Radiometric normalization and cloud detection of optical satellite images using invariant pixels

Clouds in optical satellite images can be a source of information for water measurement or viewed as contaminations that obstruct landscape observations. Thus, the use of a cloud detection method that discriminates cloud and clear-sky pixels in images is necessary in remote sensing applications. With the aid of radiometric correction/normalization, previous methods utilized temporal and spectral information as well as cloud-free reference images to develop threshold-based cloud detection filters. Although this strategy can effectively identify cloud pixels, the detection accuracy mainly relies on the successful radiometric correction/normalization and reference image quality. Relative radiometric normalization generally suffers from cloud covers, while multi-temporal cloud detection is sensitive to the radiometric normalization quality. Thus, the current study proposes a method based on weighted invariant pixels for both processes. A set of invariant pixels is extracted from a time series of cloud-contaminated images by using the proposed weighted principle component analysis, after which multi-temporal images are normalized with the selected invariant pixels. In addition, a reference image is generated for each cloud-contaminated image using invariant pixels with a weighting scheme. In the experiments, image sequences acquired by the Landsat-7 Enhanced Thematic Mapper Plus sensor are analyzed qualitatively and quantitatively to evaluate the proposed method. Experimental results indicate that F-measures of cloud detections are improved by 1.1–6.9% using the generated reference images.