Abstract
Landsat satellite images are subject to cloud cover effects resulting in erroneous analysis and observations of ground features. In this work, we present a novel algorithm (STmask) combining tasseled cap band 4 (TC4) with short wave infrared spectral band 2, SWIR2 (2.107-2.294 mu m) for generating cloud, water, shadow, snow and vegetation masks. A support vector machine (SVM) with a non-linear kernel is trained on a feature space of TC4 versus SWIR2 for generating feature masks. To develop a generic and unbiased algorithm, the SVM is trained using reference data comprised of 12891 pixels from Landsat 8 scenes from ten spatially and temporally diverse biomes including deciduous forest, rainforest, great plain, savanna, desert, ocean, freshwater, taiga, tundra, and icesheet. 960000 text pixels spanning 96 scenes across 8 biomes from the USGS Landsat cloud cover assessment data set are used for accuracy assessment of STmask as well as to compare its performance with the operational Landsat algorithm, C function of mask (CFmask). Using McNemar's statistic, STmask is shown to maximize both the precision and sensitivity of the classification of all features compared to CFmask. It addresses the challenges of CFmask through statistically significant improvement in the precision of cloud detection over snow/ice, barren, water, urban, and shrubland biomes. Aggregated over all biomes, the average improvement in cloud detection over CFmask is observed to be similar to 3.8% using the F-measure. The classification of non-cloud features exhibits promising improvements and mostly comparable performance to CFmask. Overall classification performance is promising, and thus STmask is a novel, biome-independent, parsimonious, and computationally efficient alternative (and/or a cloud screening addition) to the operational CFmask algorithm. The work is timely and is targeted as an innovative processing solution for the land surface remote sensing research community.
NSTL
Elsevier