A Deep-Learning Workflow for CORONA-Based Historical Land Use Classifications

Historical satellite imagery lacks efficient methods for automated land use mapping, particularly when working with CORONA satellite data from the Cold War era. These highresolution images from the 1960s offer valuable insights into historic land use conditions, but require intensive preprocessing and lack standardized methods for analysis. This study presents an integrated and systematic workflow for preprocessing CORONA imagery and demonstrates an automated approach for land use classification that combines imagery and terrain feature analysis. The preprocessing and classification framework was tested in a 35-kilometer radius area near Boulder, Colorado at the NEON Niwot Ridge Mountain Research Station (NIWO). Results show that georeferencing achieved satisfactory accuracy with a mean absolute error of 5.99 meters. The classification approach that uniquely incorporated terrain variation analysis achieved 89.36% overall accuracy in categorizing different land use types. This combined classification approach significantly outperformed methods that used either CORONA imagery alone (overall accuracy 77.67%) or basic terrain data alone (overall accuracy 79.72%). The combined system showed particular strength in distinguishing between different vegetation types such as forests and grasslands that appear similar in satellite images but have distinct terrain characteristics. The successful integration of terrain feature analysis with panchromatic imagery overcomes traditional spectral limitations of historical satellite data. This framework provides researchers with a robust and replicable method to optimize CORONA imagery for studying landscape transformations over the past 60 years.