A field-based thickness measurement dataset of fallout pyroclastic deposits in the peri-volcanic areas of Campania (Italy): statistical combination of different predictions for spatial estimation of thickness

Abstract

Determining the spatial thickness (z) of in situ and reworked fallout pyroclastic deposits plays a key role in volcanological studies and in shedding light on geomorphological and hydrogeological processes in peri-volcanic areas. However, this is a challenging line of research because (1) field-based measurements are expensive and time-consuming, (2) the ash might have been dispersed in the atmosphere by several volcanic eruptions, and (3) wind characteristics during an eruptive event and soil-forming and/or denudation processes after ash deposition on the ground surface affect the expected spatial distribution of these deposits. This article tries to bridge this knowledge gap by applying statistical techniques for making representative spatial thickness predictions to be used for the analysis of geomorphic processes at the catchment and sub-catchment scales. First, we compiled a field-based thickness measurement dataset (https://doi.org/10.5281/zenodo.8399487; Matano et al., 2023) of fallout pyroclastic deposits in the territories of several municipalities in Campania, southern Italy. Second, 18 predictor variables were derived mainly from digital elevation models and satellite images and were assigned to each measurement point. Third, the stepwise regression (STPW) model and random forest (RF) machine learning technique are used for thickness modeling. Fourth, the estimations are compared with those of three models that already exist in the literature. Finally, the statistical combination of different predictions is implemented to develop a less biased model for estimating pyroclastic thickness. The results show that the prediction accuracy of RF (RMSE <82.46 and MAE <48.36) is better than that of existing models in the literature. Moreover, statistical combination of the predictions obtained from the above-mentioned models through a least absolute deviation (LAD) combination approach leads to the most representative thickness estimation (MAE <45.12) in the study area. The maps for the values estimated by RF and LAD (as the best single model and combination approach, respectively) illustrate that the spatial patterns did not change significantly, but the estimations by LAD are smaller. This combined approach can help in estimating the thickness of fallout pyroclastic deposits in other volcanic regions and in managing geohazards in areas covered with loose pyroclastic materials.