Cloud-Free Satellite Image Mosaics with Regression Trees and Histogram Matching

Cloud-free optical satellite imagery simplifies remote sensing, but land-cover phenology limits existing solutions to persistent cloudiness to compositing temporally resolute, spatially coarser imagery. Here, a new strategy for developing cloud-free imagery at finer resolution permits simple automatic change detection. The strategy uses regression trees to predict pixel values underneath clouds and cloud shadows in reference scenes from other scene dates. It then applies improved histogram matching to adjacent scenes. In the study area, the islands of Puerto Rico, Vieques, and Culebra, Landsat image mosaics resulting from this strategy permit accurate detection of land development with only spectral data and maximum likelihood classification. Between about 1991 and 2000, urban/built-up lands increased by 7.2 percent in Puerto Rico and 49 percent in Vieques and Culebra. The regression tree modeling and histogram matching require no manual interpretation. Consequently, they can support large volume processing to distribute cloud-free imagery for simple change detections with common classifiers. Introduction Persistent cloud cover over many regions complicates remote sensing with optical satellite imagery. Applications may require cloud-free parts from many scene dates for each vegetation map or for each of the times that bound a change detection. The variously dated scenes or cloud-free scene parts that might compose an image mosaic will differ in atmospheric conditions, sun-target-sensor geometry, sensor calibration, soil moisture, and vegetation phenology. These differences cause the relationships between land-cover classes and pixel brightness values to vary across space over a mosaic period, which refers to the time period spanning the cloud-free scenes or scene parts that compose an image mosaic. One solution to variable relationships between landcover classes and their spectral signatures is to separately classify scene dates (Achard and Estreguil, 1995; Cohen et al., 2001; Helmer et al., 2002). Depending on the number of scenes and their degree of cloudiness, however, a scene wise approach to vegetation mapping or change detection may not be practical. For example, where cloud-free imagery is common, scene footprints yield a regular arrangement of Cloud-Free Satellite Image Mosaics with Regression Trees and Histogram Matching E.H. Helmer and B. Ruefenacht between-scene differences across space. This regular arrangement leads to a regular arrangement of the unique combinations of scene dates through time. Consequently, change detection can occur piecewise across space (Cohen et al., 2002). Piecewise change detection is analogous to scene wise image classification as a solution to between-date radiometric differences across a project area. Insofar as cloud cover determines an irregular arrangement of scene dates across space for each mosaic period, unique combinations of scene dates through time form a complex patchwork. Piecewise change detection becomes infeasible where clouds are persistent. Reducing across space these radiometric and phenological scene differences could permit change detection with one seamless image mosaic for each temporal endpoint. If the land-cover changes of interest are spectrally non-subtle, the image mosaics that bound change detection intervals may not require radiometric normalization to each other (Cohen et al., 1998; Song et al., 2001). Other alternatives exist to mosaicing cloud-free parts of scenes, but they have limitations. The spatial distribution or optical depth of clouds, along with the spatial complexity of land cover, limit how well geostatistical interpolation can predict cloud-obscured pixel values or land cover (Rossi et al., 1994). Microwave satellite imagery, which clouds do not obscure, provides an alternative to optical imagery. Yet, land-cover discrimination with microwave imagery can benefit from optical data (Rignot et al., 1997). A method to create cloud-free optical image mosaics expands the options available for satellite-based remote sensing. Approaches for cloud removal include satellite image compositing, which selects pixels for an output, composite image that are least likely to have cloud cover from among scenes acquired over a compositing period (Gatlin et al., 1984; Holben, 1986). A common compositing criterion is to select pixels for the output image with the largest values of the normalized difference vegetation index [NDVI (Near Infrared Red)/(Near Infrared Red)]. A drawback of image compositing is residual cloud contamination, but excellent methods exist that detect or correct for cloud contamination in composite images (Gutman et al., 1994; Cihlar et al., 1996), or haze in Landsat scenes (Zhang et al., 2002). For imagery with high temporal resolution, Cihlar et al. (1996), for example, detect cloud-contaminated pixels in composites with four thresholds. These thresholds include (a) the maximum red band reflectance that is present in the data set from snow and ice-free land under clear sky, (b) positive and negative deviations from the expected, median NDVI for a PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Sep t embe r 2005 1079 E.H. Helmer is with the International Institute of Tropical Forestry, USDA Forest Service, 1201 Calle Ceiba, Jardín Botánico Sur, Río Piedras, PR 00926-1113 ([email protected]). B. Ruefenacht is with Red Castle Resources, Inc., USDA Forest Service-Remote Sensing Applications Center, 2200 West, 2300 South, Salt Lake City, UT 84119-2020 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 71, No. 9, September 2005, pp. 1079–1089. 0099-1112/05/7109–1079/$3.00/0 © 2005 American Society for Photogrammetry and Remote Sensing 03-124.qxd 8/6/05 4:53 PM Page 1079