A Warning System for Rainfall-Induced Debris Flows: A Integrated Remote Sensing and Data Mining Approach

Monday, 15 December 2014
Racha Elkadiri1, Mohamed Sultan1, Ilyas Nurmemet2, Hassan Al Harbi3, Ahmed Youssef3, Tamer Elbayoumi4, Yesser Zabramwi3, Saeed Alzahrani3 and Alla Bahamil3, (1)Western Michigan University, Kalamazoo, MI, United States, (2)Xinjiang University, Resources and Environmental Science, Urumqi, China, (3)Saudi Geological Survey, Jeddah, Saudi Arabia, (4)Oklahoma State University, Stillwater, OK, United States
We developed methodologies that heavily rely on observations extracted from a wide-range of remote sensing data sets (TRMM, Landsat ETM, ENVISAT, ERS, SPOT, Orbview, GeoEye) to develop a warning system for rainfall-induced debris flows in the Jazan province in the Red Sea Hills. The developed warning system integrates static controlling factors and dynamic triggering factors. The algorithm couples a susceptibility map with a rainfall I-D curve, both are developed using readily available remote sensing datasets. The static susceptibility map was constructed as follows: (1) an inventory was compiled for debris flows identified from high spatial resolution datasets and field verified; (2) 10 topographical and land cover predisposing factors (i.e. slope angle, slope aspect, normalized difference vegetation index, topographical position index, stream power index, flow accumulation, distance to drainage line, soil weathering index, elevation and topographic wetness index) were generated; (3) an artificial neural network model (ANN) was constructed, optimized and validated; (4) a debris-flow susceptibility map was generated using the ANN model and refined (using differential backscatter coefficient radar images). The rainfall threshold curve was derived as follows: (1) a spatial database was generated to host temporal co-registered and radiometrically and atmospherically corrected Landsat images; (2) temporal change detection images were generated for pairs of successively acquired Landsat images and criteria were established to identify “the change” related to debris flows, (3) the duration and intensity of the precipitation event that caused each of the identified debris flow events was assumed to be that of the most intense event within the investigated period; and (4) the I-D curve was extracted using data (intensity and duration of precipitation) for the inventoried events. Our findings include: (1) the spatial controlling factors with the highest predictive power of debris-flow locations are: topographic position index, slope, NDVI and distance to drainage line; (2) the ANN model showed an excellent prediction performance (area under receiver operating characteristic [ROC] curve: 0.961); 3) the preliminary I-D curve is I=39.797×D-0.7355 (I: Intensity and D: duration).