Predictive mapping with small field sample data using semi‐supervised machine learning

Abstract

Abstract Existing predictive mapping methods usually require a large number of field samples with good representativeness as input to build reliable predictive models. In mapping practice, however, we often face situations when only small sample data are available. In this article, we present a semi‐supervised machine learning approach for predictive mapping in which the natural aggregation (clustering) patterns of environmental covariate data are used to supplement limited samples in prediction. This approach was applied to two soil mapping case studies. Compared with field sample only approaches (decision trees, logistic regression, and support vector machines), maps using the proposed approach can better capture the spatial variation of soil types and achieve higher accuracy with limited samples. A cross validation shows further that the proposed approach is less sensitive to the specific field sample set used and thus more robust when field sample data are small.