Exploring the Effect of Balanced and Imbalanced Multi-Class Distribution Data and Sampling Techniques on Fruit-Tree Crop Classification Using Different Machine Learning Classifiers

Abstract

Fruit-tree crops generate food and income for local households and contribute to South Africa’s gross domestic product. Timely and accurate phenotyping of fruit-tree crops is essential for innovating and achieving precision agriculture in the horticulture industry. Traditional methods for fruit-tree crop classification are time-consuming, costly, and often impossible to use for mapping heterogeneous horticulture systems. The application of remote sensing in smallholder agricultural landscapes is more promising. However, intercropping systems coupled with the presence of dispersed small agricultural fields that are characterized by common and uncommon crop types result in imbalanced samples, which may limit conventionally applied classification methods for phenotyping. This study assessed the influence of balanced and imbalanced multi-class distribution and data-sampling techniques on fruit-tree crop detection accuracy. Seven data samples were used as input to adaptive boosting (AdaBoost), gradient boosting (GB), random forest (RF), support vector machine (SVM), and eXtreme gradient boost (XGBoost) machine learning algorithms. A pixel-based approach was applied using Sentinel-2 (S2). The SVM algorithm produced the highest classification accuracy of 71%, compared with AdaBoost (67%), RF (65%), XGBoost (63%), and GB (62%), respectively. Individually, the majority of the crop types were classified with an F1 score of between 60% and 100%. In addition, the study assessed the effect of size and ratio of class imbalance in the training datasets on algorithms’ sensitiveness and stability. The results show that the highest classification accuracy of 71% could be achieved from an imbalanced training dataset containing only 60% of the original dataset. The results also showed that S2 data could be successfully used to map fruit-tree crops and provide valuable information for subtropical crop management and precision agriculture in heterogeneous horticultural landscapes.