Generating cloudmasks in spatial high-resolution observations of clouds using texture and radiance information

The detection of clouds in measurements taken by airborne and spaceborne remote sensing sensors in the visible and near-infrared is often difficult due to the high albedo of underlying surfaces such as snow- and ice-covered surfaces as well as sunglint regions of water surfaces. The authors show that the measured intensity of the reflected solar radiation together with texture information is effective in detecting clouds over water surfaces which are affected by sunglint. An automated cloud-masking technique for images measured by a compact airborne spectrographic imager (casi) during the ACE-2 CLOUDYCOLUMN experiment has been developed based on supervised learning of an artificial neural network. The neural network has been trained on radiances, texture features, and gradient-filtered radiances. The radiances were measured at a single wavelength but with high spatial resolution so that characteristic spatial features within an image can be used to discriminate clouds from sunglint, cloud shadow and ocean surface. The accuracy of the cloudmask-generating algorithm was investigated on the basis of the testing set for the neural network. Maximum errors of 3.4% and 1% occur for detecting cloudy and cloud-free pixels, respectively. The performance of the network was compared with a second network trained on radiances alone. The second network is up to 44% less efficient for cloud detection which demonstrates the improvement arising from the use of texture information together with spatial high-resolution observations.