An efficientnet-based model for classification of oil palm, coconut and banana trees in drone images

Abstract

Oil palm tree detection and classification from coconut and banana trees is vital for increasing the production of oil palm businesses globally, particularly in Malaysia. Since oil palm, coconut, and banana trees share common characteristics such as tree shape and structure, classification is challenging. Further, this work considers images captured by drones, which adds complexity to the classification problem. Unlike most existing methods that primarily detect oil palm trees, the proposed work aims to detect and classify multiple tree types. Inspired by the success of the Segment Anything Model (SAM), a generalized model for object segmentation, we adapted SAM for detecting and localizing oil palm, coconut, and banana trees in drone images. Similarly, motivated by the efficiency and effective feature extraction of EfficientNetB3, we integrated it for the classification task. The proposed model combines SAM for detection and EfficientNetB3 for classification in an end-to-end architecture. To evaluate its performance, we conducted experiments on a dataset collected from a Malaysian drone services company, featuring frames captured across diverse locations. Results demonstrate that the proposed method significantly outperforms state-of-the-art approaches. For detection, the proposed SAM achieves F1-scores of 97 %, 89 %, and 91 % for oil palm, coconut, and banana trees, respectively. For classification, the proposed model achieves F1 scores of 92 %, 88 %, and 91 % for oil palm, coconut, and banana trees, respectively. The results show that the proposed method is superior to the existing methods.