Abstract:

A typical fragmented farmland in Taihe County, Fuyang City, which is located in the main soybean producing area in the northern Anhui Province, China, was taken as the study area in the present study. Soybean planting area extraction models were constructed based on the unmanned aerial vehicle (UAV) RGB images and different machine-learning algorithms to perform fine crop mapping. In addition to the relative reflectances of R, G, B bands, the 3 components of HLS color space, 9 visible light vegetation indices, 6 texture features and 1 geometrical feature, were selected as the candidate feature variables. Then, a feature selection method coupled with classifier was employed to single out four optimum feature-subsets for the algorithms of random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost) and BP neural network (BPNN). Corresponding models were subsequently constructed based on the filtered subsets of features and algorithms for mapping, and their performances were examined. It was shown that the optimum feature-subsets of the four algorithms outperformed the original RGB bands in terms of extraction accuracy, among which RF exhibited the best performance, as its overall accuracy was 93.96%, and the Kappa coefficient reached 0.87. In general, the RF algorithm combined with optimum feature-subset showed great potential in soybean planting area extraction based on UAV platform, and the feature selection method could achieve a balance between higher classification accuracy and less data volume, which had certain practical significance.

Key words: unmanned aerial vehicle, machine learning, soybean, crop mapping, feature selection