fire-prediction
Team project (fork). Predicting Bay Area wildfires from NASA fire data and weather data.
Overview
A model that predicts wildfire occurrence in the Bay Area. It joins NASA FIRMS fire records, which have confidence scores and timestamps, with historical weather from the Open-Meteo API. The first version is a binary classifier on 5 to 6 weather parameters over one county. A second approach uses regression to predict fire severity.
My part
My commits built the train, test and validation data for the first version, compared five classifiers and wrote the first version of the README, and added an SVM model with PCA on the weather data.
Findings
- Logistic regression and an SGD classifier did not recognize fire cases well.
- Random forest was more accurate but still missed fire cases.
- AdaBoost had a much higher recall and a higher F1 score, which matters on this imbalanced dataset.
- XGBoost had better F1 and recall than AdaBoost.
- The team also trained TabNet, an attention-based network for tabular data.
- The San Jose fire dataset (2018 to 2023) was dropped: it covers vegetation fires, not forest fires.