The impacts of training data spatial resolution on deep learning in remote sensing

Abstract

Deep learning (DL) is ubiquitous in remote sensing analysis with continued evolution in model architectures and advancement of model types. However, DL is still constrained by the need for extensive training datasets, which are costly and time-consuming to produce. One potential solution is adapting training data annotations from different spatial resolutions, though the feasibility of such an application has yet to be tested. In this study, we explore the effects of using forest boundary training data derived from the 3D Elevation Program (3DEP) at 1.5m resolution and the National Land Cover Database (NLCD) at 30m to compare the effects on DL model performance. Our research covers diverse landscapes across 11 counties in Indiana (∼11,636 km²), developing 36 DL models to assess the impact of spatial resolution, model architectures, land cover, and training chip sizes. Our results show that higher-resolution training data yield more accurate models, regardless of imagery resolution, though the performance gap (F1 score) was limited to ∼2.7% even at its most extreme. We also found significant variation in performance based on land cover, with average F1 scores of 0.923 in homogeneous forested areas compared to 0.684 in complex urban settings. Despite similar training times between data sources, chipping 3DEP data took roughly five times longer. We expect that the findings from this study will assist future research in optimizing the development of DL training datasets, selection of source imagery at the proper resolution given training data availability, and application of appropriate model tuning depending on landscape complexity.