LIOUSYMP Cloud Masking without Thermal Infrared Bands: A Neural Network Approach Based on MODIS, As a Candidate for the NASA PACE Mission

Cloud masking (i.e., identifying which pixels contain clouds and which do not) is the first step in processing satellite imagery to geophysical data products across many satellite sensor types and research domains. Capabilities to detect clouds, and tolerances to cloud contamination from underscreening, vary widely between sensors and downstream algorithm needs. This mean there is no “one size fits all” cloud masking algorithm, although similar sensor types often use similar principles.

Passive multispectral imagers such as the MODerate resolution Imaging Spectroradiometers (MODIS) tend to make use of the fact that clouds are bright (reflectance is high), white (small spectral variation across the visible), bumpy (spatially heterogeneous reflectance), and cold (low brightness temperature) relative to the underlying surface. These cloud masking algorithms are often, essentially, series of tests on the observed radiance with human-specified or human-guided thresholds. NASA’s Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission (https://pace.gsfc.nasa.gov) will have launched by the time of this presentation and its primary sensor, the Ocean Color Instrument (OCI), will make hyperspectral measurements from the ultraviolet through near-infrared, with discrete bands in the shortwave infrared. This includes the range of solar wavelengths sampled by MODIS, although OCI will not make thermal infrared (TIR) measurements. Thus, some of the MODIS cloud mask tests relying on thermal bands will be inapplicable.

Initial experiments based on the MODIS cloud mask algorithm tests without TIR bands did not perform satisfactorily. Therefore, we asked: how much of the skill in cloud masking from MODIS’ TIR bands can be found in its solar bands? A multi-layer perceptron network is conceptually similar to the approach traditionally taken, and these types of machine learning methods have achieved good predictability in other work. We used the standard MODIS Aqua cloud mask product (MYD35) as labelled training data to develop a neural network cloud mask taking as input geometry, surface type, and MODIS reflective solar band measurements. Comparison against the standard MODIS cloud mask reveals agreement in more than 90% of pixels, much better than the “solar tests only” stripped-down approach we tested initially. This implies that some of the information content for cloud masking in MODIS TIR bands can indeed be obtained purely from its solar bands. Due to similarity in pixel size, spectral channels, and geometry, it is anticipated that this neural network mask could also be successful when applied to PACE OCI measurements. We present progress to date on this research.