Correction of Precipitation Records through Inverse Modeling in Watersheds of South-Central Chile

Abstract

Precipitation is the main input in the water balance of watersheds; therefore, correct estimates are necessary for water resources management and decision making. In south-central Chile there is a low density of precipitation gauges (~1 station/675 km²), most of which are located in low-altitude areas. The spatial distribution of precipitation is, therefore, not properly recorded. In this study an inverse modeling approach is used to estimate the extent to which precipitation amounts must be corrected. Using a lumped water balance model, a factor for correcting precipitation data is calculated for 41 watersheds located in south-central Chile. Then, based on a geo-statistical interpolation, a map for correcting the precipitation amounts is proposed and a validation of these corrections is achieved. The results show that in gently sloping areas, the precipitation records are more representative than in steep mountain areas. In addition, the higher the mountains, the less representative the precipitation records become.

1. Introduction

The climate is important as a regulating factor of the natural and human environment on both the population level and in terms of daily activities. In this regard, precipitation is among the climatic elements that most influence nature and its configuration [1]. Its temporal and spatial distribution determine cycles in agriculture and forestry, which are the main economic activities of south-central Chile, where both productivity and the development of the primary plant and animal species depend on water resources from mountainous areas.

Having an accurate understanding of the availability of water resources has become important and necessary. This variable depends to a large extent on the storage and regulation capacities of mountain watersheds and the climate variability that affects the area [2]. The precipitation regime in south-central Chile (lat. 34°–40° S) is associated mainly with the successive passage of frontal systems resulting from the seasonal shift of the South Pacific subtropical anticyclone (April to September) [3], which enter the continent from the Pacific Ocean. During warmer months, meanwhile, the anticyclone shifts a few degrees to the south, impeding the formation of precipitation associated with the passage of such frontal systems.

The Coastal and Andes mountain ranges generate strong precipitation gradients that can also vary among storms, locally modifying precipitation amounts recorded by precipitation gauge stations [4]. However, due to its low density, the precipitation gauge station network in Chile may prove insufficient for correctly characterizing spatial precipitation variability. For example, in south-central Chile, there is a precipitation gauge station density of approximately 1 precipitation gauge every 675 km², and most of these stations are concentrated in the Central Valley and along the coast. This density is very low in comparison to the recommendations of the World Meteorological Organization (WMO); according to the WMO [5], to have a good precipitation gauge network in terms of density, there must be at least one precipitation gauge every 250 km² in mountainous areas and one every 575 km² in interior plains [5,6]. In addition, the presence of mountainous reliefs complicates access and impedes the installation of precipitation gauge stations; therefore, there is a much lower density in areas such as the western slope of the Andes and the highlands of the Coastal Range (e.g., Nahuelbuta Range, lat. ~37°45′ S). Thus, in a watershed with areas that are mostly mountainous, precipitation may be underestimated since the increase in precipitation due to the orography may not correctly represented.

Precipitation variability in both spatial and temporal terms is influenced by atmospheric thermodynamics, which depend on topography and relief, introducing marked imbalances in the spatial distribution of precipitation [7]. Daly et al. [8] hold that the main source of uncertainty in the characterization of the spatial distribution of precipitation is related to mountain height and meteorological factors such as atmospheric stability and moisture availability, among others. Therefore, it proves necessary to estimate precipitation in watersheds with a complex topography such as those of south-central Chile—with the presence of the Coastal Range and the Andes—for appropriate hydrological management.

There are numerous simulation models in the literature that can be applied to determine total watershed input [2,9,10,11,12,13]. Water availability in a watershed can be quantified through hydrological models; however, in order to achieve this objective, it is necessary to compile information on its meteorology, streamflows and geomorphology. The objective of this study is to quantify the effect of the morphology of south-central Chile (Coastal Range–Central Valley–Andes) on the spatial distribution of precipitation and its relationship to annual precipitation records from the national precipitation gauge network in order to propose a correction alternative for the available records. To this end, a spatially distributed precipitation correction factor was determined through inverse modeling, allowing recorded precipitation amounts to be related to precipitation increases, comparing the net input and output volumes per watershed in south-central Chile.

2. Study Area

The study area includes 41 watersheds located in south-central Chile (lat. ~34.5° to 39.5° S, Figure 1). Of these, 15 are located in the Andes, 16 in the Central Valley, and 10 in the Coastal Range.

The study area is ~100,000 km² and includes the Coastal Range (with an average elevation of 1200 m.a.s.l.), the Andes (with altitudes greater than 4000 m.a.s.l.) and the Central Valley. This last area is a flat valley located between the two mountain ranges (see topography in Figure 1). The area has an average precipitation gauge density of 1 precipitation gauge every 675 km², which can be subdivided into 1 precipitation gauge every 540 km² in low zones (below 400 m.a.s.l.) and 1 precipitation gauge every 960 km² in mountainous areas (above 400 m.a.s.l.). In mountainous areas, the station density is a quarter of that recommended by the World Meteorological Organization [5].

3. Data

To carry out the study, an inverse hydrological modeling approach was used, which consists of deriving the precipitation inputs to a watershed from modeling and comparing the results of a model with streamflow records in the modeled watershed.

The modeling requires monthly series of precipitation, streamflow, air temperature and potential evapotranspiration values, in addition to the morphological characterization of each watershed, which was constructed with a digital terrain model made from 1-arc-second-resolution (30 m) ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) images.

Regarding the input data, monthly precipitation time series were collected from precipitation gauge stations managed by the National Water Directorate (Dirección General de Aguas, DGA) (see Figure 1). The inverse distance weighting (IDW) method was used to calculate the representative precipitation of each watershed. With respect to temperature records, there were no stations with continuous monthly series and the spatial distribution required for this study; therefore, monthly temperature series provided by University of Delaware (UD) Center for Climatic Research were used [14]. To estimate potential evapotranspiration (PE), the formula proposed by Thornthwaite [15] and UD temperature data series were used. The spatial distribution of these variables over each watershed was determined using Thiessen polygons. Regarding streamflows, monthly series were collected from streamflow gauging stations managed by the DGA, as indicated in Figure 1.

Due to the availability and quality of the input data, the analysis was carried out at a monthly time step over a minimum time window of 10 years.

4. Methods

4.1. Conceptual Model Description

The model used in this article is the conceptual water balance model presented in Muñoz [16] and Muñoz et al. [13] (Figure 2). The model simulates the precipitation-runoff and snowmelt-runoff processes separately. Because the pluvial regime predominates in the study area, only the precipitation-runoff module was used in this study, and it will be briefly described here.

The lumped model considers the watershed a double storage system consisting of subsurface (SS) and deep storage (US). SS represents the water stored in the unsaturated soil layer as soil moisture, while US is the water covering the saturated soil layer. The main input data for the model are precipitation (PM) and potential evapotranspiration (PET). The model output is the total runoff (ETOT) at the watershed outlet, which includes both base flow (ES) and surface runoff (EI). For a further description of the model’s characteristics, please refer to Muñoz [16] and Muñoz et al. [13]. In addition,

Table 1. Description of the model parameters, adjustment factors, and conceptual-physical range for the precipitation-runoff model.

Parameter	Description	Influence over	Range
Cmax	Maximum runoff coefficient when the sub-surface layer is saturated	EI	0.05–0.70
PLim (mm)	Precipitation limit over which PPD exists	PPD	0–200
D	Percentage of precipitation over PLim transformed into PPD	PPD	0–100
Hmax (mm)	Maximum capacity of the soil layer to retain water	Cmax and ER	200–500
PORC	Fraction of Hmax that defines the soil water content, restricting the evaporation processes	Hmax and ER	0–100
Ck	Subterranean runoff coefficient	ES	0.05–0.80
A	Precipitation data adjustment factor	PM	0.85–2.00
B	Potential evapotranspiration data adjustment factor	PET and ER	0.85–2.00

EI: direct runoff; PPD: direct deep percolation; ER: real evapotranspiration; ES: subterranean runoff; PM: precipitation; PET: potential evapotranspiration.

presents a brief description of the model parameters and their influence on the model.

4.2. Calibration and Calculation of the Precipitation Correction Factor

To calibrate the models implemented in each watershed, the Monte Carlo Analysis Toolbox (MCAT) and the identifiability analysis developed by Wagener and Kollat [17] were used. This analysis allows the estimation of a set of suitable parameters in order to obtain good-performing models [18] based on a predefined performance measure or objective function. To assess identifiability, MCAT operates by running repeated simulations with a randomly selected parameter set within a user-defined range. This range must be wide enough to cover any possible simulation combination and at the same time must be a valid conceptual range for the purposes of the modeling. Then, the parameter values are compared with the values obtained for the objective function for subsequent analyses.

The identifiability analysis consists of a graphic representation of a sensitivity analysis applied to the best 10% of the simulations and allows parameter ranges and processes that positively influence the model outputs (with respect to a given objective function) to be found, and thereby helps restrict the parameter range such that only good-performing models are kept [19,20]. In addition, it allows the uncertainty associated with predictions to be determined and reduced, improving the quality of the hydrological model and the reliability of the outputs [13].

In the present study, 10,000 simulations were carried out for each model (watershed), and the precipitation adjustment factor (A) was calculated by taking the 50th percentile (P₅₀) of the best 10% of the models according to the identifiability analysis and three objective functions. The root mean square error (RMSE), transformed root mean square error (TRMSE) and the runoff coefficient error (ROCE) functions, designed to calibrate a model as a function of high streamflows, low streamflows and long-term water balance closure, respectively, were used [21].

To ensure process and simulated-flow representativeness, it was verified that only behavioral models were considered within the best 10% models. Nash–Sutcliffe values higher than 0.6 were used as an acceptance criterion for behavioral models.

4.3. Objective Function Description

(i)

Root mean square error (RMSE) (Equation (1)): this indicator operates as a function of simple differentials of the simulated streamflows (Q_s,t) with respect to the observed streamflows (Q_o,t) and is focused on the high flows.

$$\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{m}}{\sum }_{\mathrm{j}=1}^{\mathrm{m}}{\left({\mathrm{Q}}_{\mathrm{s},\mathrm{t}}-{\mathrm{Q}}_{\mathrm{o},\mathrm{t}}\right)}^2}$$

(1)

(ii)

Transformed root mean square error (TRMSE) (Equations (2) and (3)): this function is equivalent to RMSE, with the difference that the simulated and observed values are first transformed through Box–Cox transformation, which results in low flows taking on greater importance than medium or high flows.

$$\mathrm{TRMSE}=\sqrt{\frac{1}{\mathrm{m}}{\sum }_{\mathrm{j}=1}^{\mathrm{m}}{\left({\mathrm{Z}}_{\mathrm{s},\mathrm{t}}-{\mathrm{Z}}_{\mathrm{o},\mathrm{t}}\right)}^2}$$

(2)

$$\mathrm{Z}=\frac{{\left(1+\mathrm{Q}\right)}^{0.3}-1}{0.3}$$

(3)

(iii)

Runoff coefficient error (ROCE) (Equation (4)): this indicator captures the overall water balance, as it combines the flows into just one descriptor of hydrological characteristics—the average annual runoff coefficient—defined as ($\overline{\mathrm{Q}}/\overline{\mathrm{P}}$). The differences in absolute value between the simulated $\frac{{\overline{\mathrm{Q}}}_{\mathrm{s}}}{\overline{\mathrm{P}}}$ and observed $\frac{{\overline{\mathrm{Q}}}_{\mathrm{o}}}{\overline{\mathrm{P}}}$ descriptors are compared. This descriptor calculates the volumetric differences between simulated and observed streamflows over a modeling period, and therefore is the most suitable function for the objective of this study.

$$\mathrm{ROCE}=\mathrm{abs}\left(\frac{{\overline{\mathrm{Q}}}_{\mathrm{s}}}{\overline{\mathrm{P}}}-\frac{{\overline{\mathrm{Q}}}_{\mathrm{o}}}{\overline{\mathrm{P}}}\right)$$

(4)

5. Results

Table 2. Precipitation correction factors estimated by inverse modeling for each watershed, using three objective functions. In addition, the average annual precipitation amounts from National Water Directorate (DGA) precipitation gauge records (AAP_P), isohyets (AAP_I) and inverse modeling (AAP_c) using different objective functions are shown. Made by the authors.

Area	Watershed	A [-]			AAP [mm]
		RMSE	TRMSE	ROCE	AAPP	AAPI	AAPRMSE	AAPTRMSE	AAPROCE
Andes	Chillánen Esperanza	1.38	1.38	1.43	2284	2250	3142	3151	3260
	Diguillínen San Lorenzo	1.46	1.48	1.5	2123	2500	3097	3145	3194
	Lircay en Puente las Rastras	1.03	1.08	1.12	1673	1783	1731	1798	1879
	Longavíen la Quiriquina	1.6	1.58	1.61	1727	2741	2762	2725	2775
	Perquilauquénen San Manuel	1.44	1.37	1.43	1918	2733	2754	2632	2742
	RenegadoenInvernada	0.72	0.7	0.67	2047	2483	1464	1428	1367
	Rucúe en Camino a Antuco	1.45	1.39	1.46	1833	4695	2666	2554	2679
	Estero UpeoenUpeo	1.43	1.39	1.46	1063	2000	1516	1476	1550
	Claro enlosQueñes	1.03	1.05	1.23	1293	1996	1333	1352	1586
	Colorado en Junta con Palos	1.59	1.66	1.87	1407	2109	2232	2329	2626
	Palos en Junta con Colorado	1.69	1.81	2.24	1479	2222	2494	2678	3314
	Ancoa Antes Túnel Canal Melado	1.9	1.82	2.03	2248	2500	4269	4093	4566
	Ñubleen la Punilla	1.2	1.21	1.22	2252	2753	2698	2727	2754
	Sauces Antes Junta Con Ñuble	1.24	1.22	1.21	2060	2300	2562	2509	2501
	AllipénenMelipeuco	1.56	1.59	1.62	2209	3818	3452	3521	3569
Central Valley	Claro enCamarico	0.92	0.93	1.02	1374	1544	1264	1278	1401
	Claro enRauquen	1.37	1.45	1.49	1089	1109	1491	1578	1622
	PerquilauquenenGniquen	1.11	1.09	1.15	1397	1907	1551	1523	1607
	PerquilauquenenQuella	1.1	1.08	1.19	1292	1550	1421	1395	1537
	CauquenesenDesembocadura	1.04	1.06	1.1	763	743	794	809	839
	PurapelenSauzal	0.98	1	1.04	703	800	689	703	731
	Loncomillaen Bodega	1.46	1.4	1.59	859	1059	1254	1203	1366
	Loncomillaen las Brisas	1.26	1.31	1.42	986	1335	1242	1292	1400
	ItataenTrilaleo	1.07	1.02	1.02	1731	1801	1852	1766	1766
	Itataen General Cruz	1.08	1.06	1.13	1522	1666	1644	1613	1720
	LumacoenLumaco	1.09	1.07	1.14	1020	1211	1112	1091	1163
	QuillenenGalvarino	1.13	1.22	1.29	1250	1216	1413	1525	1613
	CholCholenCholChol	1.32	1.26	1.39	1079	1259	1424	1360	1500
	QuepeenQuepe	1.5	1.43	1.5	1633	2752	2450	2335	2450
	PuyehueenQuitratue	1	0.96	0.99	2034	2000	2034	1953	2012
	DonguilenGorbea	1.11	1.07	1.07	1942	2109	2156	2078	2078
Coastal Range	CauquenesenArrayan	1.32	1.12	1.15	804	759	1062	897	875
	LoancoenDesembocadura	1.56	1.39	1.47	739	800	1157	1024	1176
	PurapelenNirivilo	1.5	1.47	1.59	699	800	1051	1027	1274
	Andalién Camino a Penco	1.74	1.7	1.6	1011	1397	1761	1718	2235
	ButamalalenButamalal	1.76	1.77	2.04	1139	2310	2006	2020	4721
	CarampangueenCarampangue	2.08	1.98	2.05	1169	2048	2432	2315	4199
	CayucupilenCayucupil	1.93	1.89	2.2	1149	2327	2222	2166	5121
	ReputoenReputo	1.09	1.08	1.23	1190	1500	1291	1277	1838
	Tucapelen Cañete	1.49	1.49	1.65	1228	2375	1823	1832	3912
	Mahuidancheen Santa Ana	1.22	1.17	1.19	1876	1825	2287	2200	2175

A: Precipitation data adjustment factor; AAP: Average annual precipitation; RMSE: Root mean squared error; TRMSE: Transformed root mean squared error; ROCE: Runoff coefficient error; AAP_P, AAP_I: Average annual precipitation calculated from isohyets; AAP_RMSE: Average annual precipitation calculated inverse modeling and RMSE objective function; AAP_TRMSE: Average annual precipitation calculated inverse modeling and TRMSE objective function; AAP_ROCE: Average annual precipitation calculated inverse modeling and ROCE objective function.

shows the results obtained from the identifiability analysis (P₅₀ of the best 10% of the models, of a total of 10,000 simulations) of factor A. As a complement, Table 2 shows: (i) average annual precipitation calculated for each watershed from DGA records using the inverse distance weighting method to estimate the representative precipitation of each watershed (AAP_P); (ii) average annual precipitation calculated from average annual precipitation isohyets from the national water balance made by the DGA [22] (AAP_I); and (iii) average annual precipitation (AAP) calculated from inverse modeling using three objective functions (AAP_RMSE, AAP_TRMSE and AAP_ROCE).

Taking the AAP_P values as a reference point for the precipitation of each watershed, the percent difference in the AAP was calculated from isohyets and inverse modeling. Thus, the greater the difference in absolute value, the less representative the recorded precipitation is of the precipitation in the watershed, while positive differences suggest an underestimation of AAP_P values and negative differences an overestimation. Consolidated results are shown in box plots grouped by watershed location (Andes, Central Valley and Coastal Range) (Figure 3).

Figure 3 shows that, in general, precipitation tends to be underestimated, although the differences are smaller in areas with flatter topography such as the Central Valley [23]. In addition, in the Andes these differences increase, probably because the increase in precipitation due to the orographic effect caused by the Andes is not properly measured by the existing stations [24]. This effect is also observed in the Coastal Range, although it is most noticeable in areas where the cordillera is higher (e.g., Nahuelbuta Range, lat. ~37°45′ S) [25]. Similar results have been obtained in hydrological studies and watershed-scale precipitation analysis in various watersheds in south-central Chile (e.g., [13,23,24,25,26,27,28]). In general, it is concluded that precipitation in cordillera zones is underestimated by between 20% and 90%.

As the model is a lumped water balance model, all the values considered in the modeling stage are assumed to be uniformly distributed in a watershed. Therefore, a point value at the geometric center of each watershed was considered representative of the precipitation and correction factor of each modeled watershed. To consolidate the results a cokriging-z (with ground elevation as auxiliary variable), geospatial interpolation of AAP_ROCE (Figure 4a) and the precipitation correction factor estimated using the ROCE function (A_ROCE, Figure 4b) was carried out. AAP_ROCE and A_ROCE values were assumed to be at the geometric center of each modeled watershed. Figure 4 leads to the same conclusions as those obtained from Figure 3. On the map it is observed that the area with the closest calculated and measured precipitation values is the Central Valley, where factor A varies around 1.

The Central Valley is the sub-area with the least spatial topographic variability and smoothest relief and, therefore, there is not high variability in precipitation caused by the interaction of frontal systems and the topography. This area is also the most accessible and thus has the greatest precipitation gauge station density (1 station every 540 km²) [23]. In addition, in the Andes and Coastal Range there is a strong interaction between the entry of frontal systems and the topography, producing an increase in precipitation as land elevation increases [26,29,30]. This variability, along with the low station density (1 station every 960 km²) explains the underestimation in the calculated precipitation amounts.

The map in Figure 4 allows the extent to which precipitation is underestimated to be seen and provides a useful tool for the correction and estimation of precipitation in south-central Chile. Along with the map, average annual precipitation isohyets every 100 mm were created (not shown in this article) for use in scientific and practical applications. The results of this study in shape file format are included with this manuscript as Supplementary Materials.

6. Conclusions

There is a low precipitation gauge station density in south-central Chile, which varies according to accessibility and is highest in the Central Valley and lowest in the mountainous areas. This density also decreases as mountain height increases, mainly due to decreasing accessibility. Although the density is lower than desired and below the level recommended by the World Meteorological Organization, the estimation and spatial representation of precipitation in the Central Valley proves to be acceptable.

In mountainous areas there is a conspicuous underestimation of precipitation due to the low precipitation gauge density. This underestimation increases along with mountain height, making it necessary to carry out precipitation corrections of up to 3 times the values recorded in the lowest areas.

Correction factors for the precipitation records from the national precipitation gauge network were determined and precipitation correction maps, along with average annual precipitation isohyets corrected for the orography, were made.

It is important to point out that this paper is focused on correcting precipitation records based on observations. It is an attempt to improve the current data availability and knowledge in the study area. Orography may not be the only variable related to spatial variability in the study area, and climate change trends and climate variability are not directly considered in the corrected precipitation amounts. Therefore, caution is necessary if climate projections, hydrological projections, or processes analysis are pursued. Including such signals or components may be necessary for medium- and long-term projections.

Although the hydrological model and its parameters, among other factors, may contribute as sources of uncertainty in the study, we deem the model not necessarily to be a major source, especially in the study area. The results show that precipitation is underestimated by up to 3.7 times in some areas. Therefore, precipitation is a great source of uncertainty for modeling and performing precipitation-related studies. In addition, it is important to point out that the methods include the simple inverse distance weighting method in the experimental design; therefore, corrected precipitation represents a further step in the precipitation estimation of the watersheds.