EVALUATION OF CLUSTERING TECHNIQUES FOR CROP AREA ESTIMATIONUSING REMOTELY SENSED DATA

Finding a suitable clustering algorithm has long been a problem when processing remotely sensed digital data from satellites for crop area estimation. In the PEDITOR software system used by the National Agricultural Statistics Service (NASS), the data presented to a clustering algorithm are usually assumed to represent a single crop or ground cover type. This assumption is justified since NASS has always had a large amount of ground information available from the enumerative surveys used to do area estimation without remotely sensed data. The addition of satellite data is intended to improve the quality of the estimates by providing additional inputs into a regression estimator [1]. The data, usually from Landsat or SPOT satellites, consist of a number of pixels (picture elements). Each pixel represents an area on the ground and has several channels, each representing a scaled value of reflectance in a particular spectral band. The scaling is between 0 and 255 so that each channel takes one byte of storage. Typical numbers of channels are seven for the Landsat Thematic Mapper (TM) and three for the French SPOT multispectral scanner. Mu1titempora1 data are often used to help distinguish crops that may have similar spectral characteristics in a single scene. The mu1titempora1 data consist of two scenes over the same area but from di fferent dates, thus containing twice as much data as the single (unitempora1) scene. In order to reduce the computational burden, a subset of the available channels is often selected and used for processing.