A Comparison of Geostatistical Algorithms for Incorporating Satellite Imagery into the Mapping of Volume
ABSTRACT The Forest Inventory and Analysis group of the USDA Forest Service is moving from a periodic survey, conducted approximately every 10 years, to an annual inventory. The ground sampled plots will be on a systematic hexagonal grid in the annual inventory. There will be at least one ground sampled plot per hexagonal grid area. The goal is to visit 1/5 of the hexagonal grid points each year. In both cases, a phase one sample is used for stratification. Two types of phase one samples are being investigated: aerial photography and Landsat thematic mapper (TM) satellite imagery. Aerial photography is used by placing a grid over the region of interest, i.e. a state. The grid points are photo interpreted as to whether they are forested or nonforested. The forested plots are further photo interpreted by volume class. Field crews visit a subset of the photo-interpreted plots and record information on the status of the resource. An annual inventory provides current information every year about the resource. However, the cost of an annual inventory is higher than for a periodic inventory. To reduce the cost, we are investigating whether satellite imagery can be used for stratification in lieu of aerial photography, which varies in quality, scale, and date. Inconsistencies naturally exist between different photo-interpreters, with regard to correct location of sample points on a photo and correct vegetation classification. Also, the process is expensive and slow and aerial photography can only be acquired every 7 years, whereas satellite imagery is available yearly. In this paper, we investigate whether remote sensing, specifically TM, used as an auxiliary variable can improve volume class estimation. The primary variable is either cubic foot volume or a surrogate: percent forest, average diameter at breast height, average basal area, trees per acre, or quadratic mean diameter. The potential secondary or auxiliary variables are: Landsat TM Band 4, tasseled cap, and normalized difference vegetation index (NDVI). The goal is to interpolate cubic foot volume or one of its surrogates at unsampled locations in space. This requires techniques from the field of geostatistics. Preliminary results comparing kriging and cokriging indicate that a type of Markov model (MM2) for colocated cokriging is superior to both ordinary kriging and cokriging. The MM2 model resulted in a 44.39% decrease in the mean square error between predicted and ground truth samples. Cokriging is also superior in cross-validation studies. Other potential geostatistical techniques include simulation and kriging with a trend model. These techniques are applied to Connecticut, a highly fragmented state, and can be extended to other regions. |