|
This paper shows an application of Digital Terrain Analysis for the purpose of enhancing and replacing the photo-interpretation in soil survey. We tested weather the supervised landform classification could be used to replace additional photo-interpretations at county level, i.e. for areas on the order of 1000 km2. Here the basic idea was to use subjective photo-interpretation in small sample areas to train a classifier, and then apply this classifier in the extrapolation, which is analogous to the remote sensing-based land cover mapping. We then assessed the classification accuracy using the error matrix, calculated by comparing the whole API maps and point samples, with the results of classification. In addition, we investigated whether visualization of landform parameters in a false color composite could improve subjective photo-interpretation in the training areas. Finally, we wanted to see if the classifier would identify features or patterns that had been omitted or missed by the photo-interpreter, i.e. details that are not visible on the photos. Here you can download the data sets and view additional illustrations and color graphics assigned to the paper. For complete details see:
Hengl, T. and Rossiter, D.G., 2003. Supervised landform classification to enhance and replace photo-interpretation in semi-detailed soil survey. Soil Sci. Soc. Am. J., 67(5): 1810-1822.
Study area and training:
The study area of 1062 km2 corresponds to the Croatian portion of the historic region of Baranya. It is located in north-eastern Croatia, in the triangle formed by the Danube River to the east, the Drava River to the southwest, and the Hungarian border to the north (centered at N 45° 42' 14', E 18° 40' 35'). The study area consists of number of different landforms and therefore was interesting how will the classification work for different types. The principal soil-forming factors differ between the essentially erosional hilland and depositional plain.
Two methods for selecting training samples were compared. In the first, the entire area of the interpreted photographs, i.e. API maps (in further text whole-API training set). In the second, training samples were created by manual selection on-screen of about 100 pixels within each photo-interpretation unit in the sample areas (in further text point-sample training set). Here by central concept we consider locations, which were in our mental model typical representatives of landform classes when observed stereoscopically. In addition, the photo-interpretation units were displayed as boundaries over false-color composites of synthetic bands and then point-samples checked to ensure that they fall in relatively homogenous facets.
Figure: False color composite made from GWD, SLOPE and CTI (left) and VSHED, DISTW and GWD (right). Different band combinations are suitable for classification of landform elements in hilland and in plain.
Figure: Selecting point-samples using the central concept, training area A.
Classification results:
From total of 167 aerial photos covering the whole study area, we selected six training photos of 2116 ha (4.6 km x 4.6 km) each, totaling 11,079 ha. These were selected subjectively to provide a representative sample of major soil landscapes. Training areas A and F covered sections of Baranja hill, the abandoned course of the Drava, and the edge of the low terrace, while the others (B, C, D, E) covered the terraces and the floodplain. Initial attempts to classify these together, gave results never better than about 50% overall accuracy. The classifications using nine predictors, either as original predictors or their principal components, and all API legend classes showed clear differences between methods but similar overall results. The maximum-likelihood classification gave 45.3% with whole-API training set (Table 6) and 36.8% with point-sample training set overall reproducibility. The corresponding figures for the classification of separate landscapes were 58.1% and 51.6% (hilland), and 39.1% and 34.4% (plain).
We then repeated the classification with the whole-API training set, resulting in overall accuracies relative to the API of 63.4% (whole area), 65.8% (plain), and 58.2% (hilland), i.e. an improvement of 26.7% in the plain and 18.1% overall; the results for the hilland were not affected. After three iterations we were able to achieve high reproducibility for the point sample itself: 90.2% (Kappa=89.3%). Thus we were able to reproduce the classification of the central concept of each landform class.
Figure: Extrapolation of the original API map to the whole hilland. In areas of higher relief, the reproducibility was about 60% of whole API.
This animation shows results of mapping landform classes over the whole study area by using representative sampling areas. In this case, the number of manual photo-interpretations that had to be prepared was reduced from 84 to 6.
| spatial-analyst.net/landforms visited:
5169 times |
|
|
|
){kind=link}