Investigating the impact of classification features and classifiers on crop mapping performance in heterogeneous agricultural landscapes

Abstract

Timely and accurate mapping crops is essential for agricultural management, policy making and food security. The smallholder agricultural systems lead to large number of fragmented and heterogeneous landscape, making fine crop mapping a huge challenge. Feature and classifier selection are two important influencing factors in crop classification. However, there are few systematic tests to determine the specific features and classifiers needed for heterogeneous agricultural landscapes. In this study, 24 candidate spectral features, 8 spatial texture features from the red-edge (RE) and near-infrared (NIR) bands, and 4 supervised classifiers (i.e. random forest (RF), support vector machine (SVM), artificial neural network (ANN), and maximum likelihood classifier (MLC)) were used for crop mapping. 60 spatially heterogeneous landscapes in Heilongjiang Province, northeastern China were selected as the study areas, evaluated by compositional heterogeneity (homogeneity index, HOI) and configurational heterogeneity (splitting index, SPLIT). The results summarized a look-up table for searching the optimum classification features and classifiers in different landscape, providing a reference for future crop classifications and preventing the consumption of computing time. The results revealed that (1) an optimal feature-subset (with reduced data volume by 65%) can achieve high-accuracy crop mapping in heterogeneous regions. (2) The optimum type and number of features and classifiers are landscape sensitive. When a specific accuracy was required, homogenous regions need a smaller number of features and a simple MLC could meet the requirement. (3) The impact of configurational heterogeneity on textural features is more significant, while compositional heterogeneity performs better on spectral Vegetation Indices (VIs). Findings from this study provide a general guideline for crop mapping in plains or fragmented landscape and areas with single or complex planting structure.