The problem CycloNet solves

Cyclone intensity estimation plays a major role in disaster management actions. The existing model that is used by ISRO for Cyclone intensity estimation is the HWRF model which has some disadvantages. The HWRF (Hurricane Weather Research and Forecast) Model requires accurate center determination of the cyclone imagery to estimate the intensity correctly, which is difficult during the initial stages of cyclone formation.
HWRF Model is run four times a day 00Z, 06Z, 12Z, 18Z (equivalent to Greenwich Mean Time) by NCEP (National Center for Environmental Prediction) with on-demand input. But INSAT3D observations are available at every 15 minute interval, therefore by using the HWRF model we lose a tremendous amount of data that can be utilised to understand the structural changes at every half hour interval. This calls for the need of a model which can compute intensity of the Cyclone IR imagery regardless of the centre determination, which can be useful to understand the instantaneous structural changes in the initial stages of the cyclonic event for making strong predictions.Deep Learning can identify more complex features in various images of the cyclones to identify certain patterns and understand a deeper relationship between the image of the cyclone and its respective intensity.Our Solution mainly targets development of a Web-App User Interface where the user can upload INSAT-3D IR Satellite Image of Cyclone which is then passed to our Deep Convolutional Neural Network built in PyTorch which is trained on Cyclone imagery of various intensities on our custom dataset curated from Raw INSAT-3D satellite captured images on MOSDAC server to compute the intensity of the uploaded INSAT3D cyclone image and forward the submitted values to the archive via the database. The CNN eliminates the need for usage of traditional methods for accurate center determination to estimate the Cyclone intensity using Satellite imagery.

Challenges we ran into

We encountered several bugs while linking our frontend with backend model and database together in Flask. We had to learn some new technologies to include special features for our model and faced several challenges doing it in the first attempt like deploying a deep learning model in Flask and using the Windy API to implement our ideas like including a Live wind map and take latitude and logitudnal coordinates to display the wind pattern on the redirected location . For training the model, the most important factor is data. GOOD DATA = BETTER AI . There were no labelled dataset for our objective on the internet. So to make the most out of the raw data from the satellite, we annotated the data ourselves to minimize any chance of error. The dataset is created by deriving of intensity of each datapoint from a graph available for each record on the MOSDAC server. The dataset now consists of all the 135 cyclone INFRARED images captured by the satellite "INSAT3D" from 2012 to 2021 with their given intensities. We, collectively organized our schedules to complete the project on time and also worked upon improvising our user interfaces, layouts and animations to give a good experience to the users. Taking the time constraints in mind, we showed significant teamwork to build the complete working app. Together with the inputs from all team members, we were able to decode the issues successfully and experienced solving problems from different point of views.