CR 79:1-8 (2019)  -  DOI: https://doi.org/10.3354/cr01582

ABSTRACT: Robust statistical techniques are required to estimate the trend in meteorological data, where data are available only for a limited period with too many missing observations. We examined the application of a dynamic linear model (DLM) for estimating changes in time series meteorological data. For this purpose, we used maximum and minimum monthly temperatures recorded over 36 yr at 6 meteorological stations representing 3 physiographic regions in the Central Himalayas. Temperature changes over time may be influenced by hidden processes, such as seasonality. To elucidate such processes, we estimated a Fourier-form seasonal model with 12 seasons with 2 harmonics. The DLM model fit was evaluated based on the distribution of standardized residuals and the p-value of Ljung-Box statistics. We reported the level of temperature change from 1980 to 2015. The DLM results were compared with more conventional, simple linear regression analysis (SLR). Although a significant trend of temperature increase was observed in the Central Himalayas, the change was not uniform across the physiographic regions. Localized changes in temperature levels may be due to geographic configuration and micro-climate. Notably, the SLR showed a similar trend of average annual temperature change to that of the DLM but overestimated the magnitude of the change. SLR is easily executed but is sensitive to outliers and non-normality in the observations. Under these circumstances, DLM may be a more robust modeling technique for estimating changes in meteorological data.

KEY WORDS: Central Himalayas · Global warming · State-space model · Seasonality and trend · Kalman filter · Markov chain · Simple linear regression (SLR) · Mann-Kendall trend test