KCL Space| Automated Detection of Hazards
Automated Detection of Hazards
Spotting Cloud Streets Using ML
Summary
The natural phenomena detected is cloud streets because they can predict turbulent weather, especially during the times of climate change. Cloud streets are clouds which form due to a specific wind pattern known as horizontal convective roll. These winds rise due to warm surface temperatures but sink when they reach the stable air above. This destabilises local atmosphere and causes variations in convective available potential energy, a necessity for weather hazards. We use a machine learning detect this phenomenon. The model takes in a RGB satellite image and predict cloud streets’ existence. The ML model is complemented by a web app, which informs the user whether cloud streets is present.
How We Addressed This Challenge
Observing satellite imagery is often the first stage of detecting major global events such as climate change. Early detection can prove useful for scientists to prepare the society to overcome such challenges and perhaps find solutions at an early stage. Our project attempts to provide such a tool which allows the user to detect the formation of cloud streets, or the underlying horizontal convective rolls. Considering the associated weather-related consequences of cloud streets, such a tool, especially the use of AI, can help predict climatic variations in a particular region drastically faster. Further extensions of this web app can provide detailed information regarding the importance of such phenomena taking place in different parts of the world.
How We Developed This Project
The motivation behind the project is to able to spread awareness about a phenomenon that is not known by many, but has countless minor impacts that can, in some form or the other, influence everyday life.
The approach consisted of developing a model for understanding and studying the data in as much detail as possible, in order to build a user-friendly, interactive web app.
Our project consists of three primary segments:
- The client side is a website developed using Vue along Javascript frameworks such as Ajax and cropper.js
- The server side is developed using Django which runs on the Python programming language and serves as a backend which is retrofitted with the ever popular django rest framework
- The brains of the of the entire project is a machine learning model which is trained using Tensorflow along Keras which is trained on the Inception V3 model. We performed transfer learning to obtain faster results in the given short duration in a hackathon
The challenges included training the model on a very small data set, with transferred learning, to save time due to lack of time, and insufficient data. Further, limited computing resources, and systems with no GPUs made it extremely difficult to train the model precisely.
The team tried to overcome this by training it on CPUs, however, a large number of epochs weren’t achieved.
How We Used Space Agency Data in This Project
We used the Prelabelled datasets from NASA to train and validate the machine learning model. Imagery from NASA WorldView was used to create the map in our web app.
Project Demo
Data & Resources
- Nasa WorldView Map
- Prelabelled datasets from NASA-IMPACT data-share were used to train and validate the model.