, thereby enhancing the engineering practicability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a rain-free X-band marine radar image;
[0033] Figure 2 It is an X-band marine radar image with less rainfall;
[0034] Figure 3 It is an X-band marine radar image with heavy rainfall;
[0035] Figure 4 It is a Cartesian box diagram of the wave monitoring area of the marine radar image;
[0036] Figure 5 is the wavenumber energy spectrum within the Cartesian frame of the rain-free X-band marine radar image;
[0037] Figure 6 is the wavenumber energy spectrum within the Cartesian frame of the light rain X-band marine radar image;
[0038] Figure 7 It is the wave number energy spectrum within the Cartesian frame of the X-band marine radar image for moderate to heavy rain;
[0039] Figure 8 It is the training result of BP neural network rainfall intensity level estimation;
[0040] Fig. 9 It is a flow chart of an implementation method of the present invention. DETAILED DESCRIPTION
[0041] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0042] Combination Fig. 9 , the present invention comprises the following steps:
[0043] Step 1, offline BP neural network training:
[0044] The original file data of X-band marine radar images and the generation time of the images were read, and the statistical results of the rain gauge were recorded according to the image time to determine the rainfall intensity level at that time. A suitable Cartesian frame was selected in the radar image for experimental research. The echo intensity matrix in the selected Cartesian frame was processed using a two-dimensional discrete Fourier transform and the wave number energy ratio of the waves was obtained. A model was established between the wave number energy ratio of the image and the rainfall intensity level through a BP neural network.
[0045] Step 1 includes the following steps:
[0046] Step 1.1, conduct observation experiments offline, complete the continuous recording of marine radar images, and synchronously record the corresponding rainfall. According to the actual rainfall situation, randomly select multiple groups of marine radar image data containing different rainfall intensity levels, and record the radar image acquisition time and the number of image sampling points. Determine the rainfall intensity level according to the statistical results of the rain gauge corresponding to the image acquisition time. According to the parameters such as the angle and radial distance of the radar image area with clear wave texture and good waveform, select the radar Cartesian frame image for calculating the rainfall intensity, and use the selected interpolation algorithm to convert the polar coordinate image into a rectangular coordinate image.
[0047] Step 1.2, for each offline image data, obtain the corresponding wave number energy ratio.
[0048] The Fourier coefficient matrix corresponding to the echo intensity matrix of the Cartesian frame area selected in the radar image is obtained through two-dimensional discrete Fourier transform. The formula is as follows:
[0049]
[0050] Where: (x, y) is the coordinate of each pixel in the Cartesian frame image, F(k x ,k y) is the Fourier coefficient of f(x,y), (k x ,k y ) is the wave number domain coordinate, k x and k y are the component values of the wave number vector, f(x,y) is the echo intensity value in the image, and N is the number of pixel points contained in the side length of the selected Cartesian box.
[0051] The formula for obtaining the energy spectrum using Fourier coefficients is as follows:
[0052]
[0053] The formula for obtaining energy based on the energy spectrum is as follows:
[0054]
[0055] Where: The resolution of wave number on x-axis and y-axis are: L x , L y are the spatial scales of the x-axis and y-axis respectively. Since the energy obtained by two-dimensional discrete Fourier transform is a discrete matrix, the integration process of the energy spectrum is the summation process of the energy spectrum matrix. n is the total number of pixel points in the selected Cartesian frame, that is, N×N. En i It is the value of the energy spectrum matrix element corresponding to each pixel point in the Cartesian box.
[0056] The wave number range is the range of the maximum and minimum wave numbers. The formula is as follows:
[0057]
[0058]
[0059] Where: |k| max is the maximum value of the wave modulus length, |k| min is the minimum value of the ocean wave modulus length, λ max is the maximum wavelength of the ocean wave, λ min is the minimum wavelength of the ocean wave. Then the range of the ocean wave number band is |k| min ~|k| max .
[0060] The wave number energy ratio of the wave is the percentage of the energy of the wave number band where the wave is located to the energy of the entire wave number band, that is, P 海浪 The formula is as follows:
[0061]
[0062] In the formula: En 海浪 It is the sum of all energies in the wave number band where the wave wavelength is located. 总En is the total energy of all wave number bands in the selected Cartesian frame, that is, the total image energy. m is the total number of pixel points in the wave number band in the selected Cartesian frame. j It is the value of the energy spectrum matrix element corresponding to each pixel point within the wave number band.
[0063] Step 1.3: All offline image data wave number energy proportion P 海浪 The BP neural network training is performed with rainfall information summary, and the number of training times, learning rate and training target minimum error are preset. 海浪 As the input of the BP neural network, the rainfall intensity level is obtained as the output through forward propagation. If the output result does not meet the preset requirements, back propagation is performed to adjust the weights of each layer. After the preset conditions are met, the trained BP neural network model is recorded.
[0064] The specific steps are as follows:
[0065] Step 1.3.1, construct BP neural network. 海浪 The vector composed of is taken as input o, and the vector composed of rainfall intensity level is d o As the corresponding output result expectation. The BP neural network forward propagation layer-by-layer output process is as follows:
[0066] net=w T o+b (7)
[0067] Where: w is the weight vector of each layer, b is the threshold, and o is the neural network input.
[0068] The activation function is as follows:
[0069]
[0070] The error vector function is as follows:
[0071] error=d o -y o
[0072] Where: d o is the expected output, y o is the output layer output vector.
[0073] The global error function is as follows:
[0074]
[0075] Where: d o is the expected output, y o is the output layer output vector.
[0076] If the training result reaches the target accuracy, the training is stopped. Otherwise, the model is corrected by backpropagation. The error between the actual output and the expected output is calculated layer by layer and the weight is adjusted. The dlogit function of backpropagation is as follows:
[0077]
[0078] First, calculate the output layer error. You need to calculate the differential and gradient of the reverse output layer to update the weights of this layer. The differential result is as follows:
[0079]
[0080] The gradient formula is as follows:
[0081] δ o =f error (d o -y o )×f2′(y i ) (13)
[0082] Where: w ho The hidden layer is connected to the output layer with weights, and then the hidden layer error is calculated. The differential and gradient of the hidden layer are also calculated to update the weights of this layer.
[0083]
[0084] δ h =(δ o w ih )*f1′(h i ) (15)
[0085] Where: w ih is the connection weight between the input layer and the hidden layer. After the above correction process, the connection weight is reset to update the hidden layer and the output layer.
[0086]
[0087]
[0088] Step 1.3.2, after training is completed, record the weight w from the input layer to the hidden layer of the BP neural network ih , the weight w from the hidden layer to the output layer ho , where the hidden layer threshold is denoted as b1 and the output layer threshold is denoted as b2.
[0089] Step 2, obtaining the energy ratio of the wave number of the radar image to be detected;
[0090] Read the marine radar image data to be detected, record the radar image acquisition time and the number of image sampling points. Determine the Cartesian area of the radar image used for calculation based on the angle and radial distance of the area with clear wave texture and good waveform in the radar image. Calculate the wave number energy proportion P of the selected area of the marine radar image to be detected 海浪 The specific method and formula are the same as step 1.2.
[0091] Step 3, estimating the rainfall intensity level of the marine radar image to be detected: inputting the wave number energy ratio of the marine radar image to be detected into the trained BP neural network model, and determining the rainfall intensity level in the image according to the output of the neural network.
[0092] An embodiment is given below in combination with specific parameters.
[0093] The marine radar used in the implementation of the present invention is an X-band marine radar, which works in short pulse mode with a pulse repetition frequency of 1300Hz. After the echo data is digitized, it is stored in polar coordinates by line. The time interval between two adjacent storage lines is less than 1ms. The radar antenna scans a circle for about 2.5s. The total number of lines in a radar image is about 3300, and there are 600 pixel points on each line. Its azimuth resolution is about 0.1° and its radial resolution is about 7.5m. The original marine radar image used in the experiment mainly comes from the observation data of the ocean observation station on Haitan Island, Pingtan County, Fujian Province in November 2013. The rain gauge used in the experiment is installed in the area near the radar, and the accumulated rainfall per minute is recorded. The measurement accuracy of the rain gauge is 0.1mm. If the accumulated rainfall within 1 minute does not reach 0.1mm, it is recorded as 0, that is, no rain.
[0094] Combined with Figures 1 to 9 , the present invention comprises the following steps:
[0095] The first step is offline BP neural network training. It includes the following steps:
[0096] Step 1.1, conduct observation experiments offline, complete the continuous recording of marine radar images, and synchronously record the corresponding rainfall. According to the actual rainfall situation, randomly select multiple groups of marine radar images containing different rainfall intensity levels, record the acquisition time of the image file, and then find the measurement data of the rain gauge in the corresponding period according to the acquisition time of the file. The classification of rainfall type and rainfall intensity in this example is shown in Table 1.
[0097] Table 1 Rainfall intensity and rainfall intensity classification standards
[0098]
[0099] The rainfall type corresponds to different rainfall intensities, so different rainfall intensity levels are set, referred to as rain intensity levels. Since the measurement accuracy of the rain gauge is 0.1mm, if the accumulated rainfall within 1 minute does not reach 0.1mm, it is recorded as no rain. The difference between light rain and no rain is small, so it is classified as almost no rain type and set to rainfall intensity level 0. Similarly, the light rain type is set to rainfall intensity level 2, and the rainfall intensity level of medium to heavy rain is set to rainfall intensity level 3. There is less data on extreme rainstorms, and this example cannot be trained well, so the rainfall intensity level is not estimated in this method. The rainfall intensity level is divided according to the specific situation as shown in Table 2.
[0100] Table 2 Classification of rainfall intensity in this example
[0101]
[0102] In the attached figure Figure 1 This is the rain-free X-band marine radar image. Figure 2 This is an X-band marine radar image with less rainfall. Figure 3 This is an X-band marine radar image with heavy rainfall. With the appearance of rainfall information, the overall echo intensity of the marine radar image increases. Observe the area with clear wave texture and good waveform in the marine radar image. In this example, a Cartesian frame area with an azimuth of 135 degrees, a radial distance of 600m to 1600m, and a size of 960m×960m is selected as a fixed wave monitoring area for experiment. The obtained polar coordinate image is converted into a rectangular coordinate image by near point interpolation replacement. The schematic diagram of selecting a Cartesian frame in this area is shown as follows Figure 4 As shown, the fan-shaped area is a schematic diagram of the selected wave monitoring area, and the rectangular area is a schematic diagram of the Cartesian frame. The point in the fan-shaped area closest to each point in the Cartesian frame is found, and the points in the Cartesian frame are interpolated one by one. After interpolation, the Cartesian frame echo intensity matrix can be obtained.
[0103] Step 1.2, calculate the wave number energy ratio for each offline image data. Use two-dimensional discrete Fourier transform to calculate the Fourier coefficient matrix corresponding to the echo intensity matrix of the Cartesian frame area selected in the radar image. The formula is as follows:
[0104]
[0105] Where: (x, y) is the coordinates of each pixel in the Cartesian frame area image of the selected area, F(k x ,k y ) is the Fourier coefficient of f(x,y), (k x ,k y ) is the wave number domain coordinate, k x and k yare the component values of the wave number vector, f(x,y) is the echo intensity value in the image, and N is the number of pixel points contained in the side length of the selected Cartesian box, which is 128.
[0106] The formula for obtaining the energy spectrum using Fourier coefficients is as follows:
[0107]
[0108] The formula for obtaining energy based on the energy spectrum is as follows:
[0109]
[0110] Where: The resolution of wave number on x-axis and y-axis are: L x , L y are the spatial scales of the x-axis and y-axis respectively, in this case 128×128 pixels, that is, 960m×960m. Since the energy obtained by two-dimensional discrete Fourier transform is a discrete matrix, the integration process of the energy spectrum is the summation process of the energy spectrum matrix. n is the total number of pixels in the selected Cartesian frame, that is, 128×128. En i It is the value of the energy spectrum matrix element corresponding to each pixel point in the Cartesian box.
[0111] Figure 5 is the wavenumber energy spectrum within the Cartesian frame of the rain-free X-band marine radar image, Figure 6 is the wave number energy spectrum in the Cartesian frame of the X-band marine radar image of Xiaoyu, Figure 7 It is the wave number energy spectrum in the Cartesian frame of the X-band marine radar image of moderate to heavy rain. As the rainfall intensity increases, the points of the wave number energy spectrum gradually move toward the origin. The formula for the maximum and minimum values of the wave number is as follows:
[0112]
[0113]
[0114] In the formula: In the formula: |k| max is the maximum value of the wave modulus length, |k| min is the minimum value of the ocean wave modulus length, λ max is the maximum wavelength of the ocean wave, λ min is the minimum wavelength of the ocean wave. Then the range of the ocean wave number band is |k| min ~|k| max According to the traditional wave number energy method theory, combined with the known length of the wave wavelength and the radar image resolution, this example selects the minimum wavelength of the wave as 52.5m and the maximum wavelength as 200m, so the wave number range of the wave is between 0.0314rad / m and 0.1197rad / m.
[0115] The formula for the energy ratio of wave number is as follows:
[0116]
[0117] In the formula: En 海浪 It is the sum of all energies in the wave number band where the wave wavelength is located. 总 is the sum of the energy in all wave number bands, that is, the total energy of the image. m is the total number of pixel points in the wave number band of the selected Cartesian frame. j It is the value of the energy spectrum matrix element corresponding to each pixel point in this range.
[0118] Step 1.3: To prevent inaccurate fitting due to different dimensions, the wave number energy ratio P 海浪 Combined with the actual rainfall intensity level into a matrix for normalization. Then construct a neural network for training, preset the number of training times, learning rate and minimum error of the training target. The input of the BP neural network is the normalized wave number energy ratio vector o. The result of the forward propagation output is the normalized rainfall intensity level estimated by the training model. The estimated result is denormalized. Considering that the estimated result may not be an integer, the training estimation result is rounded to determine the rainfall intensity level. If the output result does not meet the preset requirements, back propagation is performed to adjust the weights of each layer. Record the BP neural network model after the preset conditions are met.
[0119] The specific steps are as follows:
[0120] Step 1.3.1, construct BP neural network. 海浪 The vector composed of input o and rainfall intensity level is d o As the corresponding output result expectation. The BP neural network forward propagation layer-by-layer output process is as follows:
[0121] net=w T o+b (24)
[0122] Where: w is the weight vector of each layer, b is the threshold, and o is the neural network input.
[0123] The activation function is as follows:
[0124]
[0125] The error vector function is as follows:
[0126] error=d o -y o
[0127] Where: d o is the expected output, yo is the output layer output vector.
[0128] The global error function is as follows:
[0129]
[0130] Where: d o is the expected output, y o is the output layer output vector.
[0131] If the training result reaches the target accuracy, the training is stopped. Otherwise, the model is corrected by backpropagation. The error between the actual output and the expected output is calculated layer by layer and the weight is adjusted. The dlogit function of backpropagation is as follows:
[0132]
[0133] First, calculate the output layer error. You need to calculate the differential and gradient of the reverse output layer to update the weights of this layer. The differential result is as follows:
[0134]
[0135] The gradient formula is as follows:
[0136] δ o =f error (d o -y o )×f2′(y i ) (30)
[0137] Where: w ho The hidden layer is connected to the output layer with weights, and then the hidden layer error is calculated. The differential and gradient of the hidden layer are also calculated to update the weights of this layer.
[0138]
[0139] δ h =(δ o w ih )*f1′(h i ) (32)
[0140] Where: w ih is the connection weight between the input layer and the hidden layer. After the above correction process, the connection weight is reset to update the hidden layer and the output layer.
[0141]
[0142]
[0143] In this example, the BP neural network is configured to be trained 1000 times, the learning rate is 0.01, and the minimum error of the training target is 0.1. The rainfall intensity level estimation training results and the actual rainfall intensity level are as follows Figure 8 shown.
[0144] Step 1.3.2, after training, record the weights w from the BP neural network input layer to the hidden layer ih , the weight w from the hidden layer to the output layer ho , where the hidden layer threshold is b1 and the output layer threshold is b2. The training results of this example are shown in the following formula.
[0145]
[0146]
[0147] w ho =[0.4302 -0.0171 0.1106 -0.0479 0.7192 -0.2462 -0.1239 -0.2213](37)
[0148] b2=-0.2508 (38)
[0149] The second step is to obtain the energy ratio of the wave number of the radar image to be detected.
[0150] Read a piece of marine radar image data to be detected and record the radar image acquisition time. Observe the area in the marine radar image where the wave texture is clear and the waveform is good. In this example, a Cartesian frame area with an azimuth of 135 degrees, a radial distance of 600m to 1600m, and a size of 960m×960m is selected as a fixed wave monitoring area for the experiment. Calculate the wave number energy ratio of the selected area of the marine radar image to be detected. The specific method and formula are the same as step 1.2. In this example, the wave number energy ratio P of the marine radar image to be detected is calculated. 海浪 It is 0.7092.
[0151] The third step is to estimate the rainfall intensity level of the marine radar image to be detected.
[0152] The obtained wave number energy ratio P 海浪 Input to the trained BP neural network model, the weight w1 from the input layer to the hidden layer, the weight w2 from the hidden layer to the output layer, the hidden layer threshold b1, and the output layer threshold b2 of the model are set according to the results obtained in step 1.3. The rainfall intensity level in the image is determined based on the output of the neural network. In this example, it is estimated that the rainfall intensity level corresponding to the marine radar image to be detected is 0, that is, there is almost no rain at this time. In this example, the actual rain gauge data corresponding to the image acquisition time shows that there is no rain at this time, so the BP neural network rainfall intensity level estimation result is correct.
[0153] To avoid contingency, multiple original radar images with different statistical data were selected and experimental verification was carried out according to the above steps. For the trained BP neural network model, the test set was selected as 100 sets of marine radar image data, in which the distribution of rainfall intensity levels is as follows Figure 8 The specific statistical results of rainfall intensity levels are shown in Table 3.
[0154] Table 3 Statistical results of rainfall intensity levels
[0155]
[0156] Set the test parameters, select the test set, and test the training results. Calculate the mean square error, root mean square error, and accuracy of the test set output.
[0157]
[0158] Where: MSE is the mean square error, d o is the true value, y o is the estimated value and n is the sample size.
[0159]
[0160] Where: RMSE is the root mean square error, d o is the true value, y o is the estimated value and n is the sample size.
[0161]
[0162] Where: precision is the accuracy, n is the dimension of the test set input vector, n error is the number of estimation results with large errors.
[0163] The estimation results of the BP neural network are rounded off, and the experimental performance is as follows: the mean square error (MSE) is 0.39709, the root mean square error (RMSE) is 0.63015, and the accuracy is 0.83.
[0164] The experimental results show that the accuracy of the rainfall intensity level inversion of marine radar images based on BP neural network is high, which meets the requirements of identifying rainfall intensity level inversion. Since BP neural network has the ability of self-learning and self-adaptation, the accuracy will gradually increase with the increase of the size of the data set. Compared with the traditional neural network rainfall recognition method, this method can further estimate the rainfall intensity level and has more practical significance.
[0165] The BP neural network-based rainfall intensity level inversion method for marine radar images proposed in the present invention has a high rainfall intensity level estimation accuracy. This method overcomes the defects of the traditional wave number energy ratio algorithm that cannot be self-learned and adaptive, and can gradually improve the accuracy as the data set increases, thereby realizing remote sensing measurement of rainfall intensity levels.
Claims
1. A method for inverting rainfall intensity levels from radar images, characterized by: Offline reading of radar images and the rainfall corresponding to the images, and randomly selecting radar images containing different rainfall amounts, and determining the rainfall intensity level of the selected radar images according to the rainfall intensity level standard; Selecting Cartesian frame images respectively from the selected radar images, and converting the polar coordinate images into rectangular coordinate images; Calculate the wave number energy proportion P of each selected radar image 海浪 ; The Fourier coefficient matrix corresponding to the echo intensity matrix of the Cartesian frame area selected in the radar image is obtained by two-dimensional discrete Fourier transform: Where: (x, y) is the coordinate of each pixel in the Cartesian frame image, F(k x ,k y ) is the Fourier coefficient of f(x,y), (k x ,k y ) is the wave number domain coordinate, k x and k y are the component values of the wave number vector, f(x,y) is the echo intensity value in the image, and N is the number of pixel points contained in the side length of the selected Cartesian box; Use Fourier coefficients to obtain the energy spectrum: Find the total energy in the Cartesian box based on the energy spectrum: In the formula, the wave number resolution on the x-axis and y-axis are and L x , L y are the spatial scales of the x-axis and y-axis respectively; n is the total number of pixel points in the selected Cartesian frame, that is, N×N, En i is the value of the energy spectrum matrix element corresponding to each pixel point in the Cartesian frame; The wave number range, i.e. the range of the maximum and minimum wave numbers, is: Where: |k| max is the maximum value of the wave modulus length, |k| min is the minimum value of the ocean wave modulus length, λ max is the maximum wavelength of the ocean wave, λ min is the minimum value of the wave wavelength, then the range of the wave number band is |k| min ~|k| max ; Wave number energy ratio P 海浪 The percentage of the energy of the wave number band where the wave is located in the entire wave number band is: In the formula: En 海浪 En is the sum of all energies in the wave number band where the wave wavelength is located. 总 is the sum of the energies in all wave number bands, i.e., the total image energy; m is the total number of pixel points in the wave number band in the selected Cartesian frame, En j is the value of the energy spectrum matrix element corresponding to each pixel point in the wave beam segment in the Cartesian frame; The wave number energy ratio P 海浪 The vector o is taken as input, and the vector d composed of the rainfall intensity level of the selected radar image is o As the corresponding output result expectation, the BP neural network is trained using the vector o, and the trained BP neural network model is recorded; Select a Cartesian frame image from the radar image to be detected, and calculate the wave number energy proportion P of the radar image to be detected. 海浪 ; The wave number energy proportion P of the radar image to be detected 海浪 The data is input into the trained BP neural network model to estimate the rainfall intensity level of the radar image to be detected.
2. The method for inverting rainfall intensity levels from radar images according to claim 1, characterized in that: The method of training the BP neural network using the vector o comprises: Step a: Preset the number of training times, learning rate and minimum error of training target of BP neural network; Step b: vector o is used as the input of the BP neural network, and the rainfall intensity level is obtained as the output through forward propagation; Step c: Determine whether the error between the output result and the corresponding output result expectation meets the minimum error requirement or reaches the maximum number of training times. If so, the BP neural network training is completed; otherwise, back propagation is performed to adjust the connection weights between the hidden layer and the output layer and the connection weights between the input layer and the hidden layer, and return to step b.
3. The method for inverting rainfall intensity levels from radar images according to any one of claims 1 to 2, characterized in that: The radar is an X-band radar.
