Spatio-Temporal Agnostic Sampling for Imbalanced Multivariate Seasonal Time Series Data: A Study on Forest Fires

Area of Science:

• Environmental Science
• Data Science
• Geospatial Analysis

Background:

• Natural disasters, including forest fires, pose significant threats due to seasonal and anthropogenic factors.
• Forest fires are increasing in frequency and destruction, impacting ecosystems and economies.
• Predictive modeling for forest fires is challenged by highly imbalanced datasets (over 100,000 non-fire events per fire event).

Purpose of the Study:

• To introduce a novel data sampling technique, Spatio-Temporal Agnostic Sampling (STAS), for handling imbalanced time-series data in forest fire prediction.
• To provide a mathematical framework and complexity analysis for STAS, comparing it with existing methods like NearMiss and SMOTE.
• To evaluate the effectiveness of STAS in improving forest fire probability classification and severity assessment models.

Main Methods:

• Development and mathematical formulation of the Spatio-Temporal Agnostic Sampling (STAS) framework.
• Complexity analysis comparing STAS against NearMiss and SMOTE.
• Implementation of binary classification and regression models using STAS-generated data for fire prediction and severity assessment.
• Extensive validation through 432 experiments and additional temporal data split analysis.

Main Results:

• STAS demonstrates superior performance in handling imbalanced multivariate seasonal time-series data.
• Binary classification models built on STAS data achieved an F1-score > 0.9 in 180 out of 216 experiments.
• Regression models for fire severity assessment achieved an R2-score > 0.75 in 150 out of 216 experiments.

Conclusions:

• The Spatio-Temporal Agnostic Sampling (STAS) framework is highly effective for improving forest fire prediction models.
• STAS successfully addresses the challenge of highly imbalanced data in seasonal, multivariate time-series datasets.
• The validated performance of STAS indicates its significant applicability for real-world forest fire early warning systems.