Wildfire Analysis & Prediction for California

Course: IS-517: Methods of Data Science Team: Apurva Malpure · Prithvish Taukari · Sai Kiran Billa · Yashvi Bhatt

---

Overview

Wildfires in California are a growing threat to life, property, and the environment. This project applies machine learning to two primary California wildfire datasets to:

1. Predict daily wildfire ignition risk from weather and seasonal signals.
2. Classify unknown fire causes (≈31 % of records) as human-made or natural.
3. Predict fire duration using weather and temporal features.
4. Identify geographic wildfire hotspots and risk zones across the state.

---

Motivation

More than 31 % of CALFIRE incident records carry unknown or undocumented causes. Environmental factors such as temperature, wind speed, precipitation, and season interact non-linearly, making rule-based approaches inadequate. By combining supervised classification, regression, and spatial clustering we aim to turn raw incident logs into actionable risk intelligence.

---

Required Datasets

Three datasets are required to reproduce this project. They are not tracked in this repository due to file size. Place each file in the data/ folder before running any script.

File	Required by	Where to obtain
fires_with_max_weather_corrected.csv	All RQ2 scripts, RQ3, RQ4, Rmd	Not publicly available. Team-engineered from CALFIRE perimeters + weather APIs. Contact the course instructor or project team to obtain this file.
CA_Weather_Fire_Dataset_1984-2025.csv	research-question-4.R, Rmd (RQ1)	Publicly available on Zenodo. https://zenodo.org/records/14712845
CA_Weather_Fire_Dataset_2008_2023_with_duration.csv	Rmd (RQ3 SVM duration analysis)	Not publicly available. Team-engineered from CALFIRE perimeters with duration calculations applied. Contact the course instructor or project team.

Engineering steps already applied to fires_with_max_weather_corrected.csv:

• FIRE_DURATION calculated from alarm and containment dates.
• Latitude / Longitude extracted from polygon centroids.
• Temperature and wind data integrated via external APIs.

---

Pre-generated Outputs (Committed to Repository)

The following files are the outputs of slow training runs and are committed so the report can be re-knitted without re-running the full training pipeline.

File	Generated by	Purpose
models/5_fold_rf.rds	Method1_Final_model.R	RQ2 Method 1 — trained Random Forest
models/xgb_model.rds	Method1_Final_model.R	RQ2 Method 1 — trained XGBoost
models/svm_model_weight1.5.rds	Method2_Complete.R	RQ2 Method 2 — trained SVM (RBF, weight 1.5)
data/cause14_dual_model_predictions_5fold.csv	Method1_Final_model.R	RQ2 Method 1 — per-record raw predictions
data/cause14_final_ensemble_classification.csv	Method1_Final_model.R	RQ2 Method 1 — final ensemble classifications
data/cause14_binary_predictions_with_confidence.csv	Method2_Complete.R	RQ2 Method 2 — binary predictions with probabilities

---

Repository Structure

project-root/
│
├── data/                                          # Input datasets + intermediate outputs
│   ├── fires_with_max_weather_corrected.csv       # ⚠ not tracked — add manually
│   ├── CA_Weather_Fire_Dataset_1984-2025.csv      # ⚠ not tracked — add manually
│   ├── CA_Weather_Fire_Dataset_2008_2023_with_duration.csv  # ⚠ not tracked — add manually
│   ├── cause14_dual_model_predictions_5fold.csv   # pre-generated, committed
│   ├── cause14_final_ensemble_classification.csv  # pre-generated, committed
│   └── cause14_binary_predictions_with_confidence.csv       # pre-generated, committed
│
├── src/                                           # All analysis scripts
│   ├── Seasonal_trends.R                          # Monthly / seasonal trend plots
│   ├── NaturalvsHumanMade_Piechart.R              # RQ2 cause-type pie chart
│   ├── Pie_Chart_Classification_by_class.R        # RQ2 cause distribution donut
│   ├── Method1_Final_model.R                      # RQ2 Method 1 — RF + XGBoost ensemble
│   ├── Method1_Pie_Chart.R                        # RQ2 Method 1 — classification outcome pie
│   ├── Method2_Complete.R                         # RQ2 Method 2 — binary SVM
│   ├── Method-2_Plots.R                           # RQ2 Method 2 — additional SVM plots
│   ├── research-question-4.R                      # RQ1 — weather-based RF fire-day predictor
│   ├── Fire_duration_Plot.R                       # RQ3 — fire duration distribution plot
│   └── Spearman_Plot.R                            # EDA — Spearman correlation heatmap
│
├── models/                                        # Saved model objects (.rds)
│   ├── 5_fold_rf.rds
│   ├── xgb_model.rds
│   └── svm_model_weight1.5.rds
│
├── figures/                                       # Static images referenced by the report
│   ├── figure1_spearman.png
│   ├── figure2_Seasonal.png
│   ├── figure3_duration.png
│   ├── figure4_cause_pie.png
│   ├── figure5_flowchart.png
│   ├── figure6_method1_pie.png
│   ├── figure7_natural_human.png
│   ├── figure8_weight_accuracy.png
│   ├── figure9_auc_roc.png
│   ├── figure10_svm_grouping.png
│   ├── figure11_new_seasonal.png
│   ├── figure12_svm_example.png
│   ├── RQ3_methodology.png
│   └── Smart_flowchart.png
│
├── reports/                                       # R Markdown report and supporting files
│   ├── Group4_Project_report_Final.Rmd
│   └── float-setup.tex                            # LaTeX header for figure placement
│
├── packages.R                                     # Quick-setup entry point (also in src/)
├── .gitignore
└── README.md

---

Research Questions & Methods

RQ1 — Can weather variables predict wildfire occurrence?

Model: Tuned Random Forest (ranger) via caret with repeated 5-fold cross-validation and down-sampling to address class imbalance.

Features: MAX_TEMP, MIN_TEMP, PRECIPITATION, AVG_WIND_SPEED, TEMP_RANGE, WIND_TEMP_RATIO, lagged precipitation, lagged wind speed, SEASON, MONTH.

Split: Train 1984–2015 · Test 2016–2023.

Key findings: Temperature and month are the strongest predictors. The model generalises well on unseen data (2016–2023).

Figure 1 — Count of fire days by season.

---

RQ2 — Can unknown fire causes be classified?

Approximately 31 % of records carry cause code 14 (unknown). Two complementary methods:

Method 1 — Multiclass ensemble (RF + XGBoost)

Model	5-Fold CV Accuracy	Top Features
Random Forest	39.39 %	Max Wind Speed, Max Temp, Wind Speed, Area
XGBoost	41.24 %	Wind Speed, Duration, Temp, Area

Agreement between both models treated as a confirmed classification. ≈ 51.9 % of unknown records classified.

Figure 2 — Ensemble classification outcome for Cause-14 fires.

Method 2 — Binary SVM (RBF kernel)

Causes collapsed to Human-made vs. Natural before re-training.

Metric	Value
Accuracy	84.99 %
Sensitivity (Natural)	61.95 %
Specificity (Human-made)	94.20 %

  
Figure 3 — Cause breakdown  |  Figure 4 — SVM ROC Curve.

---

RQ3 — What weather factors drive fire duration?

Model: RBF-SVM with decision boundary visualisations across feature pairs.

Key findings:

• Longer fires are strongly associated with higher minimum temperatures.
• Peak fire durations cluster mid-year (May–July).
• Wind speed shows limited independent influence on duration once temperature is controlled for.

Figure 5 — Distribution of fire duration across the dataset.

---

RQ4 — Where are California's wildfire hotspots?

Methods: DBSCAN / HDBSCAN for geographic clustering; K-means (k = 4) for fire-behaviour clustering on duration × temperature × wind speed; composite risk score mapped geographically.

High-risk zones identified: Southern California · Central Valley · Sacramento.

Figure 6 — Spearman correlation matrix of numeric features.

---

Note on Reproducibility

This repository is substantially reproducible subject to the following conditions:

• Path updates required. All .R scripts contain hard-coded absolute paths (local to the original authors' machines) that must be changed to relative paths before running on any other machine. The path-replacement table is in the section below.

• Re-training not required for PDF rendering. Pre-generated model objects (5_fold_rf.rds, xgb_model.rds, svm_model_weight1.5.rds) and intermediate data files are committed to the repository so the Rmd can be re-knitted immediately after supplying the three raw datasets.

• Re-training is slow. Reproducing the RF (RQ1) and RF + XGBoost ensemble (RQ2) from scratch requires repeated 5-fold cross-validation and may take several hours on a standard laptop.

• RQ2 Method 1 pipeline. The full RF + XGBoost training logic is contained in src/Method1_Final_model.R. The same logic is embedded as {r} code chunks inside the R Markdown report. The standalone script is the authoritative source.

---

How to Run With the Current Files

# ── Step 1: Install TinyTeX for PDF rendering (run once, if not installed) ───
tinytex::install_tinytex()   # skip if you already have a LaTeX distribution

# ── Step 2: Install all R packages (run once) ────────────────────────────────
source("packages.R")

# ── Step 3: Place raw datasets in data/ ──────────────────────────────────────
# data/fires_with_max_weather_corrected.csv             (contact project team)
# data/CA_Weather_Fire_Dataset_1984-2025.csv            (download from Zenodo)
# data/CA_Weather_Fire_Dataset_2008_2023_with_duration.csv  (contact project team)

# ── Step 4 (optional): Re-run individual analysis scripts ────────────────────
# All scripts in src/ use paths relative to src/ as the working directory.
# Open each script in RStudio and run it from there, OR set your working
# directory to src/ before sourcing:
setwd("src")
source("Seasonal_trends.R")
source("NaturalvsHumanMade_Piechart.R")
source("Pie_Chart_Classification_by_class.R")
source("Method1_Final_model.R")    # slow — re-trains RF + XGBoost
source("Method2_Complete.R")       # re-trains SVM, writes prediction CSV
source("research-question-4.R")   # slow — re-trains tuned Ranger RF
setwd("..")                        # return to repo root when done

# ── Step 5: Render the full report ───────────────────────────────────────────
# Run from the repo root. knitr resolves all paths relative to reports/.
rmarkdown::render(
  "reports/Group4_Project_report_Final.Rmd",
  output_format = "pdf_document",
  output_dir    = "reports/"
)

---

Required Packages

Install everything in one step: source("packages.R")

Package	Purpose
ggplot2	All visualisations
dplyr · tidyr · tidyverse · readr	Data manipulation
lubridate	Date/time parsing
scales	Axis formatting
gridExtra · corrplot	Multi-panel plots & correlation matrix
maps	California state outline
sf · ggmap	Spatial data handling
caret	Model training, cross-validation, metrics
e1071	SVM (RBF kernel)
ranger · randomForest	Random Forest (fast training + importance)
xgboost	Gradient boosting (RQ2 Method 1 ensemble)
pROC	ROC / AUC computation
dbscan	DBSCAN / HDBSCAN geographic clustering
knitr · tinytex	Report rendering

---

Authors

Name	Role
Apurva Malpure	Data engineering, RQ2 multiclass classification (Method 1)
Prithvish Taukari	Visualisations, RQ2 binary SVM classification (Method 2)
Sai Kiran Billa	Weather-based fire prediction, Random Forest (RQ1)
Yashvi Bhatt	Hotspot identification, DBSCAN / K-means clustering (RQ4)

IS-517: Methods of Data Science