Aarhus University Seal

Amélie Marie Beucher

Soil map disaggregation improved by soil-landscape relationships, area-proportional sampling and random forest implementation

Research output: Contribution to conferenceConference abstract for conferenceResearch

Detailed soil information is often needed to support agricultural practices, environmental protection and policy decisions. Several digital approaches can be used to map soil properties based on field observations. When soil observations are sparse or missing, an alternative approach is to disaggregate existing conventional soil maps.
At present, the DSMART algorithm represents the most sophisticated approach for disaggregating conventional soil maps (Odgers et al., 2014). The algorithm relies on classification trees trained from resampled points, which are assigned classes according to the distribution of soil types within the polygons. Vincent et al. (2016) successfully implemented soil-landscape relationships into the algorithm.
The study tests the sensitivity of the algorithm towards the input data by using conventional soil maps for Denmark produced by Jacobsen (1984) and the Commission of European Communities (CEC, 1985) respectively, both using the FAO 1974 classification. Furthermore, the effects of implementing soil-landscape relationships, using area proportional sampling instead of per polygon sampling, and replacing the default C5.0 classification tree algorithm with a random forest algorithm were evaluated. The resulting maps were validated on 777 soil profiles situated in a grid covering Denmark.
The experiments showed that the results obtained with Jacobsen’s map were more accurate than the results obtained with the CEC map, despite a nominally coarser scale of 1:2,000,000 vs. 1:1,000,000. This finding is probably related to the fact that Jacobsen’s map was more detailed with a larger number of polygons, soil map units and soil types, despite its coarser scale.
The results showed that the implementation of soil-landscape relationships, area-proportional sampling and the random forest implementation generally improved the algorithm’s ability to predict the correct soil class. The implementation of soil-landscape relationships and area-proportional sampling generally increased the calculation time, while the random forest implementation reduced the calculation time.
In the most successful experiments, the disaggregation provided a good prediction of the soil types, despite the coarse scale of the input maps.
Original languageEnglish
Publication year2017
Publication statusPublished - 2017
EventPedometrics Conference - Wageningen, Netherlands
Duration: 26 Jun 20171 Jul 2017


ConferencePedometrics Conference
Internet address

See relations at Aarhus University Citationformats

ID: 114481539