See more from this Session: Sensor-Driven Digital Soil Mapping: II
Monday, November 1, 2010
Long Beach Convention Center, Exhibit Hall BC, Lower Level
The eminently qualitative character of methods in traditional soil survey has driven the development of quantitative models to describe, classify and study the distribution of soils in the landscape. The aim of this study was to evaluate the efficacy of artificial neural networks (ANN) for the construction of a predictive model of soils in the Tibagi’s watershed. As data sources orbital remote sensing, geological and soil pre-existing maps, drainage network and altitudes of the site were used. The methodological approach consisted of two stages: pre-processing the input data and, selecting the method of ANN for the digital classification. The selected ANN presented a Feedforward and algorithm used was the backpropagation. With the result of the digital classification the Crosstab was performace and thus obtain an image and the Cramer's V coefficient found. Once the best result in Cramer's V was found, those areas with coincident soil classes were scanned, resulting in another image that allowed the comparison of existent areas in the original soil map and the ones found by ANN. Finally, the influence of input data in the distribution of mapping units was verified. The results showed that the ANN allowed us to recognize the existing soils and identified three kinds of soil with more than 65% efficacy among five orders defined by the original soil mapping. Of the total area of Entisols, 83% was recognized , 69% of the area dominated by Oxisols and Ultisols Red-Yellow-Red Yellow was also identified by the network. In the case of Oxisols, 41% of areas were identified by the ANN. The ANN had more difficulties in recognizing areas dominated by Inceptisols (39% efficiency). ANN can be employed to identify patterns of soil in order to minimize the subjectivity of traditional soil surveys.
Indexing terms: Pedometry, GIS, Remote Sensing.