Project Overview
This project analyzes global natural disaster observations from the
NASA EONET API, with a focus on category imbalance, temporal trends,
seasonality, spatial concentration, and predictive benchmarking.
Main Research Question
How do global natural disaster observations in NASA EONET vary across
category, time, seasonality, and space, and to what extent can these
patterns support simple out-of-sample prediction?
Key Findings
- Wildfires dominates the dataset with 24,208 records (89.1%).
- Recorded observations peak in 2024 at 16,816, with a sharp post-2023
rise.
- Seasonal structure is strong and category-dependent rather than
uniform across hazards.
- In full-scope benchmarking, the best RMSE is from GAM (324.258),
while in wildfire-only benchmarking the best RMSE shifts to OLS
(571.294).
Report Summary
The written report is structured as Introduction, Methods, Results,
and Conclusions, and answers explicit research questions with
quantitative evidence.
Research question addressed in the report
- Main question: How do NASA EONET disaster observations vary by
category, time, seasonality, and space, and how does predictive
performance change between two targets (full multi-category counts vs
wildfire-only counts)?
- Supporting questions include: category dominance, temporal trend
shape, category-specific seasonality, and out-of-sample benchmark
performance.
Quantitative findings from the report
- Category imbalance is extreme: Wildfires accounts for 89.1% of all
observations.
- Temporal counts peak in 2024 at 16,816; the surge is primarily
wildfire-driven.
- Scope comparison changes model ranking: full scope best RMSE =
324.258 (GAM), wildfire-only best RMSE = 571.294 (OLS).
- Even with partial RMSE gains, out-of-sample R2 remains weak
(negative in the report benchmarks), so predictive results are
interpreted cautiously.
Featured Interactive Visualizations