Comparative analysis of three temporal smoothing algorithms on annual land cover time series

Gaps and noises are two major challenges for using time series remote sensing data for land cover mapping and change analysis. Various algorithms have been developed to applying temporal smoothing remote sensing data but effectiveness of different smoothing for mapping different land cover types have not been well researched. In this study we compared the performance of discrete Fourier transformation smoothing, Whittaker smoother and Linear-Fit averaging smoother on Landsat 5,7 and 8 based yearly composites to classify land cover in province-1 of Nepal. The performance of smoothers was also tested based on whether it was applied on composites or on intermediate probability layers generated using random forests classification. The study was performed for years 2000 to 2018. Probability distribution was examined to check the quality of intermediate layers while a confusion matrix was used for accuracy assessment of the final land cover maps. In most cases the accuracy improved with smoothing except for classes that go through frequent changes. The best results were found for the Whittaker smoothing algorithm for stable classes and Fourier transformation smoothing for other classes. In combining the best results obtained from different approaches, the final accuracy of our land cover map increased from 79.18% to 83.30%.