Sam Edds | Predicting Natural Disasters | Rise of AI Conference 2022

Ai

"Discover how AI can predict and combat natural disasters, exploring successes and challenges with data from summer-dominant datasets and future directions in pandemics and crime patterns.

Key takeaways
  • Results for model trained on point of view images and predicting on aerial images are 60% for landslide, 78% for flood, and 85% for fires, but struggle with predictions of buildings and infrastructure.
  • Points of bias in the data include dominance of images from summer, lack of seasons other than summer, and questionable labelling.
  • To fight bias, data scientists should look carefully at label data, quality check it, and bring in diverse opinions to combat biases.
  • Future directions include exploring other natural disasters and scenarios where data science and AI can be applied, including pandemics and crime patterns.
  • To scale and implement these models, it is important to refine models, use pre-trained models that can be easily applied to many scenarios, and aggregate results to make predictions easier.
  • For public good, using explainable AI and deep learning in this context can help us combat natural disasters and make our data science more transparent and accountable.
  • Bias is always present in data science and data quality, and it is our job to try to combat bias at every stage of the process.