Evaluation of Spatial interpolation techniques for mapping climate variables with low sample density: a case study using a new gridded dataset of Bangladesh

Dissertation submitted in partial fulfillment of the requirements for the Degree of Master of Science in Geospatial Technologies.

This study explores and analyses the impact of sample density on the performances of the spatial interpolation techniques. It evaluates the performances of two alternative deterministic techniques – Thin Plate Spline and Inverse Distance Weighting, and two alternative stochastic techniques – Ordinary Kriging and Universal Kriging; to interpolate two climate indices - Annual Total Precipitation in Wet Days and the Yearly Maximum Value of the Daily Maximum Temperature, in a low sample density region - Bangladesh, for 60 years – 1948 to 2007. It implies the approach of Spatially Shifted Years to create mean variograms with respect to the low sample density. Seven different performance measurements - Mean Absolute Error, Root Mean Square Errors, Systematic Root Mean Square Errors, Unsystematic Root Mean Square Errors, Index of Agreement, Coefficient of Variation of Prediction and Confidence of Prediction, have been applied to evaluate the performance of the spatial interpolation techniques. The resulted performance measurements indicate that for most of the years the Universal Kriging method performs better to interpolate total precipitation, and the Ordinary Kriging method performs better to interpolate the maximum temperature. Though the difference surfaces indicate a very little difference in the estimating ability of the four spatial interpolation techniques, the residual plots refer to the differences in the interpolated surfaces by different techniques in terms of their over and under estimation. The results also indicate that the Inverse Distance Weighting method performs better for both indices, when the sample density is too low, but the performance is questioned by the inclusion of measurement errors in the interpolated surfaces. All the error measurements show a decreasing trend with the increasing sample density, and the index of agreement and confidence of prediction show an increasing trend over years. Finally, the strong correlation between the Sample Coefficient of Variation and the performance measurements, implies that the more representative the samples are of the climate phenomenon, the more improved are the performances of the spatial interpolation techniques. The correlation between the sample coefficient of variation and the number of samples implies that the high representativity of the sample is attainable with an increased sample density.