Despite the large availability of satellite rainfall data and their increasing spatial/temporal coverage and accuracy, their use for hydrological applications is still very scarse. The reasons for this could rely on: 1) the coarse spatial/temporal resolution of the provided satellite products; 2) the inherent BIAS contained in every estimate; 3) the latency own by the product before been collected by the end-users (Serrat-Capdevilla et al., 2014); 4) a general and unjustified skepticism regarding satellite data in the hydrologist community.
If one makes a research on the literature available on the use of satellite rainfall data for hydrological applications, the result would be that the majority of the studies are carried out at the basin scale for flood simulations, mainly in developing countries. On the other hand, just few studies provided some guidelines in order to use satellite rainfall products for hydrological applications. For instance, Artan et al. (2007) used the NOAA Climate Prediction Center (CPC) product for Famine Early Warning System (FEWS, Xie and Arkin, 1997) to drive a physically- based semi-distributed hydrologic model over four basins in Asia and Africa. They found that satellite rainfall products (SRPs) can be used to force a hydrologic model provided that the recalibration of the model parameter values is carried out. Harris et al. (2007) used TMPA 3B42 real-time product to drive a hydrologic model over a catchment in Kentucky finding that a BIAS correction is needed before using real-time satellite data in flood forecasting. Stisen and Sandholt (2010) forced a distributed hydrologic model over the Senegal River Basin with different SRPs: TMPA 3B42 V6, CPC MORPHing technique (CMORPH, Joyce et al., 2004), CPC FEWS v.2, PERSIANN and a local product based on CCD relationship. They found that the SRPs need a BIAS correction because of the differences in the estimates of the analyzed products (e.g., the number of rainy days and the recorded intensity). Camici et al., 2018 stated that BIAS correction and specific model recalibration are mandatory steps, even if not always sufficient to achieve good performance, mainly in small basins. They also stated that a simple integration between state-of-the-art and SM- derived products allows to obtained results even better than those obtained by using observed data. In this respect, Massari et al. (2014) showed that a simple integration between ground observed data and satellite rainfall provided better performance in terms of flood simulation for 3 out 4 basins with respect to use ground rainfall data only.
With respect to landslide risk assessment, just a handful of studies uses satellite rainfall data. Hong et al. (2006; 2007) set up an experimental monitoring system for rainfall-induced landslides using TMPA 3B42v6 product. The precipitation data were linked to a landslide susceptibility map in order to identify the location and timing of the earth movement and the associated risk. The model was applied to 74 landslide events during the period 1998-2006 with satisfactorily results. Although satellite rainfall data have their own limitations, the authors stated that “this may form a starting point for developing an operational early warning system for rainfall-induced landslides around the globe”.
35 estimates to force a land-surface model in order to assess the stability conditions over 4 areas throughout the globe. They stated that satellite rainfall data can be a valuable source of information and a valid input for such applications. More in details, Liao et al., (2010; 2012) developed an experimental early warning system based on several satellite information. The model framework is reported in Figure 1.6.
Figure 1.6 – Experimental early warning system for rainfall-induced landslides set up in Indonesia (Courtesy of Liao et al., 2010).
As it can be seen in Figure 1.6, satellite rainfall estimates are used to force the Rain-SLIDE model. The model’s outputs are then linked to susceptibility maps in order to predict landslide events at the site scale. The authors drawn the conclusions that despite the limitations of each early warning system and the satellite rainfall estimates, the proposed framework was able to identify real events occurred in the study area, enabling researchers to develop operational early warning systems at the regional scale. Posner and Georgakakos (2015) instead defined a rainfall-soil moisture threshold for El Salvador by forcing a hydrologic model with the Global Hydro-Estimator (GHE, Scofield and Kuligowsky, 2003). The satellite data are used to simulate the soil moisture conditions associated with occurred landslide events, underlining the beneficial effect of coupling hydrologic modelling to landslide early warning systems.
A different approach was used by Farahmand and AghaKouchak (2013), who used PERSIANN rainfall estimates and 581 landslide events (Kirschbaum et al., 2010) to train a Support Vector Machine (SVM). The SVM is used to classify, through a binary classifier, between landslide events (value of 1) and non-landslides events (value of 0), by considering rainfall, slope, land cover and land use and observed landslide events. The model provided very good results, with just 2% of false
36 alarms and 7% of missed events, but needed to be calibrated and trained with specific data.
Up to now, one of the only operational satellite-based early warning system is represented by the global Landslide Hazard Assessment for Situational Awareness (LHASA, Kirschbaum et al., 2011;
2015; 2016; Stanley and Kirschbaum, 2017; Kirschbaum and Stanley, 2018;
https://pmm.nasa.gov/applications/global-landslide-model). The framework of the model is reported in Figure 1.7. The model is based on TMPA (and now GPM) rainfall real-time estimates. The model computes the Antecedent Rainfall Index based on the following equation:
n t t n t t t w p w ARI 1 1 (1.4)where t is the number of days before the running date, n is the length of the time window use for the ARI estimation, pt is the rainfall amount at day t and wt is defined by wt=t-α. In Stanley and Kirschbaum (2017) n was set to 7 and α to 0.5 after calibration. One of the most important component of this model is a susceptibility map. The map was created starting from combining information from elevation, geology, roads and infrastructures and forest cover. The map considers 5 different levels of susceptibility (very low, low, moderate, high and very high). If the 7-day precipitation is unusually high, the susceptibility map is used to discriminate if the landslide risk of the analyzed pixel is low, moderate or high.
Figure 1.7 – Global Landslide Hazard Assessment for Situational Awareness (LHASA) model framework. (https://pmm.nasa.gov/applications/global-landslide-mode )
37 Very recently, Rossi et al., (2017) defined rainfall thresholds over Umbria region by using observed and satellite rainfall data obtained through TMPA products. The authors stated that the general underestimation of precipitation provided by the remotely sensed estimates is reflected in lower rainfall thresholds with respect to the ones obtained by considering rain gauge data. One the most recent application of satellite rainfall data to landslide risk assessment is the one proposed by
Brunetti et al. (2018). In this work, the authors evaluated the capabilities of TMPA 3B42RT, CMORPH, PERSIANN real-time data and SM-derived rainfall estimates in terms of predicting the spatial-temporal occurrence of landslide events over the Italian territory during the period 2008- 2014. Specifically, for all the analyzed products, an empirical cumulated rainfall – rainfall duration (ED) threshold is estimated by analyzing more than 1400 landslide events. The ED thresholds were then used to estimate several skill scores for the assessment of the products. The results showed that satellite-based rainfall estimates generally underestimated precipitation amount with respect to the ground-based dataset used for comparison. However, this aspect is not an issue for the development of an early warning system, as it could reflect in a lower ED threshold.
On this basis, it could be stated that satellite rainfall products are able to be used for such applications, taking advantages of the short latency of the real-time products. More details about this application can be found in Appendix 1.