Disaster Risk Monitoring System - Remote Sensing and Deep Learning

Introduction

I come from a state back in India where Floods is an yearly phenomenon taking hundreds of lives every year. Similarly, Natural disasters such as wildfire, drought, and severe storms wreak havoc throughout the world, causing billions of dollars in damages, uprooting communities, ecosystems, and economies. The ability to detect, quantify, and potentially forecast natural disasters can help us minimize their adverse impacts on the economy and human lives.

In this post I will go through a real world example where I will start with a PreTrained Image Segmentation Model and create a Flood detection segmentation model using satellite imagery. Using satellites to study Flood is advantageous since physical access to flooded areas is limited and deploying instruments in potential flood zones can be dangerous. Furthermore, satellite remote sensing is much more efficient than manual or human-in-the-loop solutions.

The Disaster Risk Monitoring System is like a workflow:

1. Satellite remote sensors capture data and sends it to a Data Center
2. Data is used to (continuously) train deep learning neural network models
3. Different models and versions are managed by the model repository
4. Model inference performance is actively monitored
5. Data is passed to the inference server
6. The deep learning inference results are post-processed for either
7. Further analytics by 3rd party or
8. Raising alerts

I will use Nvidia GPU Cloud Subscription and a Pre-trained Model TAO Toolkit, TensorRT , Triton Inference Server

Then I will use the Data collected from Remote Sensors to train the Model using Transfer Learning.

Image Data Preprocessing for Deep Learning

Environment Setup and Docker Image Pull. This notebook outlines Environment Setup

Check the Exploratory Data Analysis Notebook here

Training Deep Learning Model using Transfer Learning

I will use Transfer Learning and TAO Toolkit to Train a Segmentation Model.

TAO Toolkit is an easy to use low code no code Transfer Learning Platform that helps folks like me to easily Train a Vision Transformer Model

This step comprises of taking a PreTrained Segmentation Model U-Net. U-Net model is a convolutional neural network (CNN) that uses a U-shaped architecture to perform semantic segmentation.

Semantic segmentation is a computer vision task that assigns a class or object to each pixel in an image.

I follow these steps and use Transfer Learning to Train the U-Net Model

• Prerequisites for Model Training
• Download the Pre Trained Model
• Data Preparation - this is where most of the work happens
• Model Training
• Modify Model Configuration Files
• Start Model Training
• Evaluate the Model
• Visualize Model Inference

This Notebook covers the Steps.

Segmentation Model Production Deployment

In this step, I will use the trained segmentation model and deploy it on Triton Inference Server. I will also use TensorRT. TensorRT is a highly optimized package that takes trained models and optimizes them for inference.

This Notebook covers the steps in detail.

Conclusion

Traditional Disaster Risk Monitoring requires significant domain expertise. It also relies upon many different data types, information sources, and types of models to be effective. Modern computer vision, using machine learning techniques, can perform different tasks such as classification, object detection, and segmentation . It generates Promising results but requires powerful technology to enable complex computations