top of page
After a natural disaster, a key task is to determine priority areas for responders. A critical component in this determination is the extent of damage an area has sustained, and quickly identifying the hardest hit areas can make a huge difference in saving lives.
As a result of climate change, the severity of natural disasters is rapidly increasing, causing more than $200 billion in damage worldwide each year. These disasters can cause lasting economic and agricultural disruptions, displace entire populations, and result in decades of community reconstruction. Natural disasters cannot be prevented entirely, but the effectiveness of emergency response can be improved.
The opportunity for deep learning
This current approach to damage identification relies on human experts to classify images by hand. It not only takes more time, but also requires numerous employees with the necessary training. In comparison, deep computer vision models can train faster and classify images with greater speed and accuracy.
Platform model to use
For this use case, we suggest using an image classification model.
How does the model work?
Image classification model predicts the likelihood that a particular image will have a particular label. Those labels can be as simple as "damaged" or "undamaged" or more detailed and reflect different types or degrees of catastrophic damage. As the training process proceeds, the model learns the complex relationship between the features of the image and the corresponding label. An output of the model is a probability that an image has a particular label. For example, the model might output that the image has a 34% probability of being labeled "undamaged" and a 66% probability of being labeled "damaged." We can then set the threshold that the model uses to determine if the probability is high enough. The higher the threshold, the higher the probability that the model will label an image, but determining the "right" threshold depends largely on your use case and needs.
In order to ensure a successful model, the data should reflect the characteristics of typical data collected after natural disasters. This could be, for example, satellite imagery from high or low altitude showing damaged and undamaged areas, or imagery of different types of damage such as flooding or debris. The images should all have the same shape, and ideally we should have a similar number of images for each label. In order for the model to learn the relationships between the images and the disaster types, you need a table that lists the image names and their corresponding labels.
Modell Performance and success
You can take a closer look at the performance of your model in the evaluation view. In particular, you should look at the rate of false positives and false negatives. Although priorities may vary among emergency response organizations, the consequences of incorrectly labeling a damaged area as undamaged are generally greater than those of labeling an undamaged area as damaged. If you have chosen instead to use multiple labels in your model (e.g., damage types), you should examine the model's performance for each label. Are there certain types of damage that your model detects better than others? Understanding these results will help you make the best decisions when using the model's results.
Where to learn more?
If you are interested in building this use case, our damage classifier tutorial is a great place to start. We included a tutorial inside the platform that leads you through how to build an image classifier that can predict multiple types of damage to cars.
bottom of page