A parametric model for classifying land cover and evaluating training data based on multi-temporal remote sensing data

Time series of multispectral images are widely used to monitor and map land cover. However, high dimensionality and missing data present significant challenges for classification algorithms that use multi-temporal remotely sensed data. Further, generation and assessment of high quality training data, including detection of outliers and changed pixels in training data, is difficult. In this paper we present a new statistical framework that is based on a parametric model that enables a targeted principal component analysis (PCA) to reduce the dimensionality of multi-temporal remote sensing data. In doing so, the model provides a novel basis for land cover classification and evaluating the nature and quality of training data used for supervised classifications. The methodology we describe uses a Kronecker operator to reduce the spectral dimensionality of multi-temporal images while preserving their temporal structure, thereby providing low-dimensional data that is well-suited for classification and outlier detection problems. As part of our framework, we use an expectation–maximization method to impute missing data, and propose new metrics that characterize the representativeness and pixel-to-pixel homogeneity of training sites used for supervised classification. To evaluate our approach, we use data from NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS) and extracted more than 200 training sites where the land cover has been characterized from high spatial resolution imagery. The original input data was composed of 196 features (28 dates × 7 bands), and the PCA-based approach we describe captured 91% of the variance, in these 7 bands, in 3 components. Results from maximum likelihood classification show that the retained principal components successfully distinguish land cover classes from one another, with classification results that were comparable to supervised machine learning methods applied to the original MODIS data. Analysis of our site composition metrics show that they successfully characterize the homogeneity (or lack thereof) and representativeness of individual pixels and entire sites relative to other training sites in the same class.