Improved time series land cover classification by missing-observation-adaptive nonlinear dimensionality reduction

- First nonlinear dimensionality reduction (DR) application to multi-temporal imagery
- New Laplacian Eigenmaps DR method capable of handling missing observations
- Demonstrated for Landsat time series supervised classification
- Improved land cover classification capability

Dimensionality reduction (DR) is a widely used technique to address the curse of dimensionality when high-dimensional remotely sensed data, such as multi-temporal or hyperspectral imagery, are analyzed. Nonlinear DR algorithms, also referred to as manifold learning algorithms, have been successfully applied to hyperspectral data and provide improved performance compared with linear DR algorithms. However, DR algorithms cannot handle missing data that are common in multi-temporal imagery. In this paper, the Laplacian Eigenmaps (LE) nonlinear DR algorithm was refined for application to multi-temporal satellite data with large proportions of missing data. Refined LE algorithms were applied to 52-week Landsat time series for three study areas in Texas, Kansas and South Dakota that have different amounts of missing data and land cover complexity. A series of random forest classifications were conducted on the refined LE DR bands using varying proportions of training data provided by the United States Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL); these classification results were compared with conventional metrics-based random forest classifications. Experimental results show that compared with the metrics approach, higher per-class and overall classification accuracies were obtained using the refined LE DR bands of multispectral reflectance time series, and the number of training samples required to achieve a given degree of classification accuracy was also reduced. The approach of applying the refined LE to multispectral reflectance time series is promising in that it is automated and provides dimensionality-reduced data with desirable classification properties. The implications of this research and possibilities for future algorithm development and application are discussed.