 |
| Titel |
How badly are we doing? Estimating misclassification rates of shallow landslide susceptibility maps |
| VerfasserIn |
Andreas Papritz, Jonas von Ruette, Peter Lehmann, Christian Rickli, Dani Or |
| Konferenz |
EGU General Assembly 2010
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 12 (2010) |
| Datensatznummer |
250038441
|
|
|
|
|
|
| Zusammenfassung |
| An important motivation for continuing efforts in landslide susceptibility mapping is the need
for reliable maps of landslide-prone areas. “Reliable” means that a map should not
systematically under- or overpredict landslide incidence and provide a fair measure of its
predictive power. Probabilistic susceptibility maps may be generated by various
statistical methods (logistic regression, neural networks, classification trees, etc.). These
methods must be “trained” with data on past landslide occurrence and information
about conditioning features (terrain, geology, land use, vegetation, soil, -¦). Once
trained, most approaches fit observed landslide incidences in training areas reasonably
well. If probabilistic predictions are thresholded, the error rate (total percentage
of misclassifications), the true positive rate (sensitivity) and false positive rate
(1- specificity) Â —Â which form the receiver operating characteristics curve (ROC)
when plotted against each other for several thresholds — provide, seemingly, a
favourable picture of the predictive power of the methods. However, these apparent
misclassification rates give too small estimates of the true rates. Remedy for bias is
cross-validation, which provides nearly unbiased estimates, but suffers from large
random variation. The large variance of cross-validation estimates can be mitigated by
bootstrapping. Efron and Tibshirani [1] proposed the .632+ bootstrap for estimating
the true error rate. Adler and Lausen [2] extended the method for estimating ROC
curves.
We use the .632+ bootstrap to estimate misclassification rates (ROC curve, bias score [3])
of landslide susceptibility maps, generated by logistic regression and a random forest
classifier. The methods are trained with data on incidence of shallow landslides, released in
a pre-alpine catchment in Switzerland during a heavy rainfall in summer 2005.
Geomorphological terrain attributes and information on land use are used as conditioning
features. Two event-based landslides surveys in another two catchments in the same region
are used as independent test data to estimate true misclassification rates. These
estimates will be compared with the apparent, the cross-validation and the .632+
bootstrap estimates of misclassification rates computed from the training data only. This
comparison will reveal whether bootstrapping offers some advantage over cross-validation
for obtaining honest estimates of the predictive power of landslide susceptibility
maps.
References
[1]Â Â Â Efron, B. and Tibshirani, R. (1997). Improvement on cross-validation: The
.632+ bootstrap method. Journal of the American Statistical Association, 92,
548–560.
[2]Â Â Â Adler, W. and Lausen, B. (2009). Bootstrap estimated true and false positive
rates and ROC curve. Computational Statistics and Data Analysis, 53, 718–729.
[3]   Wilks, D. S. (2006). Statistical Methods in the Atmospheric Sciences.
Academic Press, second edition. |
|
|
|
|
|
|