14
P. Muñoz et al.
(overparameterization and intense computational requirements) and the predictive
capability for extreme flows (Brath et al. 2004; Chang and Hwang 1999; Dawson
and Wilby 2001; Galelli and Castelletti 2013; Gupta et al. 2008; Willems 2014).
An arising alternative is the use of Machine Learning (ML) techniques for predictive modeling. ML popularity has recently and remarkably increased mainly due to
its flexibility and scalability properties in pattern extraction (Bontempi et al. 2012).
ML models are characterized by a more compact representation and high predictive
potential, with considerably fewer parameters to calibrate when compared to fully
distributed models (Galelli and Castelletti 2013; Mosavi et al. 2018).
Moreover, since there is no assumption on a global function describing the data,
ML techniques are particular relevant for problems of non-stationarity, missing
features, and measurement errors (Bontempi et al. 2012). Several ML methods can
be used for flood forecasting: artificial neural networks (ANNs (Kim et al. 2016)),
support vector machines (SVMs) (Martens et al. 2007), and decision trees (DTs)
(Kubal et al. 2009; Wang et al. 2015) among others. These methods exhibit some
weaknesses such as overfitting for ANNs (Jin et al. 2005), the complexity of mathematical functions for SVMs (Martens et al. 2007) and the considerable effort needed
for pre-processing data for DTs (Kubal et al. 2009). Generally, there is a global
tendency for using the Random Forest (RF) algorithm due to its simplicity, robustness
and capacity to cope with complex data structures (Kühnlein et al. 2014).
The objective of this study is to construct flood and drought forecasting models of
varying time duration (4, 8, 12 and 24 h). All modes are based on the Random Forest
(RF) algorithm. We selected two catchments, the Tomebamba and the Yanuncay
catchments, to be representative of the Tropical Andes in Ecuador.
2.2 Study Sites and Dataset
The first Tropical Andean catchment, the Tomebamba catchment, is delineated
upstream the Matadero-Sayausí station, in the Tomebamba river. It is located in
the south-eastern flank of the Andes, and the outlet is positioned 10 km away from
the city of Cuenca. The Tomebamba catchment has a drainage area of 300 km
2
,
and its elevation ranges between 2800 and 4100 m above the sea level (m asl). The
Tomebamba catchment is part of the Cajas National Park, declared by UNESCO as
a World Biosphere Reserve in 2013. The second catchment is allocated right next to
the first one, in the Yanuncay river at the Yanuncay A.J. Tarqui station. The Yanuncay
catchment has an area of 420 km
2 , spanning from 2480 to 4280 m asl. The outlet of
the catchment is located at the entrance of the city of Cuenca. Both the Tomebamba
and Yanuncay catchments drain to the Amazon river toward the Atlantic ocean.
Data comprise rainfall and discharge hourly timeseries for the periods Jan/2015 to
Sep/2018 and Jan/2015 to Jul/2017 for the Tomebamba and Yanuncay catchments,
respectively. From Jul/2017 on, measurements of precipitation and discharge at the
Yanuncay catchment are not reliable, and thus limited the use of a similar data
period for both catchments. However, this issue will not compromise the analyses
P. Muñoz et al.
(overparameterization and intense computational requirements) and the predictive
capability for extreme flows (Brath et al. 2004; Chang and Hwang 1999; Dawson
and Wilby 2001; Galelli and Castelletti 2013; Gupta et al. 2008; Willems 2014).
An arising alternative is the use of Machine Learning (ML) techniques for predictive modeling. ML popularity has recently and remarkably increased mainly due to
its flexibility and scalability properties in pattern extraction (Bontempi et al. 2012).
ML models are characterized by a more compact representation and high predictive
potential, with considerably fewer parameters to calibrate when compared to fully
distributed models (Galelli and Castelletti 2013; Mosavi et al. 2018).
Moreover, since there is no assumption on a global function describing the data,
ML techniques are particular relevant for problems of non-stationarity, missing
features, and measurement errors (Bontempi et al. 2012). Several ML methods can
be used for flood forecasting: artificial neural networks (ANNs (Kim et al. 2016)),
support vector machines (SVMs) (Martens et al. 2007), and decision trees (DTs)
(Kubal et al. 2009; Wang et al. 2015) among others. These methods exhibit some
weaknesses such as overfitting for ANNs (Jin et al. 2005), the complexity of mathematical functions for SVMs (Martens et al. 2007) and the considerable effort needed
for pre-processing data for DTs (Kubal et al. 2009). Generally, there is a global
tendency for using the Random Forest (RF) algorithm due to its simplicity, robustness
and capacity to cope with complex data structures (Kühnlein et al. 2014).
The objective of this study is to construct flood and drought forecasting models of
varying time duration (4, 8, 12 and 24 h). All modes are based on the Random Forest
(RF) algorithm. We selected two catchments, the Tomebamba and the Yanuncay
catchments, to be representative of the Tropical Andes in Ecuador.
2.2 Study Sites and Dataset
The first Tropical Andean catchment, the Tomebamba catchment, is delineated
upstream the Matadero-Sayausí station, in the Tomebamba river. It is located in
the south-eastern flank of the Andes, and the outlet is positioned 10 km away from
the city of Cuenca. The Tomebamba catchment has a drainage area of 300 km
2
,
and its elevation ranges between 2800 and 4100 m above the sea level (m asl). The
Tomebamba catchment is part of the Cajas National Park, declared by UNESCO as
a World Biosphere Reserve in 2013. The second catchment is allocated right next to
the first one, in the Yanuncay river at the Yanuncay A.J. Tarqui station. The Yanuncay
catchment has an area of 420 km
2 , spanning from 2480 to 4280 m asl. The outlet of
the catchment is located at the entrance of the city of Cuenca. Both the Tomebamba
and Yanuncay catchments drain to the Amazon river toward the Atlantic ocean.
Data comprise rainfall and discharge hourly timeseries for the periods Jan/2015 to
Sep/2018 and Jan/2015 to Jul/2017 for the Tomebamba and Yanuncay catchments,
respectively. From Jul/2017 on, measurements of precipitation and discharge at the
Yanuncay catchment are not reliable, and thus limited the use of a similar data
period for both catchments. However, this issue will not compromise the analyses