Forecasting seasonal rainfall with copula modelling approach for agricultural stations in Papua New Guinea

Forecasting seasonal rainfall with copula modelling approach for agricultural stations in Papua New Guinea

Developing innovative forecasting tools is important to address issues related to climate change, agriculture, and economy of small Pacific Island nations. Papua New Guinea, PNG is a developing nation that is vulnerable to the imminent threats of climate change and influences agricultural sector that supports a majority of its citizens. Accurate modeling and forecasting methods for both monthly and seasonal rainfall (that influences agricultural and other human activities) by employing large-scale climate mode indices (linked to rainfall events) are significant predictive tools for developing climate resilience and productivity in agricultural activities.

Copula statistical models, developed in this Master’s study, are considered as viable alternative tools to fulfill this objective. This Masters by Research Thesis utilizes the D-vine copula-based quantile regression methods that are developed to create a model between statistically significant lagged relationships and joint influences of large-scale climate mode indices such as the El-Niño Southern Oscillation (ENSO) and Indian Ocean Dipole- on seasonal rainfall data across four major agricultural-based weather stations. Copula techniques allow the respective model to fully capture the dependence structure between input(s) and the target variable regardless of the marginal distribution of each variable. The D-vine copula-based quantile approach, used in this study, through Akaike information criterion (AIC)-corrected conditional log-likelihood (cllAIC) can also enable researchers to identify the most influent predictor variables for seasonal rainfall forecasting.

To forecast the monthly and the respective seasonal rainfall for PNG, an agricultural-reliant nation, the statistically significant lagged correlations between ENSO indicators (e.g., SOI, Nino3.0, etc.) and the IOD indicator (i.e., DMI) with a three-monthly total rainfall were established for up to 7 months ahead time. For example, in a 'lead-0' timescale case study for seasonal rainfall forecasting, this study has utilized the January to March average SOI (as a model input) relative to the April to June total rainfall (as the target variable) deduced by the Kendall rank correlation coefficients established between the input and the target variable.

In terms of the results of this study, a correlation analysis performed between the most optimal lead times considering climate mode indices and the three-monthly total rainfall were found to be consistent with the most influent predictor variables identified from the D-vine copula-based quantile model (as a basis to generate bivariate models that captured ENSO impacts on rainfall). To further explore any improvements in rainfall forecast model accuracy, particularly, the extreme rainfall events, the study has also considered the impact of Indian Ocean Dipole (IOD) index by embedding the DMI into the bivariate model to finally construct a trivariate forecast models that accounts for compound effects of ENSO and IOD on extreme rainfall events.

To ascertain the versatility of the proposed copula-based forecast models as a major contribution of this study, a number of statistical score metrics based on the Willmott's Index (d), Nash–Sutcliffe Efficiency (ENS), Legates-McCabe’s Index (L), root-mean-square-error (RMSE), and mean absolute error (MAE), including the Relative Root Mean Square Error (RRMSE) and Mean Absolute Percentage Error (MAPE) are computed from forecasted and observed rainfall data in the testing phase. It was evident that the station Aiyura attained the best result for both the bivariate and the trivariate model, exhibiting r = 0.63, RMSE = 105.99, MAE = 89.75, ENS = 0.63, d = 0.38, L=0.20 with, the RRMSE =15.39% for the bivariate study, whereas the trivariate model evaluations generated a score metric of 0.68, 0.42, 0.28 and 14.84%, respectively.

In summary, the copula statistical modelling approaches contributed by this study, can be enabling mechanisms for climate change resilience, measuring and implementing risk management strategies. These predictive tools can have significant implications for applications in many socioeconomic sectors such as water resources management, better farming practices for crop health, and other agricultural management not only in the present study region but also in the other agricultural-reliant nations where rainfall prediction is often challenging task.