An approach for collecting and evaluating land cover training data using time series of Landsat data

Land cover and land cover change are important attributes of terrestrial ecosystems that influence land surface atmosphere energy exchanges, the carbon cycle, and by extension, weather and climate at both regional and global scales. As a result, high quality land cover data is required at large spatial scales and with sufficiently high spatial resolution to realistically capture land cover properties and dynamics. However, such data are often difficult to generate due to limitations associated with collecting high-quality training data required for remote sensing-based mapping.

The main goal of this work is to develop an approach for collecting, evaluating, and enhancing training data used for mapping land cover at continental scale based on long-term time series of Landsat data. To address this challenge, we present a three stage approach consisting of: 1) unsupervised clustering of spatio-temporal features generated from Landsat time series based on a large sample of locations randomly selected across the area of interest; 2) evaluation of the value of existing training data for mapping the region-specific land cover; and 3) supervised identification of locations in the region of interest that are under-sampled in training data and appropriate for additional training data collection.

We tested our approach over 4 areas representing different US EPA Ecoregions, each of which spanned areas defined by 5000 by 5000 grids of Landsat pixels. Our analysis used Landsat time series consisting of all available observations acquired from 1998 to 2018. Using random samples selected within each ecoregion, we extracted spectral and temporal characteristics of land cover at each pixel using the Continuous Change Detection (CCD) algorithm, and assessed the representativeness of training data within each region based on results from an unsupervised clustering algorithm applied to the random sample of locations.

Our results show that this approach is useful for tailoring training data designed to capture regional characteristics of land cover. In particular, our methodology provides an efficient basis for: 1) evaluating the representativeness of training data used in supervised land cover mapping efforts, and 2) efficient augmenting land cover training data in support of ecoregion to continental scale land cover mapping.