Abstract
Flood prediction for ungauged karst wetland is facing a great challenge. How to build a wetland hydrological model when there is a lack of basic hydrological data is the key to dealing with the above challenge. Napahai wetland is a typical ungauged karst wetland. In ungauged wetland/condition, this article used the wetland open water area (OWA) extracted from Landsat remote sensing images during 1987-2018 to characterize the hydrological characteristics of Napahai wetland. The local daily precipitation in the 1987-2018 rainy season (June-October) was used to set the variables. Based on the following hypothesis: in the rainy season, the OWA of the Napahai wetland rises when there is an increase in accumulated precipitation (AP), two data-driven models were established. The study took the area difference (AD) between two adjacent OWAs as the dependent variable, the accumulated precipitation (AP) within the acquisition time of two adjacent OWAs, and the corresponding time interval (TI) of the OWA as explanatory variables. Two data-driven models (a piecewise linear regression model and a decision tree model) were established to carry out flood forecasting simulations. The decision tree provided higher goodness of fit while the piecewise linear regression could offer a better interpretability between the variables which offset the decision tree. The results showed that: (1) the goodness of fit of the decision tree is higher than that of the piecewise linear regression model (2) the piecewise linear model has a better interpretation. When AP increased by 1 mm, the average AD increased by 2.41 ha; when TI exceeded 182 d and increased by 1 d, the average AD decreased to 3.66 ha. This article proposed an easy decision plan to help the local Napahai water managers forecast floods based on the results from the two models above. In addition, the modelling method proposed in this article, based on the idea of difference for non-equidistant time series, can be applied to karst wetland hydrological simulation problems with data acquisition difficulty.