Semantic segmentation and quantification of trees in an orchard using UAV orthophoto

Field inspection to determine tree counts in orchards is a common practice, requiring significant time and labor. While imaging systems integrated with the Unmanned Aerial Vehicle (UAV) have provided significant opportunities in recent years, examining the images remains a daunting task. In addition, being able to comprehend the state of the tree from the pictures is a task that requires attention and experience. Deep learning algorithms have shown great potential for counting plants in UAV-derived images. This study presents a convolutional neural network (CNN) architecture for semantic segmentation of trees, shadows and soil in orchards using high resolution orthophoto produced from UAV images. In the accuracy assessment, recall, precision, IoU and F1-Score rates of the tree class were calculated as 97.02%, 87.44%, 85.15% and 91.98%, respectively. In addition, considering the land inventory, it is seen that all 475 trees in the study area are classified. It was concluded that the applied CNN architecture is an effective strategy to replace the traditional land count or visual inspection method to detect the number and location of trees in orchards.

Automated summary

Field inspection of tree counts in orchards is time-consuming, but using imaging systems integrated with UAVs and deep learning algorithms can provide a more efficient and accurate method. The study presented a CNN architecture for semantic segmentation of trees, shadows, and soil in orchards using high-resolution orthophoto. The results showed high recall and precision rates, indicating the effectiveness of the proposed method.