A prototype for automation of land-cover products from Landsat Surface Reflectance Data Records
Tuesday, 16 December 2014
Landsat data records of surface reflectance provide a three-decade history of land surface processes. Due to the vast number of these archived records, development of innovative approaches for automated data mining and information retrieval were necessary. Recently, we created a prototype utilizing open source software libraries for automatically generating annual Anderson Level 1 land cover maps and information products from data acquired by the Landsat Mission for the years 1984 to 2013. The automated prototype was applied to two target areas in northwestern and east-central North Dakota, USA. The approach required the National Land Cover Database (NLCD) and two user-input target acquisition year-days. The Landsat archive was mined for scenes acquired within a 100-day window surrounding these target dates, and then cloud-free pixels where chosen closest to the specified target acquisition dates. The selected pixels were then composited before completing an unsupervised classification using the NLCD. Pixels unchanged in pairs of the NLCD were used for training decision tree models in an iterative process refined with model confidence measures. The decision tree models were applied to the Landsat composites to generate a yearly land cover map and related information products. Results for the target areas captured changes associated with the recent expansion of oil shale production and agriculture driven by economics and policy, such as the increase in biofuel production and reduction in Conservation Reserve Program. Changes in agriculture, grasslands, and surface water reflect the local hydrological conditions that occurred during the 29-year span. Future enhancements considered for this prototype include a web-based client, ancillary spatial datasets, trends and clustering algorithms, and the forecasting of future land cover.