Rain Scanner Radar and Optical Flow Combination For Early Identification of The Incoming Bow Echo Heavy Precipitation

In this study, heavy precipitation in the form of bow echo is observed, identified, and tracked using a rain scanner and optical flow method. Two case studies of bow echoes with a range of properties from 15 to 33 km were observed in the afternoon on 1st February and 4th February 2021. The rain scanner developed based on X-band radar could obtain the first detection of the bow echo shape. The difference between these two bow echoes events is from their initiation. The first bow echo cases are initiated from weakly organized cells. Meanwhile, the second case of bow echo is initiated from a squall line event. Bow echo is initially observed in the western part of Bandung city. Approximately twenty minutes after the first detection, the bow echo passed the center of Bandung city. Image processing using edge detection could track the bow echo from the radar data. Applied Lucas-Kanade optical flow was then used to retrieve the bow echo movement. Vector flow from optical flow showed the dominant movement toward the east direction in both case studies. A simple method to predict the incoming time of the bow echo using the estimated speed from vector flow and the image translation is also presented.


Introduction
Heavy precipitation in the form of a bow echo is associated not only with high rainfall intensity but also with strong wind occurrence on the surface [1,2,3].Information on the early detection of the bow echo could become valuable information for the incoming heavy precipitation.
The initiation of bow echo comes from sources that evolve and transform into a bow echo shape.A previous study by Klimowski et al [1] classified the initiation phase of the bow echo that comes from three sources: weakly organized cells, squall lines, and supercells.The first two sources (weakly organized cell and squall line) are more common to occur compared with the last source (supercell).The weakly organized cell comes from the merger of several cells into a bow-echo shape.The bow echo 1245 (2023) 012031 IOP Publishing doi:10.1088/1755-1315/1245/1/012031 2 movement tends to follow the cell that initiates the merger process [1].A squall line is also another possibility of the source that could transform into a bow echo shape.
To detect the initiation phase, it is essential to be able to detect the bow echo pattern.Several studies have proposed methods using various techniques.Kamani et al [4] use the skeleton matching method to detect the bow echo as an indicator of severe weather.The bow echo feature from the radar data is extracted using the skeletonization method and skeleton context to match the shape pattern of the bow echo.A different approach is done by Fanlin and Jinyi [5] which measured the contour to find the curve of the bow echo.Another method used pseudo reflectivity combined with the Convolutional Neural Network to detect the bow echo [6].
In this study, we present an alternative method to detect the bow echo using image processing.Two case studies of bow echo observed by RSR are used as an example of the applied alternative method.The objective of this study is to detect the early bow echo shape that capture by radar and then predicted the movement using optical flow and image translation.The structure of this paper is described as follows.Bow echo detection is discussed in section 2. Meanwhile, section 3 presents the bow echo nowcast.Two case studies to test the proposed method are described in section 4. The conclusion is given in section 5.

Overview of the Bow Echo from Rain Scanner Radar
The two main data used in this study are Rain Scanner Radar (RSR) and Automatic Weather Station (AWS).RSR is a modified marine X-band radar dedicated to rain observation using single polarization (horizontal).The maximum coverage area is 44 km with a spatial resolution of 120 x 120 meters.The RSR is located in Bandung city (107.59°E, 6.89° S) which has been tested concerning the accuracy and reliability of rain detection [7,8,9].The RSR observation dataset is collected from 2019 to 2021 to obtain the bow echo event.However, in this study, we present only two case studies of the bow echo event (February 1 st and February 4 th , 2021).The confirmation of the heavy precipitation carried out by the bow echo event is obtained from AWS measurement data.

Bow Echo Detection
Two methods are used to determine the bow echo, a visual method and a proposed method of shape recognition.The visual method is applied by analyzing the image sequence visually.Meanwhile, shape recognition using image processing is used to find the bow-like shape from the image sequence which resembles the bow echo.
Shape recognition consists of three steps, a) identifying the object, b) measuring the properties of the object, and c) determining the bow shape.The first step of identifying the object consists of three parts, convert the image into grayscale, image filter by removing the small object, and labeled the object.The second step is to measure the properties of the object.This is done by utilizing the regionprops function [10].Four parameters from the regionprops output (perimeters, filled area, circularities, eccentricity, orientation, and equivalent diameter) are used as a threshold to determine whether the object suits the criteria of a bow echo shape or not.
Perimeter is the distance between each adjacent pair of pixels surrounding the region's border by measuring the distance between them.The filled area specifies the number of pixels within a region where all holes are filled in.Circularities are the measure of the roundness of an object, with a value of 1 representing a circled object and more than 1 showing a less circular object.Circularities are calculated by using two parameters (Filled area and Perimeter) using Equation 1. Eccentricity is defined as the ratio of the lengths of the focal point and major axis of an ellipse, with a range value between 0 and 1. Eccentricity with the value of 0, is interpreted as a circle object, meanwhile, if the eccentricity with the value of 1 represents the line segment object.Orientation is defined as the angle between the major axis of the object with the x-axis.Circularities= perimeters 2  4 × π × Filled Area (1) Several tests using the combination of circularities, eccentricity, and orientation are carried out towards several bow echo shape objects.Thirty images with a possible of bow shape object are tested using the combination of these three parameters.From this test, we proposed a new method to detect the bow echo by using these three parameters with certain thresholds described in Table 1.The proposed method of shape recognition using threshold from three parameters (circularities, eccentricity, and orientation) is then applied to two case studies of bow echo occurrence observed by the RSR.

Lucas Kanade Optical Flow
The advantage of the RSR is its high temporal resolution [11].The high interval of the image sequence generated by the radar could give valuable information on the movement of the precipitation echo.One method to calculate the precipitation movement from the image sequence is the optical flow.Optical flow is a method to identify the movement of an object from a sequence of images [12].Optical flow is defined as the apparent motion of individual pixels on the image plane.It often serves as a good approximation of the true physical motion projected on the image plane.The more rapid the interval of the image sequence, the more information about the object movement could be retrieved.Several approaches could be applied to the optical flow such as the local differential, global variational, and phase-based [13].Two common optical flow categories often used are global variation and local differential.Global variation utilizes the entire image pixel in the calculation of the vector field.Meanwhile, the local differential only uses a small neighborhood area in the calculation.Lucas Kanade is the optical flow type that uses the local differential principle [14].Lucas Kanade intention is to calculate the vector of optical flow using an assumption that the small area near the neighborhood will be similar to the vector.

Prediction
In general, the nowcast process consists of three elements, namely identification, tracking, and prediction.Identification and tracking of the bow echo are already been done by shape recognition.Meanwhile, prediction is the process of guessing or predicting the next bow echo movement, in this case on a short-term scale, namely the next 20 minutes (prediction interval of 2 minutes).20 minutes is considered sufficient to provide information for users in making decisions.Since the movement of the bow echo in both case studies is linear, a simple image translation is carried out.The input for the pixel movement is taken from the average motion flow value calculated by Lucas Kanade optical flow.
Image translation is a simple tool to shift the image towards a certain pixel location.In this study, we use the assumption that the bow echo is commonly has a bigger, uniform shape and only affected by the environmental wind condition.The properties of how many translations of the object are based on the average motion flow.The prediction result was tested using RMSE (Root Mean Square Error) and IoA (Index of Agreement).AWS instrument captured the incoming heavy precipitation that was carried out by the bow echo.The graphical data of rainfall from AWS started to increase after 17:30 LT which the nearest location of the bow echo towards the AWS location is at that time.The bow echo started to unravel after 17:38 LT is probably because the rainfall occur at the radar location which is the same location as the AWS.This condition creates high attenuation around the radar location.However, the wind speed parameter from AWS data is not capturing any high fluctuation of wind speed.The dominant movement of the bow echo towards the east direction is strengthened by the AWS measurement which shows a wind direction value of 270° at 17:40 LT.

. Bow echo detection and prediction
In order to quantify the exact detection of the bow echo, the proposed method is applied to the RSR data.The image from RSR is then processed using image processing to obtain the object feature.By using this procedure, we detect and track the bow echo shape from 17:16 LT to 17:38 LT.With an interval of two minutes, the proposed method could track eleven from a possible twelve-bow echo shape (Fig 2c).The reason why the object at 17:28 LT is that the circularity value is more than the criteria on the proposed method in section 2.2.
The early information of the bow echo at 17:16 LT could be utilized to predict the incoming heavy precipitation.Optical flow using Lucas Kanade is utilized to obtain the average motion flow value (Fig 3 panel A).The image used is from 17:10 LT to 17:18 LT to gather the information before, during, and after the early detected bow echo.The average motion flow is 7.36 (rounded to 7) pixels per 2 minutes.Panel B showed the nowcast result with the evaluation comparing the actual and prediction result depicted in panel C in Fig 3 .Since the prediction method only use a simple translation method (without considering the growth decay of the precipitation echo), the RMSE and the IoA did not show a good result.Several minutes after that, the bow echo shape is formed with the movement is still similar to the previous squall line.
The bow echo approached the AWS location at 16:24 LT and was followed by the increased reading of AWS wind speed value (9.71 m/s) at 16:30 LT.Later, rainfall is detected at 16:40 LT (3.6 mm) with the maximum reached at 16:50 LT (4.3 mm).Wind direction showed small changes during the period from 16:20 (294°) until 16:30 LT (225°).

Bow echo detection and prediction
Investigating the first bow echo in case study 2, we compared the image radar data at 15:54 LT and 15:56 LT.From a visual point of view, these two images from panel A (Fig. 5a and Fig. 5b) look like a bow-type shape.After we processed it using the same procedure as case study 1 (converting to grayscale, .05which is below the threshold in Table 1).We then checked the bow echo sequence and track it from 15:54 to 16:24 LT.However, the bow echo only detects until 16:12 LT.The proposed method did not consider the image from 16:14 to 16:24 as a bow echo shape type.The proposed method also did not detect the bow echo pattern at the 16:04 LT image.The reason is similar to the previous case study number one, where the circularities value is lower than the threshold of the proposed method.Again, we utilized the information of the early detection of bow echo at 15:56 LT.The dataset used to calculate the average motion flow are from 15:48 to 15:56 LT.Panel A in Figure 6 showed the average motion flow with the value of 5.15 (rounded to 5) pixels per 2 minutes.The translated image for the next twenty minutes starting from 15:58 LT are depicted on panel B. Meanwhile, the evaluation for the RMSE and IoA are shown in panel C Figure 6.

Discussion
The proposed method using a combination of three parameters (circularities, eccentricity, and orientation) is presented in this study.This method is different from other previous research [4,5,6] since it uses a combination of three parameters from the shape measurement of an object image to detect bow echo.In this section, the limitation of this study, as well as future works, are discussed.

Limitation of this study
Two case studies demonstrate the ability of the proposed method to detect heavy precipitation in the form of bow echo coming from the west towards the east side of the Bandung basin.The incoming heavy precipitation is thought to be influenced by the west monsoon condition since occurs in the February period [15].However, other patterns of precipitation distribution occurred in the Bandung basin such as the precipitation distribution between the north-south area [16].It will be very interesting to see whether the proposed method could observe the bow echo in different distribution patterns.Additional case studies are needed to ensure the capability and accuracy of the proposed method could detect the bow echo with different patterns.

Future Works
The early identification of bow echo has the potential to be utilized for early detection and prediction of extreme rainfall.Two case studies demonstrated a lead time of more than 20 min that could be used not only for early mitigation but also for nowcast application.This study presented the preliminary nowcast using optical flow and image translation.Evaluation of the nowcast result with the actual RSR observation using RMSE and IoA showed low accuracy and needs more improvement on the prediction method.Improvement should be done by considering the factor of growth and decay.Machine learning could also be considered to be applied to the nowcast radar image [17].

Conclusion
In this paper, we proposed an alternative method to detect the bow echo pattern using the combination of the measured circularities, eccentricity, and orientation of the object.The threshold value of this combination is then tested in two case studies of a bow echo.The bow echo event has a similar characteristic where is detected from the west direction moving east towards the AWS location in Bandung city.The difference is the initiation of the bow echo where the first case study comes from several weak convective cells that combine into a cell with a bow echo shape.Meanwhile, the bow echo is generated from the squall line in the second case study.The incoming heavy precipitation carried out by the bow echo is confirmed by the AWS reading on both case studies.Furthermore, an early increase in wind speed is also observed in the second case study.The proposed method could determine and quantify the first bow echo several minutes before it reaches the AWS location.Simple predictions for the next 20 minutes are carried out starting from the first detection of the bow echo.

1 .
Case study bow echo 1 st February 2021 4.1.1.Observation from RSR and AWS Visual analysis is conducted toward the image of the RSR to locate the bow echo sequence.The result shows that the bow echo sequence occurred from 17:16 -17:38 LT (Local Time) on February 1 st , 2021.The source of the bow echo was also captured by the radar at 17:04 -17:14 LT.Several convective cells that merge into one big cell are moving in the east direction depicted at 17:04 LT.The merging cell is established at 17:16 LT in the form of a bow shape alike.The bow echo shape is well preserved until 17:38 LT and started to unravel after that.

Figure 3 .
Figure 3. Nowcast bow echo case study 1, a) Motion flow from Lucas Kanade Optical Flow, b) Nowcast result, c) Evaluation between prediction and actual result.

4. 2 .
Case study bow echo 4 th February 20214.2.1.Observation from RSR and AWSThe same procedure is also carried out to the image sequence in case study number two.The bow echo is observed on February 4 th , 2021 from 15:56 to 16:24 LT based on the visual image sequence of the RSR (Figure4).The source of this bow echo is well observed from 15:42 until 15:52 LT.Long stretched precipitation echo or a squall line moving towards the east direction are observed during this period.

Figure 4 .Figure 5 .
Figure 5. Bow echo detection on case study 2, a) Bow echo at 15:54 LT, b) Bow echo at 15:56 LT, c) Bow echo track.The different between Fig 5a and Fig 5b is that the parameter label value have a different color which shows that Fig 5a is not in a bow echo criteria (circularity = 6.05 which is below the threshold in Table1).

Figure 6 .
Figure 6.Nowcast bow echo case study 2, a) Motion flow from Lucas Kanade Optical Flow, b) Nowcast result, c) Evaluation between prediction and actual result.

Table 1 .
Proposed threshold for bow echo detection using shape recognition