An evaluation of time-series smoothing algorithms for land-cover classifications using MODIS-NDVI multi-temporal data

- Four smoothing algorithms were examined for MODIS land-cover classification.
- All smoothing algorithms can significantly reduce intra-class variability.
- Smoothed data resulted in large inconsistencies of Jeffries-Matusita (JM) measures.
- Fourier smoothing algorithm performed best in improving classification accuracy.

In this study we compared the Savitzky-Golay, asymmetric Gaussian, double-logistic, Whittaker smoother, and discrete Fourier transformation smoothing algorithms (noise reduction) applied to Moderate Resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI) time-series data, to provide continuous phenology data used for land-cover (LC) classifications across the Laurentian Great Lakes Basin (GLB). MODIS 16-day 250 m NDVI imagery for the GLB was used in conjunction with National Land Cover Database (NLCD) from 2001, 2006 and 2011, and the Cropland Data Layers (CDL) from 2011 to 2014 to conduct classification evaluations. Inter-class separability was measured by Jeffries-Matusita (JM) distances between selected cover type pairs (both general classes and specific crops), and intra-class variability was measured by calculating simple Euclidean distance for samples within cover types. For the GLB, we found that the application of a smoothing algorithm significantly reduced image noise compared to the raw data. However, the Jeffries-Matusita (JM) measures for smoothed NDVI temporal profiles resulted in large inconsistencies. Of the five algorithms tested, only the Fourier transformation algorithm and Whittaker smoother improved inter-class separability for corn-soybean class pair and significantly improved overall classification accuracy. When compared to the raw NDVI data as input, the overall classification accuracy from the Fourier transformation and Whittaker smoother improved performance by approximately 2-6% for some years. Conversely, the asymmetric Gaussian and double-logistic smoothing algorithms actually led to degradation of classification performance.