Generating a calibrated image from an optical sensor involves several processing steps, like debayering, pixel calibration, lens correction, and more. This preprocessing is typically applied to the raw data directly by a hardware component close to the sensor. However, in specific scenarios, greater control over the preprocessing pipeline is desired, like when the data must be high-quality for science or when the data must be geo-corrected (georectification, orthorectification, and georeferencing).

Thus, data collected by earth observation satellites are traditionally sent to the ground for processing, requiring large transmission budgets and adding a longer turn-around time to react to interesting observations. Especially in time-critical scenarios like quasi-real-time disaster detection and recovery, preprocessing and analyzing the data onboard is desired to reduce transmission budgets and turn-around time for decision-making.

In this joint research project with armasuisse S+T, we designed a computation-graph-based image preprocessing pipeline using TensorFlow. We chose TensorFlow to keep the ability to flexibly re-arrange the pipeline, change and evaluate different preprocessing methods, and compare their results. We also compiled a trade-off analysis of other methods and hardware combinations, including image calibration, geocorrection, and geoprojection.