Investigations of Mesoscale Convective Systems over the Indian Region with Image Processing Techniques

Abstract

Mesoscale Convective Systems (MCS) are convective clouds responsible for precipitation, severe newlineweather conditions, and property damage. Studying MCS patterns is crucial, which are varied by newlineglobal warming and pollution, leading towards extreme weather conditions. Detection and tracking newlinemethods are essential for understanding MCS characteristics, based on cloud temperature and size newlinethresholds. Numerous studies have contributed to the identification and tracking of MCS, with newlinelimited methods for identification. This research uses images from the Indian satellite Kalpana-1 to newlineunderstand their typical lifecycle. newlineThree MCS identification techniques based on the Hue channel of the Hue Saturation Value (HSV) newlinecolor model are proposed and implemented with an assessment of their effectiveness. These newlinetechniques execute segmentation based on Hue channel Thresholding (HT), K means clustering newlinecombined with Hue channel Thresholding (KMCHT), and the modified Source Apportionment newlineTechnique combined with Hue channel Thresholding (SATHT). Image pixel values are used to newlinedepict infrared brightness temperature data obtained from the Indian geostationary satellite Kalpana- newline1. The generated ground truth images and performance measurements are used to assess the newlineeffectiveness of the methods. The proposed SATHT method for multiple cloud segmentation results newlinein superior performance metrics than the HT and KMCHT approaches. Recently Deep learning newlinemethods are provided state-of-the-art results for image processing tasks such as Image segmentation, newlineimage classification and their analysis. Therefore Deep learning (DL) methods are used to segment newlineMCS from the IR BT images using a transfer learning (TL) strategy. VGG-16 has been improved in newlinethe present work by fusing encoder-decoder architecture and taking cues from UNet architecture. newlineThe design that is generated is called ENDE-VGG. The proposed approach enables the collection of newlinemore pertinent data over a wider region. The TL technique produced a Dice coefficient of 0.935 on newlinethe validation data set, an intersection of union (IoU) of 0.875 on the K1-IR data set, and a mean IoU newlineof 0.93. Additionally, using the test data, it obtained a 0.933 Dice coefficient and 0.875 IoU. ENDEVGG newlineperformed better than the most sophisticated cloud image segmentation methods when using newlineIoU as the loss function, based on numerous research studies. The proposed work focuses on newlinedetecting the center of the cloud, a deformable object with irregular shapes. Two techniques are newlineused: regular shape fitting and moments-based technique. Results show spatial moments outperform newlineshape fitting for cloud location detection. newlineThe area overlapping technique is a widely used method for tracking convective storms in infrared newlinesatellite images, often combined with other techniques to overcome its drawbacks of splitting and newlinemerging or lack of cloud overlap. Under MCS tracking 2 methods are implemented, Area newlineoverlapping and proposed deep learning-based Deep MCS i.e. DMCS tracker. A novel DMCS newlinetracker to identify and track MCS is proposed, which automatically learns the features of MCS using newlinean unsupervised Deep Neural Network model convolutional autoencoder. Convolutional autoencoder newlinegives a compressed representation of MCS features utilizing a trained encoder part of it. MCSs are newlinetracked using learned features across image sequences using a Structural SIMilarity index (SSIM) newlinemetric among two applied feature vectors. Results show that applying autoencoders to learn the deep newlinefeatures of the MCS helps to find the association among the MCS present in satellite image newlinesequences. It tracks MCSs efficiently though they vary in shape, size, and location. The method is newlinetested using half-hourly BT images for a month s data with training data from the eleven months newlineimages. Results are compared with the state-of-the-art method of area overlapping. The proposed newlinemethod works well with splitting and merging clouds as well as non-overlapping convective systems newlinewhere the area overlapping method fails to track. Thresholding and area overlapping methods are newlinewidely used to study MCS characteristics over the different regions of the world. Using the same newlinemethods, MCS characteristics are studied over an Indian region which will be useful for the newlinemeteorologists to understand the pattern of MCS occurrences. newline

Description

Keywords

Citation

item.page.endorsement

item.page.review

item.page.supplemented

item.page.referenced