Predicting Canopy Damage in Sonoma County using Convolutional Neural Networks

Presenter: Tyler Whitmarsh

Presenter Status: Undergraduate student

Academic Year: 22-23

Semester: Spring

Faculty Mentor: Gurman Gill

Department: Computer Science

Funding Source/Sponsor: Class Project

President's Strategic Plan Goal: Sustainability and Environmental Inquiry

Screenshot URL: https://drive.google.com/uc?id=1t_b8sRR1me_-QdIH5C_fnlmKKCDe3DCC

Abstract:
This research project aims to classify the severity of forest canopy damage caused by fires in Sonoma County using computer vision techniques. The project uses training and testing data derived from detailed information from Sonoma vegetation maps, which provide precise fire damage classification information about the 2017 Sonoma County fires. The data is preprocessed using ArcGIS's deep-learning toolset to transform it into usable training and testing datasets for the computer vision pipeline. The project employs a pipeline of custom models based on a Unet architecture, which is powered by TensorFlow, Keras, and OpenCV. The models are trained separately and then used in a soft voting ensemble method to create a segmented image prediction of the fire damage classes. The resulting ensemble model shows promise in accurately predicting different types of fire damage, providing essential insights in the areas most vulnerable to fire damage. This project highlights the potential of computer vision techniques in analyzing complex environmental phenomena and has significant implications for the prevention and management of wildfires