Fine-scale classification of horticultural crops using Sentinel-2 time-series images in Linyi country, China

Satellite imagery holds great potential for crop mapping. However, the high degree of similarity of remote sensing features and fragmented plots of horticultural crops challenges their fine-scale classification. In this study, we mapped horticultural crops by fusing Sentinel-2 time-series images, object-based method and machine learning. Features including original reflectance, vegetation index (VI), and texture were first constructed from Sentinel-2 images. ReliefF feature selection techniques were then implemented to score the features according to their importance for the classification purposes. Finally, the classification performance of pixel-based and object-based methods was evaluated by combining them with Classification and Regression Tree (CART), Random Forest (RF), and Support Vector Machine (SVM) algorithms. The results indicate that both the segmentation methods yield favorable outcomes, with the pixel-based method achieving a slightly higher accuracy (OA = 83.82 %, Kappa = 0.76) compared with the object-based method (OA = 79.99 %, Kappa = 0.70), but the object-based method is able to delineate the boundaries of certain orchards in detail as well as to ensure the consistency of classification results within the orchard. In addition, the findings reveal that training a random forest model using all features leads to exceptional accuracy, with apple, peach, and persimmon exhibiting the most effective classification performance. The Producer Accuracy (PA) and user accuracy (UA) scores surpassed 80 %. This study employed time-series features and object-based method to perform a multi-objective fine classification of horticultural tree crops, and provided valuable insights for remote sensing classification of visually similar crops in fragmented plot scenes.