Surface runoff response to climate change based on artificial neural network (ANN) models: a case study with Zagunao catchment in Upper Minjiang River, Southwest China

Climate change and its hydrological consequences are of great concern for water resources managers in the context of global change. This is especially true for Upper Minjiang River (UMR) basin, where surface runoff was reported to decrease following forest harvesting, as this unusual forest–water relationship is perhaps attributed to climate change. To quantify the hydrological impacts of climate change and to better understand the forest–water relationship, an artificial neural network (ANN)based precipitation–runoff model was applied to Zagunao catchment, one of the typical catchments in UMR basin, by a climate scenario-based simulation approach. Two variables, seasonality and CTsm (cumulative temperature for snow melting), were devised to reflect the different flow generation mechanisms of Zagunao catchment in different seasons (rainfall-induced versus snow meltingoriented). It was found that the ANN model simulated precipitation–runoff transformation very well (R1⁄4 0.962). Results showed runoff of Zagunao catchment would increase with the increase in precipitation as well as temperature and such a response was season dependent. Zagunao catchment was more sensitive to temperature rise in the non-growing season but more sensitive to precipitation change in the growing season. Snow melting-oriented runoff reduction due to climate change is perhaps responsible for the unusual forest–water relationship in UMR basin. This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/). doi: 10.2166/wcc.2018.130 om https://iwaponline.com/jwcc/article-pdf/10/1/158/533171/jwc0100158.pdf 2020 Yong Lin National Environmental Monitoring Center, State Oceanic Administration, Dalian 116023, China Hui Wen College of Urban and Environmental Sciences, Peking University, Beijing 100871, China Shirong Liu (corresponding author) Institute of Forest Ecology, Environment and Protection, Chinese Academy of Forestry, Beijing 100091, China E-mail: liusr@forestry.ac.cn


INTRODUCTION
Global climate change caused by growing atmospheric concentration of CO 2 and other trace gases has become evident.
The acceptance that increasing CO 2 concentration in the atmosphere will cause global climate change, especially the change in precipitation and temperature, has led to an increased interest in the impact of climate change on region hydrology among scientists (Guo et al. ; Jiang et al. a; Teutschbein & Seibert ; Naz et al. ). Global climate change is very likely to affect the hydrological cycle and consequently water resources by increasing evaporation due to rising air temperature and changing precipitation (Guo et al. ; Huntington ). In addition, global warming or its increased variability is expected to alter the timing and magnitude of runoff, the frequency and intensity of floods and droughts, rainfall patterns, extreme weather events, and the quality and quantity of water availability (Guo et al. ; Jiang et al. a). These changes, in turn, influence the water supply system, power generation, sediment transport gical models behave better than traditional hydrological models (conceptual, physically based model or statistical model) in runoff prediction (Dawson &  Like many places in China and the world, water shortage is a big problem in the Upper Minjiang River (UMR) basin in southwest China. Due to its special geographic location (the transition zone from Sichuan basin to Qinghai-Tibet plateau), the UMR basin is prone to climate change and so the study of the hydrological response of the UMR basin to climate change has many implications for the sustainable utilization and management of water resources. The UMR basin has experienced large-scale land-use and land-cover change (LUCC) resulting mainly from over-logging. Several forest-water studies in the UMR basin showed that the water yield decreased with the reduction of forest cover (Ma ; Zhang et al. ; Sun et al. ), which is inconsistent with many other similar studies (Brown et al. ). Climate variability may account for this particular hydrological phenomenon. In fact, LUCC (forest cover change in the UMR) and climate change in a basin are usually mixed, making it difficult to determine the pure hydrological effect of vegetation, especially forest landscape change on runoff. Due to the characteristics of high elevation and the corresponding low temperature in the UMR, snow-melting runoff is greatly affected by global warming, which may be a reason for the unusual forest-water relationship found in the UMR basin. The knowledge of surface runoff response of UMR to climate change, therefore, is expected to be helpful for understanding the eco-hydrological function of forest vegetation.
In this paper, an ANN-based precipitation-runoff model was used to study the hydrological response of Zagunao catchment, a typical catchment in the UMR basin, to climate change. The objectives of this study are as follows: (1) to study the runoff response of the UMR basin to climate change and provide baseline information for water resources management; and (2) to help understand the unusual forest-water relationship in the UMR basin by analyzing the hydrological consequences of climate change.

Study area
As an important branch of the Yangtze River, the UMR has a total drainage area of 22,900 km 2 basin (102-104 0 E, 31-33 0 N) and a total length of 340 km with an annual mean discharge of 469 m 3 /s. Zagunao River, one of the main branches of the UMR, has a drainage area of 2,528 km 2 (102.58-103.22 0 E, 31.18-31.93 0 N). The elevation of Zagunao catchment ranges from 1,823 to 5,769 m above sea level and the area above the elevation of 3,800 m accounts for 56.80% of the whole catchment with permanent snow and ice cover scattered in the catchment. There is great spatial variation in precipitation and temperature as a result of large topographic variation. Mean annual precipitation ranges from 627.5 mm to 1,478.0 mm whereas mean annual air temperature varies between À1.7 C and 12.2 C. Due to large spatial variation in temperature means, snow-melting occurs in both winter-spring (non-growing) season (in low elevation areas) and summer-autumn (growing) season (in high elevation areas). The precipitation is usually manifested in low to middle intensity of rainfall in summer and autumn, and snowfall occurs in winter and spring (Sun et al. ).
Thanks to cold and humid climate conditions, subalpine conifer forest, which is mainly composed of Picea asperata

Hydrological and meteorological data
Three rain gauge stations located in Miyaluo, Zagunao, and Shangping are available for hydrological research in Zagunao catchment. The former two stations (Miyaluo and Zagunao) are within the study catchment whereas the last one (Shanping) is outside of Zagunao catchment. To better reflect the spatial variation of precipitation in Zagunao catchment, the precipitation record from Shangping rain gauge station was also used in this study. Many methods are available for estimating mean areal precipitation over an area (e.g., catchment) based on the observation records of stations, which include spline, inverse distance weight (IDW), trend surface, kriging and Thiessen polygons. However, as only three rain gauge stations were available for this study it means that many of the above-mentioned methods are not suitable for use. Here, the monthly precipitation data (Pcp) from the three stations were just simply averaged with equal weight to get monthly precipitation data.
There is only one climate station (Li county climate station) within the study area, and it is virtually in the same position as Zagunao rain gauge station, meaning that its precipitation data are of no use in this study. Monthly average relative humidity and monthly evaporation were calculated as input for the ANN model (below). Instead of the traditional monthly average air temperature, monthly cumulative temperature for snow melting (CTsm), a new temperature variable devised by us, was used as the input for the ANN model. CTsm is calculated by the following formula: where T i is daily average temperature, n is the number of days in a month of interest. CTsm is a monthly counterpart to the variable of degree-day widely used in a snow-melting runoff model on a daily basis (Singh & Kumar ). In a degree-day based snow-melting runoff model, daily snowmelting runoff yield is calculated on active temperature above 0 C (degree-day) rather than daily average air temperature to reflect the snow-melting physical mechanism, and a similar idea was applied to monthly snow-melting runoff calculation in this study. CTsm is expected to behave much better than monthly average temperature in predicting the impacts of climate change on surface runoff in Zagunao catchment given that the temperature change below zero, for instance, from À30.0 C to À10.2 C, contributes nothing to snow-melting.  also employed to assess model performance. The formula for the indicators of PRMSE and MAE is given as follows:

ANN-based precipitation-runoff model
The  Table 1). In addition, it is clear that model behavior in the growing season was much better than in the non-growing season and had the best goodness of fit on annual basis (R 2 ¼ 0.962) in terms of R 2 . However, when judged by  Table 2 and Figure 3). It was also found that the performance of the established ANN model varied with season and criterion. In terms of R 2 , the ANN model had the best performance on an annual basis (R 2 ¼ 0.915) and the worst performance (R 2 ¼ 0.719) in the non-growing season. However, when it comes to PRMSE, the ANN model resulted in the best goodness of fit in the growing season which is also supported by MAE when weighted by the mean values shown in parentheses in suggested that the established model was eligible for the study of surface response to climate change, especially considering the fact that surface runoff of Zagunao catchment (26.84 m 3 /s) in the non-growing season was much smaller than its counterparts in the growing season (100.17 m 3 /s) and on annual basis (63.50 m 3 /s).

Runoff response to climate change
Runoff response of Zagunao catchment to temperature change and precipitation change scenarios is shown in   and would decrease by 3.28-9.48% when the precipitation decreased by 5-15%. Similarly, runoff response to precipitation change was also season-dependent. The response of runoff to precipitation change was much more sensitive in the growing season than in the non-growing season.
Runoff would increase by 7.44% for the 10% increase in precipitation scenario in contrast with 2.65% increase in the non-growing season for the same scenario.   Some studies found that the impact of climate change on regional or catchment water resources are model dependent ( Jiang et al. a), which can be attributed to the assumptions regarding the various processes in the hydrological model. ANN models are distribution-free models and do not require any a priori assumption regarding the processes involved. Therefore, the ANN-based precipitationrunoff model used here is expected to do a better job in simulating the runoff response of the Zagunao catchment to climate change. In addition, in the era of large data and in the context of increased concern regarding climate change impacts on water resources, data-driven ANN models should be highly appreciated, especially when considering the fact that water resources managers are mainly concerned about how regional water resources respond to climate change rather than the mechanism behind such a response. After all, ANN-based hydrological models have a better performance than traditional hydrological models Zagunao catchment, a typical one in the UMR basin, was used as a case to study the impact of climate changes on water resources in the UMR region in this study. Given that water resources is becoming a limiting factor for sustainable development of the UMR region and that the impacts of climate change on water resources are increasingly self-evident, the result from this study is expected to provide scientific foundation for future water resources planning and management.

DISCUSSION AND CONCLUSIONS
The main conclusions are summarized as follows: 1. With the variables of CTsm and seasonality incorporated into the input variable list, the proposed ANN precipitation-runoff model was capable of simulating precipitation-runoff transformation process reasonably well.
2. The runoff of Zagunao catchment increased with temperature increase and such response was season dependent. In comparison with the growing season, the non-growing season was more sensitive to global warming.
3. The runoff of Zagunao catchment to precipitation change was also season dependent with runoff in the growing season being more sensitive to precipitation change.