<![CDATA[Temporal estimation, specifically predicting the exact continuous duration a natural disaster will last from day one, remains a major operational bottleneck in global emergency management, despite rapid advancements in spatial mapping. This study aims to evaluate and compare the predictive performance of two structurally distinct non-linear machine learning algorithms: the Gradient Boosting Regressor and Support Vector Regression (SVR) utilizing a Radial Basis Function (RBF) kernel. The methodology employs a non-linear quantitative predictive pipeline executed via Python, utilizing empirical historical crisis records extracted from the Global Disaster Alert and Coordination System (GDACS) repository. The final feature matrix integrates localized country profiles, hazard typologies, dynamic environmental severity scales, and monthly seasonal components. The empirical results reveal a sharp contrast in computational performance between the two architectures. The sequential tree-based Gradient Boosting framework demonstrated exceptional estimation capacity, achieving an outstanding Coefficient of Determination (R2) score of 0.9818, with a minimal Mean Absolute Error (MAE) of only 4.1045 days and a robust Root Mean Squared Error (RMSE) of 6.1378 days. Conversely, the distance-based SVR architecture experienced a structural collapse, yielding an R2 score of only 0.1587 and a severely inflated RMSE of 41.7480 days, exposing its critical vulnerability and disorientation when handling heavily skewed target distributions and heavy-tailed temporal outliers. These findings underscore that embedding high-performance ensemble tree structures into digital Decision Support Systems (DSS) can provide highly precise emergency timeline projections from day one, effectively enabling a proactive, data-driven orchestration of humanitarian logistics.]]>
Copyrights © 2026
