A Jupiter dawn storm event detection method and system based on machine learning

Abstract

The invention discloses a Jupiter dawn storm event detection method and system based on machine learning, the method comprising: obtaining Jupiter aurora images, inputting a pre-established and trained detection model, and outputting a dawn storm region aurora arc area image extraction result and dawn storm event start and end time information; the detection model comprises an aurora arc segmentation module based on a UNet network and an aurora arc morphological parameter calculation module; wherein the aurora arc segmentation module is used to extract a dawn storm aurora arc mask image from the Jupiter aurora image; the aurora arc morphological parameter calculation module is used to calculate the aurora arc area and perimeter ratio according to the dawn storm aurora arc mask image, extract the change of aurora arc morphological parameters in the window based on a time sliding window method, and output the dawn storm aurora arc event.

Description

Technical Field

[0001] The present invention relates to the application research of machine learning technology in the field of space physics, and specifically to a Jupiter dawn storm event detection method and system based on machine learning. Background Art

[0002] Dawn Storm is one of the brightest events in Jupiter's auroras. As a rare bright aurora that appears near dawn local time, Dawn Storm is manifested as an enhancement in the brightness of the dawn side arc on the main auroral belt and an increase in the width of the arc in the latitudinal direction. These enhancements extend toward the poles and eastward within tens of minutes, then appear to be fixed a few hours before dawn, and finally return to the typical aurora intensity. Dawn storms may be related to the reconnection process of the magnetosphere, but the precise mechanism of their generation and evolution is currently unclear. The study of dawn storm events has important scientific value.

[0003] At present, the Hubble Telescope and Juno satellite on Earth have carried out a large number of Jupiter auroras. With the further implementation of Jupiter exploration programs at home and abroad, the accumulated aurora observation data volume is getting larger and larger. The early method of identifying dawn storm events by human eyes is inefficient. The existing method of automatically detecting aurora dawn storm events requires solving the center of gravity of the auroral ellipse after polar coordinate projection, and calculating the change in the width of the auroral arc on the main emission band perpendicular to the auroral ellipse at each scanning angle in turn. This method also has the following limitations. Since the auroral ellipse is not circular, the scanning axis is inevitably not always perpendicular to the ellipse, which will introduce deviations in the width calculation, and the definition of the arc width depends on the experience of the researcher. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the prior art and propose a Jupiter dawnstorm event detection method and system based on machine learning.

[0005] In order to achieve the above object, the present invention proposes a Jupiter dawn storm event detection method based on machine learning, comprising:

[0006] Obtain Jupiter's aurora image, input the pre-established and trained detection model, and output the image extraction results of the aurora arc area in the dawn storm area and the start and end time information of the dawn storm event;

[0007] The detection model includes an auroral arc segmentation module and an auroral arc morphological parameter calculation module based on the UNet network; wherein,

[0008] The aurora arc segmentation module is used to extract the dawn storm aurora arc mask image from the Jupiter aurora image;

[0009] The auroral arc morphological parameter calculation module is used to calculate the ratio of the auroral arc area and perimeter according to the dawn storm auroral arc mask image, extract the changes of the auroral arc morphological parameters in the window based on the time sliding window method, and output the dawn storm auroral arc event.

[0010] Preferably, the method comprises a training step of an auroral arc segmentation module, comprising:

[0011] Step S1) collecting Jupiter aurora imaging observation data, screening, marking dawn storm aurora arcs, and making a training set;

[0012] Step S2) The training set data is sequentially input into the UNet network, the network parameters are adjusted, and the Dice coefficient is used as an evaluation index to obtain a trained auroral arc segmentation module.

[0013] Preferably, the step S1) comprises:

[0014] Data collection: Obtain the Jupiter aurora imaging observation data accumulated by the Hubble Telescope in the far ultraviolet observation band over the years from the official website of Aurora Planet Imaging and Spectroscopy, select the observation data after projection of the northern hemisphere polar region, and select the set time span, set time resolution and image pixels;

[0015] Image screening: Screen and obtain dawn storm event samples;

[0016] Auroral arc contour annotation: Use polygons to annotate the auroral arc boundary, annotate the pixels in the auroral arc area of ​​each image as 1, and annotate the pixels in the background area as 0, obtain the corresponding annotation results of the image, and save them in json format as the label of the image, and expand them by data enhancement;

[0017] Data set division: Divide the data into training set, validation set and test set according to a certain ratio.

[0018] Preferably, the calculation of the Dice coefficient in step S2) satisfies the following formula:

[0019]

[0020] Among them, A gt and A pre are the baseline and auroral arc regions in the image, respectively. |*| is the number of pixels in the region. The Dice coefficient is in the range of [0,1]. The larger the value, the better the model segmentation effect.

[0021] Preferably, the auroral arc morphology parameter calculation module includes an auroral arc morphology parameter calculation unit, and the processing process includes:

[0022] Get the coordinates of each pixel from the auroral arc mask image. For pixel a, the coordinates are mask[i][j], where mask represents the mask image, i and j represent the pixel row and column numbers respectively. If the pixel values ​​of the four points adjacent to point a, mask[i-1][j], mask[i+1][j], mask[i][j-1], and mask[i][j+1] are all 1, then point a is not a boundary pixel of the mask. Otherwise, point a is a boundary pixel of the mask.

[0023] After counting the number of pixels with a pixel size of 1 and calculating the mask area and perimeter, the ratio F of the mask area to the perimeter is used as the key indicator of the characteristic change of the auroral arc width during the dawn storm.

[0024] Preferably, the auroral arc morphological parameter calculation module includes an auroral arc width variation characteristic calculation unit, and the processing process includes:

[0025] Step T1) setting the size of the sliding time window to L, the sliding step size to S, and the aurora arc width change threshold to T;

[0026] Step T2) Count the average value F1 of all ratios F within the first L / 2 time period;

[0027] Step T3) Count the average value F2 of all ratios F within the L / 2 time period;

[0028] Step T4) Determine the size of F2 / F1 and T. If F2 / F1>T, there is a dawn storm event in the current window, otherwise there is no dawn storm event;

[0029] Step T5) After calculating the dawn storm events in all sliding windows, the times in different sliding windows are deduplicated and merged, and the start and end times of the dawn storm events in a continuous period of time are output.

[0030] Preferably, the step T1) further includes a verification and adjustment processing step, including:

[0031] According to F, draw a trend graph, find the place where F increases, count its time range and F size, and preliminarily determine the size of the sliding time window L, the sliding step S and the initial value of the auroral arc width change threshold T;

[0032] Adjust L, S and T separately, evaluate the detection effect, and get a set of values ​​with the best overall effect.

[0033] In another aspect, the present invention provides a Jupiter dawn storm event detection system based on machine learning, comprising:

[0034] The detection output module is used to obtain Jupiter aurora images, input the pre-established and trained detection model, and output the image extraction results of the aurora arc area in the dawn storm area and the start and end time information of the dawn storm event;

[0035] The detection model includes an auroral arc segmentation module and an auroral arc morphological parameter calculation module based on the UNet network; wherein,

[0036] The aurora arc segmentation module is used to extract the dawn storm aurora arc mask image from the Jupiter aurora image;

[0037] The auroral arc morphological parameter calculation module is used to calculate the ratio of the auroral arc area and perimeter according to the dawn storm auroral arc mask image, extract the changes of the auroral arc morphological parameters in the window based on the time sliding window method, and output the dawn storm auroral arc event.

[0038] Compared with the prior art, the advantages of the present invention are:

[0039] 1. The present invention introduces machine learning technology for the first time to conduct information mining on massive Jupiter aurora observation data, realize the semantic reasoning process from bottom to top, identify the semantic category of each area block in the image, and finally obtain a segmented image with pixel-by-pixel semantic annotation, extracting the area blocks that may contain the semantics of the dawn storm event.

[0040] 2. The aurora arc is not a strict ellipse, and the measurement of its arc width is prone to errors. This paper obtains the mask area and perimeter through pixel-level semantic segmentation, and proposes a variation feature of the arc width based on the area / perimeter ratio, avoiding the error introduced by directly solving the arc width. At the same time, the time-varying feature of the area / perimeter ratio is calculated based on the time sliding window, and the aurora dawn storm events are screened accordingly, with good detection performance.

[0041] 3. The model is highly robust. After the model is trained, there is no need to adjust any parameters. Just input the image into the model, and it can directly output the corresponding dawn storm auroral arc area image extraction results and the start and end time information of the dawn storm event. Therefore, it is suitable for automatic analysis and processing of a large number of Jupiter aurora observation images. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of a Jupiter dawn storm event detection method based on machine learning of the present invention;

[0043] Figure 2 It is the evaluation effect on the validation set during the training process of the aurora arc segmentation model. DETAILED DESCRIPTION

[0044] In order to achieve fast and accurate automatic detection of auroral dawn storm events in massive Jupiter aurora observation images, a Jupiter aurora dawn storm event detection method based on machine learning was constructed to automatically, quickly and accurately identify events from long-term series of observation data, extract the start and end time of the auroral arc widening, and meet the statistical analysis needs of Jupiter aurora researchers for long-term and large sample sizes.

[0045] The present invention uses a large amount of aurora observation image data accumulated by the Hubble Telescope to construct a Jupiter aurora dawn-storm automatic detection method based on machine learning, which can accurately segment the aurora arc area in the northern hemisphere of Jupiter at the pixel level, accurately calculate the morphological change characteristics, and support researchers to flexibly adjust the event detection sensitivity. The Jupiter aurora dawn storm event detection method based on machine learning saves time and effort, can complete the workload that is difficult to complete by manual identification in a short time, and is helpful for conducting big data statistical analysis; on the other hand, the automatic identification results are objective, which can support some existing conclusions or provide new ideas and insights for Jupiter aurora research.

[0046] This technical solution first builds a machine recognition model for the Jupiter aurora image dawn storm aurora arc based on the UNet network, then calculates the ratio of the aurora arc area and circumference, calculates the change characteristics of the aurora arc width through a time sliding window, and outputs the dawn storm event information with increased aurora arc width.

[0047] The main steps include the following:

[0048] 1. Data collection: Select the polar projection data of the Hubble telescope Jupiter far ultraviolet auroral imaging observation data http: / / apis.obspm.fr / .

[0049] 2. Dataset creation: Based on continuous images, samples containing auroral dawn storm events in the northern hemisphere of Jupiter were manually screened, and the outline of the auroral arc on the dawn side was annotated using the labelme open source image annotation tool to construct the annotated dataset required for training the auroral arc morphology extraction model;

[0050] 3. Auroral arc segmentation model selection and testing: Verify the applicability of the model on auroral imaging images, and select a network model structure with better performance for dawn-side auroral arc segmentation based on the characteristics of auroral imaging images and dawn storm auroral arc imaging.

[0051] 4. Evaluation of the auroral arc segmentation model: The sample data of the labeled dataset is used as the input of the selected dawn storm auroral arc morphology extraction model, the parameters are adjusted to train the recognition model, and the model recognition results are obtained and evaluated.

[0052] 5. Calculation of auroral arc morphological parameters: Based on the dawn storm auroral arc mask extracted by the auroral arc recognition model, the area and perimeter of the mask are calculated, and the area / perimeter ratio is used as the judgment criterion for the change of auroral arc width characteristics during the dawn storm.

[0053] 6. Calculation of auroral arc width variation characteristics: Based on the time sliding window method, the changes in auroral arc morphological parameters within the window are extracted, and the dawn storm auroral arc event information is output.

[0054] 7. Evaluation of dawn storm event detection: Based on real observation data, the performance of the entire algorithm in identifying dawn storm events is tested.

[0055] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0056] Example 1

[0057] like Figure 1 As shown, Embodiment 1 of the present invention proposes a Jupiter dawn storm event detection method based on machine learning, comprising:

[0058] Step 1: Data collection. Obtain the Jupiter aurora imaging observation data accumulated by the Hubble Telescope in the far ultraviolet observation band over the years from the official website of Aurora Planet Imaging and Spectroscopy http: / / apis.obspm.fr, and select the observation data after projection of the northern hemisphere polar region. The data time span is from July 3, 1997 to November 5, 2022. The image integration time is 100 seconds, the time resolution is 1 minute and 40 seconds, the image pixel size is 1481*1386, and the bit depth is 8 bits.

[0059] Step 2: Dataset creation. Based on continuous aurora observation data over a period of time in the northern hemisphere of Jupiter, combined with domain expert knowledge, samples of dawn storm events are screened out to construct a standard dataset of dawn storm events. The dawn-side aurora arc samples containing dawn storm characteristics are labeled using the open source software labelme (https: / / github.com / wkentaro / labelme). The specific steps include 2-1, 2-1, and 2-3.

[0060] Step 2-1: Image screening. The dawn storm event occurred before 12 noon local time on Jupiter. When screening aurora images, we only need to pay attention to the brightness and morphological changes of the aurora arc area before 12 o'clock. Check the continuous aurora images. When we find that the aurora arc in the images at different times before and after has poleward movement and the arc width increases along the latitude direction, save the relevant images as dawn storm event samples. A total of 178 samples were collected.

[0061] Step 2-2: Labeling the contours of auroral arcs. Use polygons to label the boundaries of auroral arcs. Label the pixels in the auroral arc area of ​​each image as 1, and the pixels in the background area as 0. Obtain the corresponding labeling results for the image and save them as JSON files as the labels for the images. Since there are few auroral arc samples in the dawn storm event, data enhancement is used to expand the samples. The auroral mask image after the contours are labeled and the auroral observation image before the annotation are rotated counterclockwise and clockwise in steps of 10 degrees to expand the samples to 980.

[0062] Step 2-3: Dataset division: 588 samples are used as training set, 200 samples are used as validation set, and 192 samples are used as test set.

[0063] Step 3: Auroral arc segmentation model selection and experiment. Considering the characteristics of auroral arcs with varied shapes and scales, and the small proportion of dawn storm event samples in real scenes. After experiments, the UNet network model with excellent performance in small sample learning tasks was used as the network model for Jupiter auroral arc morphology segmentation. The UNet model has many significant advantages in image segmentation tasks, including simple structure, information transmission of upsampling paths, end-to-end learning ability, efficient training and inference speed, powerful feature learning ability, high-resolution prediction results, and strong scalability.

[0064] Step 4: Evaluation of the auroral arc segmentation model. The samples of the Dawn Storm auroral arc annotation dataset are used as the input of the UNet network model, and the parameters are adjusted to train the Jupiter Dawn Storm auroral arc morphology extraction model. The model extraction results are obtained and evaluated. The commonly used evaluation indicators in the field of image classification are used to measure the model performance, including the Dice coefficient (Dice, Dice'scoefficient) evaluation indicator commonly used in the field of image classification to quantify the Dawn Storm auroral arc morphology segmentation results. The calculation formula of the evaluation indicator is as follows:

[0065]

[0066] Among them, A gt and A preare the baseline and the auroral arc area obtained based on the method in the image, respectively, and |*| is the number of pixels in the area. The value calculated by the indicator is in the range of [0,1]. The Dice coefficient indicates the similarity between the auroral arc area obtained based on the method and the baseline. The larger the Dice coefficient value, the better the model segmentation effect. As shown in Table 1.

[0067] Table 1 Evaluation results of the auroral arc segmentation model

[0068]

[0069] Step 5: Calculation of auroral arc morphological parameters. Based on the dawn storm auroral arc mask extracted by the auroral arc recognition model, calculate its area A, perimeter P and the ratio F of the two. Count the number of pixels with a pixel size of 1 in a single auroral arc mask image to obtain the area of ​​the mask. From the mask pixels, filter out the contour boundary pixels, and use the number of all boundary pixels as the perimeter of the auroral arc. The specific screening method of boundary pixels is as follows: Assume that the coordinates of mask pixel a are mask[i][j], where mask represents the mask image, i and j represent the pixel row number and column number respectively. If the pixel values ​​of the four points adjacent to point a, mask[i-1][j], mask[i+1][j], mask[i][j-1], and mask[i][j+1] are all 1, then point a is not a boundary pixel of the mask. Otherwise, point a is a boundary pixel of the mask. After calculating the mask area and perimeter, the area / perimeter ratio F is used as the key indicator of the change in the width characteristics of the auroral arc during the dawn storm.

[0070] Step 6: Calculation of the change characteristics of the auroral arc width. Based on the time sliding window method, the changes in the auroral arc morphological parameters within the window are extracted, and the dawn storm auroral arc event is output. The size of the time sliding window is set to L, the sliding step is set to S, and the auroral arc width change threshold is set to T. In actual application scenarios, the above three parameters can be set based on the sensitivity requirements of the test of the auroral arc width change characteristics to achieve the detection of dawn storm events of different intensities. The specific method for judging dawn storm is as follows: (1) The average value F1 of the area / perimeter ratio of all mask images in the L / 2 time period (0-L / 2) before statistics; (2) The average value F2 of the area / perimeter ratio of all mask images in the L / 2 time period (L / 2-L) after statistics; (3) When F2 / F1>T, there is a dawn storm event in the current window, otherwise it does not exist. After calculating the dawn storm events in all sliding windows, the time in different windows is deduplicated and merged, and the start and end time of the dawn storm event in a continuous time is output.

[0071] Step 7: Dawn storm event detection effect evaluation. Based on the auroral arc segmentation model trained in step 4 and the auroral arc width change feature calculation method in step 6, the detection effect of the algorithm is evaluated on real observation data. Based on the detection effect, repeated experiments are performed and the sliding window calculation parameters in step 6 are optimized. The selected evaluation indicators include accuracy, recall, precision and F1 value. For samples with an unbalanced ratio of positive and negative categories, the F1 value, which combines precision and recall, is a more objective and comprehensive evaluation standard than accuracy in binary classification problems.

[0072]

[0073] Among them, TP (True positive) is the true positive case, which correctly predicts the positive class as the positive class number; FP (False positive) is the false positive case, which incorrectly predicts the negative class as the positive class number; FN (False negative) is the false negative case, which incorrectly predicts the positive class as the negative class number; TN (True negative) is the true negative case, which correctly predicts the negative class as the negative class number.

[0074] The specific steps include: (1) First, based on the typical dawn storm event observation data, draw the trend chart of the mask area / perimeter ratio F in step 6, find the place where F increases, count its time range and F size, and preliminarily determine the values ​​of the sliding window T (8 minutes, 10 minutes, 20 minutes), sliding step S (1 minute, 5 minutes, 8 minutes), and threshold T (1.05, 1.08, 1.1); (2) Fix the step S to 1 minute and the threshold T to 1.05, set the sliding window size L to 8 minutes, 10 minutes, and 20 minutes respectively, and detect and evaluate the detection effect of the dawn storm event; (3) Fix the step S to 1 minute, modify the threshold T to 1.08, and evaluate the detection effect of the model under different L conditions again; (4) Similarly, modify the step S, fix the threshold, and evaluate the detection effect of the model under different L conditions again. Finally, generate the dawn storm event test evaluation matrix of the model under different L, S and T conditions. Take the parameters with the best comprehensive effect.

[0075] The observation data used for testing and evaluation include 37 image sequences of dawn storm events and 30 image sequences of non-dawn storm events. Through experimental tests, L is set to 8 minutes, S is set to 1 minute, and polar T is set to 1.08 seconds, which can achieve good aurora dawn storm event detection results. As shown in Table 2. Figure 2 This is the evaluation result on the validation set during the training process of the auroral arc segmentation model.

[0076] Table 2 Dawn storm event detection evaluation results

[0077]

[0078] According to the structural characteristics of the auroral arc of the dawn storm event in the Jupiter aurora observation image, this paper proposes an automatic auroral arc morphology extraction model based on the supervised deep network model UNet, and a judgment standard for the dawn storm event based on the auroral arc morphology parameter measurement. Finally, the segmented image of the auroral arc of the dawn storm event with pixel-by-pixel semantic annotation is obtained, as well as the mining of the information of the presence and occurrence time of the dawn storm event.

[0079] The proposed technical solution uses an end-to-end approach. After the model is trained, there is no need to adjust any parameters. Just input the image into the model to directly output the corresponding dawn storm area auroral arc morphology extraction results and dawnstorm event detection results. It is suitable for automatic analysis and processing of a large number of Jupiter aurora images.

[0080] Example 2

[0081] Embodiment 2 of the present invention provides a Jupiter dawn storm event detection system based on machine learning, which is implemented based on the method of embodiment 1. The system includes:

[0082] The detection output module is used to obtain Jupiter aurora images, input the pre-established and trained detection model, and output the image extraction results of the aurora arc area in the dawn storm area and the start and end time information of the dawn storm event;

[0083] The detection model includes an auroral arc segmentation module and an auroral arc morphological parameter calculation module based on the UNet network; wherein,

[0084] The aurora arc segmentation module is used to extract the dawn storm aurora arc mask image from the Jupiter aurora image;

[0085] The auroral arc morphological parameter calculation module is used to calculate the ratio of the auroral arc area and perimeter according to the dawn storm auroral arc mask image, extract the changes of the auroral arc morphological parameters in the window based on the time sliding window method, and output the dawn storm auroral arc event.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A Jupiter dawn storm event detection method based on machine learning, comprising: Obtain Jupiter's aurora image, input the pre-established and trained detection model, and output the image extraction results of the aurora arc area in the dawn storm area and the start and end time information of the dawn storm event; The detection model includes an auroral arc segmentation module and an auroral arc morphological parameter calculation module based on the UNet network; wherein, The aurora arc segmentation module is used to extract the dawn storm aurora arc mask image from the Jupiter aurora image; The auroral arc morphological parameter calculation module is used to calculate the ratio of the auroral arc area and perimeter according to the dawn storm auroral arc mask image, extract the changes of the auroral arc morphological parameters in the window based on the time sliding window method, and output the dawn storm auroral arc event.

2. The Jupiter dawn storm event detection method based on machine learning according to claim 1 is characterized in that: The method comprises a training step of an auroral arc segmentation module, comprising: Step S1) collecting Jupiter aurora imaging observation data, screening, marking dawn storm aurora arcs, and making a training set; Step S2) The training set data is sequentially input into the UNet network, the network parameters are adjusted, and the Dice coefficient is used as an evaluation index to obtain a trained auroral arc segmentation module.

3. The Jupiter dawn storm event detection method based on machine learning according to claim 2 is characterized in that: The step S1) comprises: Data collection: Obtain the Jupiter aurora imaging observation data accumulated by the Hubble Telescope in the far ultraviolet observation band over the years from the official website of Aurora Planet Imaging and Spectroscopy, select the observation data after projection of the northern hemisphere polar region, and select the set time span, set time resolution and image pixels; Image screening: Screen and obtain dawn storm event samples; Auroral arc contour annotation: Use polygons to annotate the auroral arc boundary, annotate the pixels in the auroral arc area of ​​each image as 1, and annotate the pixels in the background area as 0, obtain the corresponding annotation results of the image, and save them in json format as the label of the image, and expand them by data enhancement; Data set division: Divide the data into training set, validation set and test set according to a certain ratio.

4. The Jupiter dawn storm event detection method based on machine learning according to claim 2 is characterized in that: The calculation of the Dice coefficient in step S2) satisfies the following formula: Among them, A gt and A pre are the baseline and method-based auroral arc regions in the image, |*| is the number of pixels in the region, and the Dice coefficient is in the range of [0,1].

5. The Jupiter dawn storm event detection method based on machine learning according to claim 1, characterized in that: The auroral arc morphological parameter calculation module includes an auroral arc morphological parameter calculation unit, and the processing process includes: Get the coordinates of each pixel from the auroral arc mask image. For pixel a, the coordinates are mask[i][j], where mask represents the mask image, i and j represent the pixel row and column numbers respectively. If the pixel values ​​of the four points adjacent to point a, mask[i-1][j], mask[i+1][j], mask[i][j-1], and mask[i][j+1] are all 1, then point a is not a boundary pixel of the mask. Otherwise, point a is a boundary pixel of the mask. After counting the number of pixels with a pixel size of 1 and calculating the mask area and perimeter, the ratio F of the mask area to the perimeter is used as the key indicator of the characteristic change of the auroral arc width during the dawn storm.

6. The Jupiter dawn storm event detection method based on machine learning according to claim 5, characterized in that: The auroral arc morphological parameter calculation module includes an auroral arc width variation characteristic calculation unit, and the processing process includes: Step T1) setting the size of the sliding time window to L, the sliding step size to S, and the aurora arc width change threshold to T; Step T2) Count the average value F1 of all ratios F within the first L / 2 time period; Step T3) Count the average value F2 of all ratios F within the L / 2 time period; Step T4) Determine the size of F2 / F1 and T. If F2 / F1>T, there is a dawn storm event in the current window, otherwise there is no dawn storm event; Step T5) After calculating the dawn storm events in all sliding windows, the times in different sliding windows are deduplicated and merged, and the start and end times of the dawn storm events in a continuous period of time are output.

7. The Jupiter dawn storm event detection method based on machine learning according to claim 6 is characterized in that: The step T1) also includes a verification and adjustment processing step, including: According to F, draw a trend graph, find the place where F increases, count its time range and F size, and preliminarily determine the size of the sliding time window L, the sliding step S and the initial value of the auroral arc width change threshold T; Adjust L, S and T separately, evaluate the detection effect, and get a set of values ​​with the best overall effect.

8. A Jupiter dawn storm event detection system based on machine learning, characterized in that: include: The detection output module is used to obtain Jupiter aurora images, input the pre-established and trained detection model, and output the image extraction results of the aurora arc area in the dawn storm area and the start and end time information of the dawn storm event; The detection model includes an auroral arc segmentation module and an auroral arc morphological parameter calculation module based on the UNet network; wherein, The aurora arc segmentation module is used to extract the dawn storm aurora arc mask image from the Jupiter aurora image; The auroral arc morphological parameter calculation module is used to calculate the ratio of the auroral arc area and perimeter according to the dawn storm auroral arc mask image, extract the changes of the auroral arc morphological parameters in the window based on the time sliding window method, and output the dawn storm auroral arc event.

Images