Mapping Recovery After Hurricane Katrina, 2004-2025: A Multi-Temporal Machine Learning and Trajectory-Based Analysis of Greater New Orleans

Abstract Hurricane Katrina (2005) and the subsequent levee failures caused one of the largest urban disasters in U.S. history. However, the spatial extent, timing, and environmental consequences of rebuilding across the Greater New Orleans region remain poorly quantified at the metropolitan scale. This study maps land cover for seven epochs (2004–2025) using Landsat 5, Landsat 8, and Sentinel-2 imagery. Four machine learning classifiers (Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine, and K-Nearest Neighbors) were trained on a 20-feature spectral stack and evaluated using spatial-split cross-validation, yielding overall accuracies of over 90%. XGBoost performed comparably to the other classifiers and was used to produce the final maps. SHAP analysis showed that the spectral features driving classification shifted across epochs, with the blue band dominating in 2005 (consistent with post-flood standing water). Pixels were grouped into seven change trajectories. The largest, “Decline then Recovery” (built-up in 2004, non-built-up after the storm, and built-up again by 2013 or later), covered 97.88 km², compared with 1.96 km² for “Built-up Loss” (built-up before the storm and never recovered). Global Moran’s I showed significant spatial clustering across all trajectories (I = 0.14–0.55, p < 0.001). Local Indicators of Spatial Association (LISA) identified 39 tracts with high recovery shares surrounded by similar tracts, and 62 tracts with low recovery shares surrounded by similar tracts (mainly in the Lower Ninth Ward and New Orleans East). Overall, post-Katrina recovery has been extensive but spatially uneven, with some neighborhoods rebuilt and others still lagging.