Finding cloud-free satellite imagery for large-scale ecological studies is a notoriously difficult task. The obstacles include clouds, haze, extraneous phenological changes between dates, and data corruption. These problems are especially difficult to resolve in humid tropical forests, where many images are either partially or totally obscured by cloud cover during much of the year. This inhibits our ability to systematically study the great changes underway in the humid tropics-- deforestation, the establishment of agriculture, and forest degradation. As the international community moves to preserve what is left of tropical forests worldwide, clean, clear data across large areas will be critical.
The entire image base of Landsat imagery has recently been released to the public, and this family of satellites would seem to hold promise for long-term low-cost studies of forest cover. Adding to the difficulties already posed by cloud cover, Landsat 7 experienced a permanent failure in 2003 that has caused its images to contain more than 20% black stripes of null data. Meanwhile, Landsat 5 data covers only part of North America. For many users, data from other satellites are not practical: the other free satellites of the MODIS family are too coarse to reliably detect land cover change in the tropics, and radar data, while extremely useful in peering through clouds, is costly. Across huge areas like the island of Borneo, our principal study region, recent data is, for all practical purposes, unusable.
Results/Conclusions
In this project we present a simple, accessible algorithm for the automatic creation of high-quality Landsat image mosaics that are both free of clouds and of the stripes of missing data. Our algorithm uses R and Fortran scripts and, depending on the data size, can be employed on a desktop computer. We use elements of existing “best-pixel” algorithms, histogram equalization, and spectral thresholding to find the best combination of images in a given area to produce a largely cloud-free, stripe-free mosaic. Because of the very large amount of data needed to produce such a mosaic for the island of Borneo, we employed an optimization algorithm to test for and record the best threshold values among multiple bands in hundreds of images. In this presentation, we will show results of the mosaicking process, the behavior of the optimization algorithm, and the image thresholds data base we have created to aid others in this crucial effort.
