Digital terrain modeling has revolutionized the way topography is characterized and analyzed. Its applicability has widened to almost anything where topography has a role to play. On the other hand, digital soil mapping has become the pedological paradigm of the time as it is making tremendous improvements in the ways soil information is obtained, stored, retrieved and manipulated. This research was conducted in Vestfold County of south-eastern Norway to use digital terrain analysis aided by statistical modeling and remote sensing image classification algorithms to make digital soil maps.
A digital elevation model of 25 meter resolution and digitized soil map of part of the study area accompanied by data on some analytical properties of soils were used as original data for the terrain and soil respectively. Fifteen terrain attributes were derived from the digital elevation model through digital terrain analysis. There were thirteen WRB soil classes in the surveyed area of the study site. Besides, five most important topsoil properties (the soils content of Clay, Organic carbon, Keldjahl’s Nitrogen, KNHO3- and pH) for limited number of soil profiles were also used.
The relationship between soil properties and the terrain attributes were analyzed using multiple linear regression in SPSS. The significant regression models were then fed into ARCGIS to predict the spatial distribution of the soil properties. The performance of this prediction was evaluated by comparing it with validation-based ordinary kriging interpolation of the soil properties, which was conducted in ARCGIS. The prediction of soil classes using digital terrain analysis was conducted using two conceptually different approaches. First, soil classes were considered as discrete objects and analysis of variance was used to check if there was significant difference among them in their terrain attribute values. Then, in analogy with satellite image channels, the terrain attributes were used as channels and object-oriented supervised classification algorithm was applied in eCgnition by collecting training areas from the reference soil map. To know the relative performance of this object-oriented approach, ordinary pixel-based supervised classification was conducted in ARCGIS using the same training areas. Second, the spatial variation of soil classes was conceptualized as gradual and fuzzy logic approach was employed for the prediction. Here, the relationship between the soil classes and the terrain attributes was first modeled using multinomial logistic regression in SPSS to identify the most influential terrain attributes and to construct logit models for each soil class. The logit models were used to derive probability prediction models which were then used in ARCGIS to predict the probability of existence of each of the soil classes as fuzzy variables. The reliability of this approach was evaluated qualitatively using expert knowledge, empirical soil map of the area and theoretical background of the soil classes, and quantitatively through correlation study of the probability values.
The result from the spatial prediction of topsoil properties using terrain attribute showed that the approach predicted topsoil clay content, KHNO3 content and extractible nitrogen content with better accuracy compared to the validation-based ordinary kriging. Besides, it showed that about 60% of each of their spatial variation can be attributed to terrain. On the other hand, insignificant correlation was found between the terrain attributes and organic carbon content and pH of the soils of the area.
All of the terrain attributes, with the exception of plan curvature, were found significantly influential in the spatial distribution of soils both by the ANOVA and the logistic regression analysis. Elevation, flow length, duration of daily direct solar radiation, slope, aspect and topographic wetness index were found to be the most significant terrain attributes. The crisp approach to the prediction of soil classes showed that the object-oriented approach performed better than the pixel-based terrain classification approach. The overall accuracy for the object-oriented approach was 30% while it was only 14% for the pixel-based. However, the accuracies of some soil classes reached up to 75% in the first approach. Higher accuracies were obtained for soil classes with higher spatial coverage in the area. The probability prediction for each soil class using logit models was found to be reliable when evaluated against the empirical soil maps except for those soil classes which are not greatly influenced by topography but by other factors such as human activity.
In general, the study revealed that digital terrain analysis has a great potential in digital mapping of soils and their properties. Fuzzy probability mapping and object-oriented approach were found to be reliable to a considerable extent in the prediction of soil classes and deserve further research and application.