Predicting flood stages in watersheds with different scales using hourly rainfall dataset: A high-volume rainfall features empowered machine learning approach

Accurate prediction of instantaneous high lake water levels and flood flows (flood stages) from micro-catchments to big river basins are critical for flood forecasting. Lake Carl Blackwell, a small-watershed reservoir in the south-central USA, served as a primary case study due to its rich historical dataset. Bearing knowledge that both current and previous rainfall contributes to the reservoirs' water body, a series of hourly rainfall features were created to maximize predicting power, which include total rainfall amounts in the current hour, the past 2 h, 3 h, …, 600 h in addition to previous-day lake levels. Notedly, the rainfall features are the accumulated rainfall amounts from present to previous hours rather than the rainfall amount in any specific hour. Random Forest Regression (RFR) was used to score the features' importance and predict the flood stages along with Neural Network - Multi-layer Perceptron Regression (NN-MLP), Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), and the ordinary multi-variant linear regression (MLR) together with dimension reduced linear models of Principal Component Regression (PCR) and Partial Least Square Regression (PLSR). The prediction accuracy for the lake flood stages can be as high as 0.95 in R2, 0.11 ft. in mean absolute error (MAE), and 0.21 ft. in root mean square error (RMSE) for the testing dataset by the RFR (NN-MLP performed equally well), with small accuracy decreases by the other two non-linear algorithms of XGBoost and SVR. The linear regressions with dimension reductions had the lowest accuracy. Furthermore, our approach demonstrated high accuracy and broad applicability for surface runoff and streamflow predictions across three different-sized watersheds from micro-catchment to big river basins in the region, with increases of predicting power from earlier rainfall for larger watersheds and vice versa.

Detection of hidden model errors by combining single and multi-criteria calibration

Environmental models aim to reproduce landscape processes with mathematical equations. Observations are used for validation. The performance and uncertainties are quantified either by single or multi-criteria model assessment. In a case-study, we combine both approaches. We use a coupled hydro-biogeochemistry landscape-scale model to simulate 14 target values on discharge, stream nitrate as well as soil moisture, soil temperature and trace gas emissions (N₂O, CO₂) from different land uses. We reveal typical mistakes that happen during both, single and multi-criteria model assessment. Such as overestimated uncertainty in multi-criteria and ignored wrong model processes in single-criterion calibration. These mistakes can mislead the development of water quality and in general all environmental models. Only the combination of both approaches reveals the five types of posterior probability distributions for model parameters. Each type allocates a specific type of error. We identify and locate mismatched parameter values, obsolete parameters, flawed model structures and wrong process representations. The presented method can guide model users and developers to the so far hidden errors in their models. We emphasize to include observations from physical, chemical, biological and ecological processes in the model assessment, rather than the typical discipline specific assessments.

Improved Land Evapotranspiration Simulation of the Community Land Model Using a Surrogate-Based Automatic Parameter Optimization Method

Land surface evapotranspiration (ET) is important in land-atmosphere interactions of water and energy cycles. However, regional ET simulation has a great uncertainty. In this study, a highly-efficient parameter optimization framework was applied to improve ET simulations of the Community Land Model version 4.0 (CLM4) in China. The CLM4 is a model at land scale, and therefore, the monthly ET observation was used to evaluate the simulation results. The optimization framework consisted of a parameter sensitivity analysis (also called parameter screening) by the multivariate adaptive regression spline (MARS) method and sensitivity parameter optimization by the adaptive surrogate modeling-based optimization (ASMO) method. The results show that seven sensitive parameters were screened from 38 adjustable parameters in CLM4 using the MARS sensitivity analysis method. Then, using only 133 model runs, the optimal values of the seven parameters were found by the ASMO method, demonstrating the high efficiency of the method. For the optimal parameters, the ET simulations of CLM4 were improved by 7.27%. The most significant improvement occurred in the Tibetan Plateau region. Additional ET simulations from the validation years were also improved by 5.34%, demonstrating the robustness of the optimal parameters. Overall, the ASMO method was found to be efficient for conducting parameter optimization for CLM4, and the optimal parameters effectively improved ET simulation of CLM4 in China.

The plumbing of land surface models: is poor performance a result of methodology or data quality?

The PALS Land sUrface Model Benchmarking Evaluation pRoject (PLUMBER) illustrated the value of prescribing a priori performance targets in model intercomparisons. It showed that the performance of turbulent energy flux predictions from different land surface models, at a broad range of flux tower sites using common evaluation metrics, was on average worse than relatively simple empirical models. For sensible heat fluxes, all land surface models were outperformed by a linear regression against downward shortwave radiation. For latent heat flux, all land surface models were outperformed by a regression against downward shortwave, surface air temperature and relative humidity. These results are explored here in greater detail and possible causes are investigated. We examine whether particular metrics or sites unduly influence the collated results, whether results change according to time-scale aggregation and whether a lack of energy conservation in flux tower data gives the empirical models an unfair advantage in the intercomparison. We demonstrate that energy conservation in the observational data is not responsible for these results. We also show that the partitioning between sensible and latent heat fluxes in LSMs, rather than the calculation of available energy, is the cause of the original findings. Finally, we present evidence suggesting that the nature of this partitioning problem is likely shared among all contributing LSMs. While we do not find a single candidate explanation for why land surface models perform poorly relative to empirical benchmarks in PLUMBER, we do exclude multiple possible explanations and provide guidance on where future research should focus.

Land Surface Model and Particle Swarm Optimization Algorithm Based on the Model-Optimization Method for Improving Soil Moisture Simulation in a Semi-Arid Region

Improving the capability of land-surface process models to simulate soil moisture assists in better understanding the atmosphere-land interaction. In semi-arid regions, due to limited near-surface observational data and large errors in large-scale parameters obtained by the remote sensing method, there exist uncertainties in land surface parameters, which can cause large offsets between the simulated results of land-surface process models and the observational data for the soil moisture. In this study, observational data from the Semi-Arid Climate Observatory and Laboratory (SACOL) station in the semi-arid loess plateau of China were divided into three datasets: summer, autumn, and summer-autumn. By combing the particle swarm optimization (PSO) algorithm and the land-surface process model SHAW (Simultaneous Heat and Water), the soil and vegetation parameters that are related to the soil moisture but difficult to obtain by observations are optimized using three datasets. On this basis, the SHAW model was run with the optimized parameters to simulate the characteristics of the land-surface process in the semi-arid loess plateau. Simultaneously, the default SHAW model was run with the same atmospheric forcing as a comparison test. Simulation results revealed the following: parameters optimized by the particle swarm optimization algorithm in all simulation tests improved simulations of the soil moisture and latent heat flux; differences between simulated results and observational data are clearly reduced, but simulation tests involving the adoption of optimized parameters cannot simultaneously improve the simulation results for the net radiation, sensible heat flux, and soil temperature. Optimized soil and vegetation parameters based on different datasets have the same order of magnitude but are not identical; soil parameters only vary to a small degree, but the variation range of vegetation parameters is large.

The challenge of predicting flash floods from thunderstorm rainfall

A major characteristic of the hydrometeorology of semi-arid regions is the occurrence of intense thunderstorms that develop very rapidly and cause severe flooding. In summer, monsoon air mass is often of subtropical origin and is characterized by convective instability. The existing observational network has major deficiencies for those regions in providing information that is important to run-off generation. Further, because of the complex interactions between the land surface and the atmosphere, mesoscale atmospheric models are currently able to reproduce only general features of the initiation and development of convective systems. In our research, several interrelated components including the use of satellite data to monitor precipitation, data assimilation of a mesoscale regional atmospheric model, modification of the land component of the mesoscale model to better represent the semi-arid region surface processes that control run-off generation, and the use of ensemble forecasting techniques to improve forecasts of precipitation and run-off potential are investigated. This presentation discusses our ongoing research in this area; preliminary results including an investigation related to the unprecedented flash floods that occurred across the Las Vegas valley (Nevada, USA) in July of 1999 are discussed.