Wf sblm2012 ilf

Wf sblm2012 ilf

• 1.
  Artificial Intelligence –Bridging the Gaps in Soil Mapping Fonseca, I.L.1 , Freire, S.2 , Brasil, R. 1 , Rocha, J. 1 &Tenedório, J.A.2 1 Centro de Estudos Geográficos, IGOT Universidade de Lisboa, Edf. Fac. Letras, , Alameda da Universidade, 1600-214 Lisboa, Portugal. [email protected] 2 Centro de Estudos de Geografia e Planeamento Regional, Faculdade de Ciências Sociais e Humanas da Universidade Nova de Lisboa, Av. de Berna 26-C, 1069-061 Lisboa, Portugal. Scientific Area: 1. Landscape and Land Management Keywords: Soil Digital Maps, Artificial Neural Networks, Geo-Self-Organizing Map, Multi-layer Perceptron Model, Sampling, Spatial Autocorrelation. There is a growing need for digital soil maps at scales suitable for land management and regional planning. Soils modulate hydrological fluxes, regulate ecosystems and play a very important role in mediating the impact of climate change. However, most European countries still lack complete soil map coverage at medium to large scales because soil surveys are very expensive and time consuming. Artificial Neural Networks (ANNs) have started to be applied as a means of fast, cheaply and accurately predicting soils by learning rules that can be extended to unmapped areas. In this study two different types of ANNs (GeoSom and Multi-layer Perceptron) and five sampling strategies are investigated in order to predict soil types across two catchments in Northern Portugal and implement, at a later stage, the best model to complete the soil map coverage of Portugal at 1:100000, which currently stands at c. 45% of the country. Landscape variables that are also factors of soil formation, such as parameters derived from DEMs (altitude, slope, profile and plan curvatures, upslope area, dispersal area, wetness index and potential solar radiation), land use and lithology were combined with existing soil map data to train the ANNs. Results indicate that sampling has a higher impact on model accuracy and performance than the parameterisation of the models themselves because the landscape variables most commonly used to predict soils are spatially autocorrelated. Thus, ANNs learn and converge faster to a better solution if the sampling strategy takes into account that close neighbours are more likely to have similar soil types.