Precipitation climatologies for the period 1961–1990 were generated for all climatic regions of Costa Rica using an irregular rain-gauge observational network comprised by 416 rain-gauge stations. Two sub-networks were defined: a high temporal resolution sub-network (HTR), including stations having at least 20 years of continuous records during the study period (157 in total); and a high spatial resolution sub-network (HSR), which includes all HTR-stations plus those stations with less than 20 years of continuous records (416 in total). Results from the kriging variance reduction efficiency (KRE) objective function between the two sub-networks, show that ordinary kriging (OK) is unable to fully explain the spatio-temporal variability of precipitation within most climatic regions if only stations from the HTR sub-network are used. Results also suggests that in most cases, it is beneficial to increase the density of the rain-gauge observational network at the expense of temporal fidelity, by including more stations even though their records may not represent the same time step. Thereafter, precipitation climatologies were generated using seven deterministic (IDW, TS2, TS2PARA, TS2LINEAR, TPS, MQS and NN) and two geostatistical (OK and KED) interpolation methods. Performance of the various interpolation methods was evaluated using cross validation technique, selecting the mean absolute error (MAE) and the root-mean square error (RMSE) as agreement metrics. Results suggest that IDW is marginally superior to OK and KED for most climatic regions. The remaining deterministic methods however, considerably deviate from IDW, which suggests that these methods are incapable of properly capturing the true-nature of spatial precipitation patterns over the considered climatic regions. The final generated IDW climatology was then validated against the Global Precipitation Climatology Centre (GPCC), Climate Research Unit (CRU) and WorldClim datasets, in which overall spatial and temporal coherence is considered satisfactory, giving assurance about the use this new climatology in the development of local climate impact studies.

This study aimed to assess the impacts of climate change on streamflow characteristics of five tropical catchments located in Costa Rica. An ensemble of five General Circulation Models (GCMs), namely HadGEM2-ES, CanESM2, EC-EARTH, MIROC5, MPI-ESM-LR dynamically downscaled by two Regional Climate Models (RCMs), specifically HadRM3P and RCA4, was selected to provide an overview of the impacts of different climate change scenarios under Representative Concentration Pathways (RCPs) 2.6, 4.5 and 8.5 using the 1961–1990 baseline period. The GR2M hydrological model was used to reproduce the historical monthly surface runoff patterns of each catchment. Following calibration and validation of the GRM2 model, the projected impact of climate change on streamflow was simulated for a near-future (2011–2040), mid-future (2041–2070) and far-future (2071–2100) for each catchment using the bias-corrected GCM-RCM multimodel ensemble-mean (MEM). Results anticipate wetter conditions for all catchments in the near-future and mid-future periods under RCPs 2.6 and 4.5, whereas dryer conditions are expected for the far-future period under RCP 8.5. Projected temperature trends indicate consistently warmer conditions with increasing radiative forcing and future periods. Streamflow changes across all catchments however are dominated by variations in projected precipitation. Wetter conditions for the near-future and mid-future horizons under RCPs 2.6 and 4.5 would result in higher runoff volumes, particularly during the late wet season (LWS). Conversely, dryer conditions for the far-future period under RCP8.5 would result in considerably lower runoff volumes during the early wet season (EWS) and the Mid-Summer Drought (MSD). In consequence, projected seasonal changes on streamflow across all catchments may result in more frequent flooding, droughts, and water supply shortage compared to historical hydrological regimes.

Deep Learning and transfer learning models are being used to generate time series forecasts; however, there is scarce evidence about their performance prediction that it is more evident for monthly time series. The purpose of this paper is to compare Deep Learning models with transfer learning and without transfer learning and other traditional methods used for monthly forecasts to answer three questions about the suitability of Deep Learning and Transfer Learning to generate predictions of time series. Time series of M4 and M3 competitions were used for the experiments. The results suggest that deep learning models based on TCN, LSTM, and CNN with transfer learning tend to surpass the performance prediction of other traditional methods. On the other hand, TCN and LSTM, trained directly on the target time series, got similar or better performance than traditional methods for some forecast horizons.

This work aims to examine the effect of bias correction (BC) methods on the development of Intensity–Duration–Frequency (IDF) curves under climate change at multiple temporal scales. Daily outputs from a 9-member CORDEX-CA GCM-RCM multi-model ensemble (MME) under RCP 8.5 were used to represent future precipitation. Two stationary BC methods, empirical quantile mapping (EQM) and gamma-pareto quantile mapping (GPM), along with three non-stationary BC methods, detrended quantile mapping (DQM), quantile delta mapping (QDM), and robust quantile mapping (RQM), were selected to adjust daily biases between MME members and observations from the SJO weather station located in Costa Rica. The equidistant quantile-matching (EDQM) temporal disaggregation method was applied to obtain future sub-daily annual maximum precipitation series (AMPs) based on daily projections from the bias-corrected ensemble members. Both historical and future IDF curves were developed based on 5 min temporal resolution AMP series using the Generalized Extreme Value (GEV) distribution. The results indicate that projected future precipitation intensities (2020–2100) vary significantly from historical IDF curves (1970–2020), depending on individual GCM-RCMs, BC methods, durations, and return periods. Regardless of stationarity, the ensemble spread increases steadily with the return period, as uncertainties are further amplified with increasing return periods. Stationary BC methods show a wide variety of trends depending on individual GCM-RCM models, many of which are unrealistic and physically improbable. In contrast, non-stationary BC methods generally show a tendency towards higher precipitation intensities as the return period increases for individual GCM-RCMs, despite differences in the magnitude of changes. Precipitation intensities based on ensemble means are found to increase with the change factor (CF), ranging between 2 and 25% depending on the temporal scale, return period, and non-stationary BC method, with moderately smaller increases for short-durations and long-durations, and slightly higher for mid-durations. In summary, it can be concluded that stationary BC methods underperform compared to non-stationary BC methods. DQM and RQM are the most suitable BC methods for generating future IDF curves, recommending the use of ensemble means over ensemble medians or individual GCM-RCM outcomes.

There is a gap of knowledge about the conditions that explain why a method has a better forecasting performance than another. Specifically, this research aims to find the factors that can influence deep learning models to work better with time series. We generated linear regression models to analyze if 11 time series characteristics influence the performance of deep learning models versus statistical models and other machine learning models. For the analyses, 2000 time series of M4 competition were selected. The results show findings that can help explain better why a pretrained deep learning model is better than another kind of model.

Ocho métodos de optimización global y ocho métodos de optimización local fueron utilizados para calibrar el modelo hidrológico conceptual HBV-TEC en la cuenca alta del río Toro en Costa Rica para cuatro diferentes periodos de calibración (4, 8, 12 y 16 años). Con el propósito de evaluar la sensibilidad de quedar atrapados en mínimos locales, cada método fue probado contra 50 sets de parámetros iniciales generados aleatoriamente. Todos los métodos fueron entonces evaluados en términos del desempeño de optimización y el costo computacional. Los resultados muestran un desempeño comparable entre varios métodos locales y globales dado que se correlacionan fuertemente entre ellos. Sin embargo, los métodos locales son generalmente más sensitivos a quedar atrapados en mínimos locales independientemente de la duración del periodo de calibración. El desempeño de optimización parece ser independiente del número total de llamas del modelo, el cual puede variar varios órdenes de magnitud dependiendo del método de optimización seleccionado. La selección final de un método de optimización está grandemente influenciada por su eficiencia y el nivel de recursos computacionales disponible indistintamente de clase local o global.