Canopy extraction of mango trees in hilly and plain orchards using UAV images: Performance of machine learning vs deep learning

Abstract

Mango is an important fruit widely grown in tropical and subtropical regions. Intelligent and accurate pesticide spraying for mango orchard can significantly improve yield and quality of mango. To obtain the information of mango canopy accurately is the key to realize the precision pesticide spraying of mango orchard. However, it is still a challenge to use the remote sensing technology of unmanned aerial vehicle (UAV) to accurately extract canopy information in orchards with different landforms. The visible light images of mango orchards with different geomorphological characteristics were collected by a UAV, and the canopies were accurately extracted, and their canopy areas were accurately predicted based on deep learning method in this study. Firstly, visible light images collected by a UAV were used to segment and extract mango tree canopies using various machine learning (ML) and deep learning (DL) models. Based on their accuracy, the best-performing models, HR-Net from DL and Extra Trees Classification (ETC) from ML were selected. Subsequently, Mixed Dataset-HR-Net and ETC-CHM (Canopy height model) models were developed based on these optimal models, and their performance was evaluated for canopy segmentation and area extraction in four representative regions. Finally, the influences of different environmental factors, datasets, and Elevation features on the models were discussed. The results indicate that under the influence of factors such as terrain variation, shadows, weeds, and planting density, the Mixed Dataset-HR-Net outperformed the ETC-CHM model. Specifically, the ETC-CHM model was simultaneously affected by shadows, weeds, and planting density, achieving an average segmentation accuracy of 85.56 % and an average rRMSE of 14.53 % for canopy area extraction across the four regions. In contrast, the Mixed Dataset-HR-Net, trained on a diverse dataset, demonstrated strong generalization ability and superior canopy extraction performance. It was solely affected by planting density, achieving an average segmentation accuracy of 94.55 % and an average rRMSE of 8.50 % for canopy area extraction across the four regions. The results provide new perspectives for the accurate extraction of fruit tree canopies in different topographies, which can facilitate precision pesticide spraying in orchards.