Automated cloud cover assessment for Landsat TM images

This paper describes a method for cloud cover assessment using computer-based analysis of multi-band Landsat images. The objective is to accurately determine the percentage of cloud cover in an efficient manner. The ''correct'' value is determined by an expert's visual assessment. Acceptable error rates are +/-10% from the visually-determined coverage. This research improves upon an existing algorithm developed for use by the EROS Data Center several years ago. The existing algorithm uses threshold values in bands 3, 5, and 6 (red, middle infrared, and thermal, respectively) based on the expected frequency for clouds in each band. While this algorithm is reasonably fast, the accuracy is often unsatisfactorily. The dataset used in developing the new method contained 329 subsampled, 7-band Landsat browse images with wide geographic coverage and a variety of cloud types. This dataset, provided by the EROS Data Center, also specifies the visual cloud cover assessment and the cloud cover assessment using the current automated algorithm. Mask images, separating cloud and non-cloud pixels, were developed for a subset of these images. The new approach is statistically based, developed from a multi-dimensional histogram analysis of a training subset. Images form a disjoint test set were then classified. Initial results ar significantly more accurate than the existing automated algorithm.