A TEST OF AN ARTIFICIAL NEURAL NETWORKALLOCATION PROCEDURE USING THE CZECH SOILSURVEY OF AGRICULTURAL LAND DATA

摘要

Artificial neural networks (ANN) can be used for the development of models for auto-mated soil allocation to predefined soil units. This chapter tests a minimum input data number for reliable ANN model development, and allocation improvement by including terrain data in the model. Results of the Soil Survey of Agricultural Land (SSAL) carried out in the Czech Republic in the period 1960-1972 were used as soil data. Primary terrain attributes (altitude, aspect, and slope) were used as covariates. Increasing the number of training data leads to better allocation results. Nevertheless, a number of 20-30 input profiles showed to be sufficient for most soil units under study; increasing this number did not bring an important improvement in allocation performance of the models. For a good allocation, the classes should be clearly defined and distinguished from each other. Similarities between soil units (e.g. between Luvisols (LV) and Albeluvisols (AB) in some characteristics) increase the proportion of incorrectly allocated soils. Using auxiliary data should improve the allocation results. Nevertheless, the predictors (both soil attributes and covariates) and their structure should be selected according to what is the most important for soil classes to be predicted. Development of a useful ANN allocation model requires good training-data selection, suitable model structure selection and thorough training and exhaustive validation