A Novel Automatic Learning-based Method of Training Sample Selection Using Multiple Datasets for Time-series Land Cover Mapping

Abstract:

Huge volumes of long-record Landsat image availability offer an opportunity to monitor the earth environment and understand the interactions between climate change and human activities through time. Continuous Change Detection and Classification (CCDC) was developed to detect land cover change using all available Landsat images. Consequently, large amounts of training samples across the geographical region are required to facilitate the progress of land cover mapping with the cooperation of the high dimension temporal features and advanced change detection and classification algorithms. With the implementation of CCDC, the U.S. Geological Survey (USGS) Earth Resources Observation and Science Center has initiated a new revolutionary project, called Land Change Monitoring, Assessment, and Projection (LCMAP), to characterize annual land cover and change for any location across the United States based on Landsat Analysis Ready Data (ARD). LCMAP adopted the USGS National Land Cover Database (NLCD) for the year 2001 (2011 Edition) as the training source from which obvious errors and uncertainties are carried over to LCMAP Collection 1.0 product. The most efficient way to reduce the errors of LCMAP Collection 1.0 land cover product is to improve the quality of training samples. However, the challenge is how to choose reliable training samples based on the multidate temporal information automatically, as a large number of training samples across a large geographic region is required to facilitate the progress of mapping. In this study, we proposed a method that integrates time-series analysis and existing scientific datasets including LANDFIRE's (LF) Existing Vegetation Type (EVT), NLCD, National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL), and National Wetlands Inventory (NWI) to build the training dataset automatically. The hypothesis of the method is that the temporal profiles of samples within the same land cover type are the same or similar at a local scale. Therefore, the method focuses on building the temporal patterns for each land cover category based on the consensus samples among all the mentioned products and calculating the Time-Weighted Dynamic Time Warping (twDTW) distance between the undefined sample and known temporal patterns in the same geographical region as prior knowledge. The twDTW distance is used as an indicator to select the optimal land cover category for each candidate sample from the land cover products based on the designed criteria iteratively. The method was implemented using several Landsat ARD tiles in the eastern and western conterminous United States (CONUS) to produce annual land cover maps from 1985 to 2017. The results were compared with LCMAP Collection 1.0 product. The comparison revealed that the method could generate reliable training samples without any manual interpretation, producing much better land cover results especially for the vegetation types than the current product. More land cover dynamics were captured in the prototype tests. The preliminary results also suggest that the proposed method can improve the quality of land cover product for LCMAP as it takes advantage of the strengths of different products for the training sample selection. Moreover, we have implemented this method for Hawaii with minor edits according to the available products in the region, and the results demonstrate the robustness of the method.