UAV-based Wildfire Risk Assessment Through Geospatial Machine Learning

By: Kevin P.
Year: 2024
School: Sage Hill
Grade: 11
Science Teacher: Aaron Soffa

Wildfires have increasingly become a significant threat, wreaking havoc on ecosystems and communities. With the growing urgency to prevent these destructive events, innovative solutions are imperative. Kevin’s project introduces an approach to wildfire risk prediction through the development of a novel drone platform. This platform is designed to classify areas into wildfire risk categories by conducting aerial surveys that measure various environmental variables such as temperature, humidity, vegetation health, and topography.

The primary objective of the project is to improve the precision and reliability of wildfire risk assessments. Traditional methods relying on satellite imagery face limitations in spatial and temporal resolution, which can hinder timely and accurate predictions. In contrast, unmanned aerial vehicles (UAVs) offer high-resolution data and flexible deployment, making them ideal for such tasks. By leveraging UAV technology, Kevin aimed to provide a more dynamic and detailed analysis of wildfire risk areas, thereby facilitating better-prepared firefighting and prevention strategies.

Our platform comprises two key components:

  1. Geospatial Machine Learning Model: At the heart of our system is a sophisticated machine learning model that processes inputs from various sensors. This model, built using a convolutional neural network (CNN) architecture, excels at extracting spatial features from the data. By integrating information such as vegetation burnability indices, topography, and land surface temperature, the model can predict wildfire risk with remarkable accuracy.
  2. Custom Drone Platform: To support our machine learning model, Kevin designed a custom hexacopter equipped with a suite of high-resolution sensors. These include thermal and visible spectrum cameras and LiDAR scanners. The drone captures detailed environmental data, which is then processed in real-time using an onboard AI processor. This integration allows for rapid and precise risk assessment across diverse geographical areas.

The journey of creating a predictive model involved meticulous data collection and analysis. Using Google Earth Engine, Kevin assembled a comprehensive geospatial dataset, incorporating historical wildfire data and environmental variables. Kevin’s model underwent rigorous training and validation, demonstrating exceptional performance:

  • First Model: Focused on vegetation burnability indices, this model achieved an accuracy of 100% in classifying wildfire risk areas. It exhibited perfect precision, F1, and recall values, indicating its capability to correctly identify all instances without false positives or negatives.
  • Second Model: Enhanced with additional features such as topography and land surface temperature, this iteration maintained a high accuracy of 99.71%. Despite a slight dip, likely due to the stochastic nature of the test dataset, the model continued to generalize effectively across different geographical contexts.

Both models were tested on independently acquired datasets, confirming their robustness and reliability. The close alignment of training and validation values underscored the models’ ability to avoid overfitting, ensuring consistent performance on new, unseen data.

Kevin’s drone platform represents a significant advancement in wildfire risk prediction technology. By combining high-resolution aerial surveys with powerful machine learning algorithms, we can deliver precise and timely risk assessments. This innovation not only enhances the accuracy of predictions but also supports proactive firefighting measures, ultimately contributing to the preservation of natural habitats and human communities.