Impact of Poor Data Quality in Remotely Sensed Data

Abstract

Poor data quality impacts negatively on various classification approaches. Classification is part of an important aspect of research, and often there are missing values found in data, which poses a problem in critical decision-making in terms of accuracy. Often selecting the best approach to handle missing data can be a difficult task, as there are several conditions to consider. Remote sensing as a discipline is quite susceptible to missing data. The objective of this study is to evaluate the robustness and accuracy of four classifiers when dealing with the incomplete remote sensing data problem. Two remote sensing data sets are utilised for this task with a four-way repeated-measures design employed to analyse the results. Simulation results suggest k-nearest neighbour as a superior approach to handling missing data, especially when regression imputation is used. Most classifiers achieve lower accuracy when listwise deletion is used. Nonetheless, RF is much less robust to missing data compared to other classifiers such as ANN and SVM.