15 Apr 2026

Satellite Image Quality Classification with ImageNet Transfer Learning and Data Fusion

Abstract: This Coastal and Hydraulics Engineering Tech Note (CHETN) documents the development of a convolutional neural network (CNN) to automate quality control on image classification, a process previously done by subject matter experts (SMEs), within the Littoral Zone Maneuver Support Tool (LZMST). LZMST was created to support rapid exploration of an unknown littoral region by analyzing global satellite data and wave and current models to best estimate the coastal conditions and help identify potential hazards. In support of this mission, images from Landsat-8 (Roy et al. 2014) and Sentinel-2a/2b (Drusch et al. 2012) are graded on their predicted usefulness for LZMST, which is usually done by expert selection. A CNN model is developed to automate this task, by utilizing transfer learning on a CNN using ImageNet (Krizhevsky et al. 2017) weights combined with a small data set of classifications from the CoastSat (Vos et al. 2019) python application. Because the expert selection of images is incredibly time consuming, the data set used to develop this tool was small (approximately 3,500 images), which can make creation of a data-driven algorithm difficult. This CHETN highlights the usefulness of using transfer learning to eliminate the need for large data sets and demonstrates that ImageNet weights can be successfully used to assist in quality detection on multispectral imagery from the Landsat-8 and Sentinel-2a/2b missions.