Geoscience Reference
In-Depth Information
agricultural use in a previous study (Torbick
2004
) CAUV proved to be the most
accurate representation of agricultural land use because data were consistent
between the counties and within each dataset.
The parcels identified as agriculture were used to apply a mask to the Landsat
5 data. A Maximum Likelihood classification was performed on the nonagricultural
data by first selecting 22 training sites representing eight land cover types: imper-
vious, industrial/highly reflective surfaces, forest, shrub, grass, bare soil, water, wet
vegetation. The training sites were identified using the DOQ and were selected
based on their ability to accurately represent each land cover type. Because coll-
ection of training sites is a highly subjective process, accuracy was ensured by
(1) selecting only those sites that, upon visual inspection, exhibited land cover
agreement between the Landsat data and the DOQ and (2) selecting only homoge-
neous areas. Based on the training sites, the image was classified using a maximum
likelihood algorithm. Industrial and impervious were then combined to produce an
impervious class. Forest was combined with shrub to produce a general forest class.
Water and wet vegetation were combined into one class as well.
Agriculture land was classified using the above methods. Twenty-seven training
sites were selected representing green agriculture (winter wheat), early agriculture
(corn or soy), dry bare soil, wet bare soil, forest, wet vegetation, and water. These
classes were also combined to produce four general land cover types: agriculture,
forest, wet vegetation and water.
10.3.5 Accuracy Assessment
To test the accuracy of a land cover classification, stratified random sample of
500
validation pixels (30-50 per class for each of the two classifications) was classified
and compared against the actual classification. To reduce bias in sampling, these
points were collected separately from the original set of training data used for
the classification. Each of the random pixels was then systematically inspected
within the context of the false color May 1992 Landsat TM image. In many cases,
the pixel's corresponding class could be determined based solely on visual inter-
pretation of the Landsat image. However, each validation pixel was cross checked
with the 1994-1995 1 m DOQs to ensure the correct assignment of each validation
point. The classified validation pixels served as ground-truthed samples and were
then compared to the values of the actual classifications within a confusion matrix
using ENVI. This test produced results for overall accuracy, kappa coefficient,
errors of commission and omission, and producer and user accuracies. Accuracy for
both samples was about 88% (Table
10.1
).
Table
10.2
shows the proportions of correctly classified pixels and the proportion
of misclassified pixels for each class. The most common misclassifications in the
agriculture subset were between green agriculture and forest with 20% of green
agriculture being classified as forest and 6% of forest classified as green agriculture.
Additionally, 8% of bare soil was misclassified as early agriculture. The most
significant misclassification in the non-agriculture dataset was between bare soil
