Automatic Detection of Local Cloud Systems from MODIS Data

Abstract

This paper describes an algorithm that is aimed at the identification of cloudy and clear pixels in Moderate-Resolution Imaging Spectroradiometer (MODIS) images to support earth science and nowcasting applications. The process from geolocated and calibrated data allows one to obtain cloud masks with four clear-sky confidence levels for five different cloud system types. The technique has been developed using the MODIS cloud-mask algorithm heritage, but the threshold tests performed have been executed without comparing solar reflectances and thermal brightness temperatures with thresholds determined in advance, but instead with thresholds carried out from classification methods. The main advantage of this technique is that the thresholds are obtained directly from the images. Seventy-five percent of the spectral signatures (known as end members) derived from the winter images in the detection of the various cloud types and 80% of the summer end members can be considered as being well discriminated. Furthermore, it seems that the end members characterizing the different cloud systems are constant throughout the various seasons of the year (they vary with a confidence level of 60%), whereas those describing clear sky change in a notable manner (the associated confidence level is 99%). The algorithm is able to produce cloud masks pertinent to limited regions at a mesoscale level, which may be a key factor for nowcasting purposes. This work shows that the use of end members and spectral angles, as opposed to spectral thresholds, should be carefully examined because of the fact that it might be simpler or that higher performances may be achieved at a regional scale.