|
MEMO -TASK 89.1
<br />classification, the overall accuracy was reported to be between 80 to 92 percent for the districts. There is
<br />no specification of accuracy for individual crops in the Western Slope water districts, which have a very
<br />high proportion of grass pasture, between 80 and 88 percent.
<br />2.2 Other classifications of irrigated lands
<br />In a study conducted by Maxwell and Hoffer (1996) within the Poudre River Basin 25 kilometers east of
<br />Fort Collins, Colorado, eleven different crops were evaluated for various combinations of one, two, or
<br />three date classification of Landsat TM imagery and Colorado Agricultural Statistics. In general, the
<br />multi-date classifications obtained higher absolute overall accuracies, 90.5 percent, as compared to best
<br />single date classification, which obtained an overall accuracy of 82 percent. In particular, the May-
<br />September two-date classification was most effective in discriminating spring, midsummer and late
<br />summer maturing crops. Nevertheless, the high accuracies obtained in this study can be explained in part
<br />by the relatively small size of the study area (i.e., 16 by 16 miles), as well as by the very detailed field
<br />sampling conducted. On the other hand, the authors suggest that a thorough selection of additional
<br />imagery dates may contribute to a better discrimination of hard-to-classify crops, like alfalfa, winter
<br />wheat spring grains, beans and onions.
<br />An examination of the effectiveness ofmulti-temporal classification of various crops (i.e., potatoes, sugar
<br />beet, wheat, fallow, onions) carried out in the United Kingdom by Vieira and Matter (2000) showed that
<br />multi-temporal data (i.e., a combination of 4 dates of Landsat TM and 4 dates of SPOT data) significantly
<br />increases classification accuracy as compared with the use of any single data image.
<br />In 1997, the Northern Colorado Water Conservation District (NCWCD) mapped irrigated parcels using
<br />single date Landsat TM and IRS satellite imagery. The 6-meter ground resolution panchromatic IRS
<br />imagery was used for manual mapping of parcel boundaries, while the 28-meter multi-spectral Landsat
<br />imagery was used for interpreting irrigation status and crop type. No accuracy assessment was performed
<br />in this survey (Scott Bartling, 2002, personal communication).
<br />Other classifications of irrigated lands in the Western United States include a U.S. Bureau of Reclamation
<br />(USBR) study by Williams and Eckhardt (1997). The authors developed a methodology to inventory
<br />irrigated lands in the Upper Colorado River Basin using spectral-space decision thresholds derived from
<br />Tasseled Cap transformation of Landsat TM imagery. The image data was used to update a previously
<br />developed map of field boundaries to produce the irrigated lands inventory. The land cover classes
<br />mapped included cloud-covered, water, wet soil, moderate vegetation, vigorous vegetation, and not
<br />irrigated in 1995. Overall map accuracy for the irrigated areas was estimated to be 95.4 percent, however,
<br />this study did not provide accuracy information for specific crop types.
<br />Congalton et al. (1998) produced maps of the agricultural crops in the Lower Colorado River Basin using
<br />Landsat TM imagery. For this purpose an automated signature extraction process and data exploration
<br />techniques were developed. This process used a combination of ESRI ArcInfo, ERDAS Imagine and
<br />image segmentation algorithms to produce the training sites. The final crop classification labels were
<br />determined by the anticipated use of the maps (i.e., a model to determine consumptive water use). The
<br />maps were produced four times a year, each with an adjusted overall accuracy of 93 percent and higher.
<br />The high accuracy obtained was possible in part through acreage weighting techniques as well as some
<br />adjustments and corrections made in the accuracy table to compensate for confusion among the least
<br />accurate crop types. It not clear if these adjustments were made because the project dictated that the
<br />accuracy of the agricultural crop classification should be at least 93 percent in order to meet the
<br />requirements of the Lower Colorado River Accounting System (LCRAS) model.
<br />In a study conducted in thirteen counties in Nebraska cover 3.35 million hectares, Maxwell et al. (1999)
<br />classified Landsat TM pixels corresponding to agricultural land into their likelihood to be occupied by
<br />corn (i.e., highly likely to be corn, likely to be corn, or unlikely to be corn). According to the
<br />investigators, ground reference data from three counties were used to assess the accuracy of the
<br />Page 3 of 13 ~-~~versfde TecAnafagy, lnc.
<br />W:rler Resources Frr~mec.•rng ,end L'nr+surlrn2
<br />
|