Influence of Compositing Criterion and Data Availability on Pixel-Based Landsat TM/ETM+ Image Compositing Over Amazonian Forests

Persistent cloud cover is an important obstacle to studying the ground surface of tropical rain forest areas using high-resolution optical data, such as those obtained with Landsat satellites. The identification and masking of the cloud-affected parts of such images is a necessary preprocessing step, but it easily leads to impractical fragmentation of the informative image area. Pixel-based multitemporal image compositing solves the fragmentation problem, but depends on a predefined compositing period length and compositing criterion. Here, we evaluate the radiometric consistency of Landsat TM/ETM+ composite images over undisturbed Amazonian forests and test to what degree it varies with the number of available multitemporal observations per pixel and the compositing criterion. Five compositing criteria were tested: maximum NDVI, median red, median near-infrared, multidimensional medoid, and minimum aerosol optical thickness. Each was applied to datasets consisting of 3-30 observations per pixel. Compositing quality was assessed both visually and with quantitative measures using the overlap area of neighboring WRS-2 scenes. We found that the medoid approach generated the most radiometrically consistent composite images. Composite image quality increased monotonically with the number of observations, but with diminishing returns. Satisfactory results were generally obtained with 10-15 observations per pixel.