Implications of classification of methodological decisions in flooding analysis from Hurricane Katrina

Abstract

Recent climatic patterns indicate that extreme weather events will increase in frequency and magnitude. Remote sensing offers unique advantages for large-scale monitoring. In this research, Landsat 5 remotely sensed imagery was used to assess flooding caused by Hurricane Katrina, one of the worst natural disasters in the US over the past decades. The objective of our work is to assess whether decisions associated with the classification process, such as location of reference data and algorithm choice, affected flooding results and subsequent analysis using census data. Maximum Likelihood (ML) and Back Propagation Neural Network (NN) were the tested algorithms, the former reflecting a simple and popular classifier, and the latter an advanced but complex method. Flooding estimations were almost identical within the reference sample area, 124.4 km2 for the ML classifier and 123.7 km2 for the NN classifier. However, large discrepancies were found outside the reference sample area with the ML predicting 462.5 km2 and the NN identifying 797.2 km2 as flooded, almost twice the amount. Further investigation took place to evaluate the influence of the classification method to a social study, namely the racial characteristics of flooded areas. Using Census 2000 data, our study area was segmented in census tracts. Results indicated a strong positive correlation between concentration of African Americans and proportional residential flooding. Pairwise T-Tests also verified that flooding among different African American concentrations was statistically different. There were no significant differences between the ML and NN methods in the results interpretation, which is mostly attributed to the significant geographic overlap between reference sample area and the examined census tracts. This study suggests that emergency responders should exercise significant caution in their decision making when using classification products from undersampled geographic areas in terms of classification reference data. © 2012 by the authors.