Performance of MobileNetV3 Transfer Learning on Handheld Device-based Real-Time Tree Species Identification

Detailed information on tree species constitutes an essential factor to support forest health monitoring and biodiversity conservation. Current deep learning-based mobile applications for tree and plant identification require excessive computation. They largely depend on a network connection to perform computing tasks on powerful remote servers in the Cloud. Many forestry areas are remote with limited or no cellular coverage, which is an obstacle for these applications to recognize trees and plants in these areas in real-time. This paper investigates existing CNN-based machine learning applications for plant identification tailored for handheld device usages. Driven by network independence, reduced computation, size and time requirements, we propose the use of MobileNet (a mobile computer vision architecture) transfer learning to improve the accuracy of offline leaf-based plant recognition. We then carry out experimental validation of state-of-the-art MobileNet. Our findings reveal that using MobileNetV3 transfer learning, accuracy up to 90% can be achieved within fewer iterations than end-to-end CNN-based models for plant identification. The lightweight model comes with reduced computation that runs independently within a smartphone application without internet access, ideal for tree species identification in rural forests.