Street Level Analysis

Use of machine learning in damage assessment has been implemented by FEMA and emergency responders in the recovery process. Damage assessment machine learning algorithms help users better understand where damage is located and how to prioritize aid. Yet, many current methods reinforce existing inequalities. We asked questions about the data collection process, composition of training data, and annotation methods to better understand  who this algorithm leaves behind and what the implications when machine learning predictions are incorrect. We curated new datasets and new methods to address those concerns. Our goal was to prevent bias in machine learning tools while making it more accessible and understandable to a wide variety of users. 

Machine Learning Basics & Tutorials

These tutorial videos will guide you through all the necessary steps to create your damage assessment model. The platforms included in the videos provide an accessible walkthrough for model design and image annotations, and Google Colaboratory notebook provides pre-written Python code to run a custom YOLOv5 model through the Roboflow platform. These videos highlight the ease in access to machine learning platforms for open-sourced utilization for anyone interested in this field

New and worthly
3 min read

GitHub Repository

Our analysis has shown many of the available data sets contain inherent biases and inaccuracies for disaster recovery. We're sharing our curated and disaster specific datasets on GitHub.

New and worthly
3 min read

Google Colaboratory Notebook

We’re sharing our code and our trained damage assessment model for anyone to replicate and use.

New and worthly
3 min read

Roboflow

We found computer vision needs better labeled data sets. Roboflow is a tool we used to curate the data sets above.

New and worthly
3 min read

Lobe.ai

There's a movement towards making AI accessible to everyone. We used Lobe.ai  to label, train, and test with ease.

Project Documentation

Working Papers