Identifying and mapping individual plants in a highly diverse high-elevation ecosystem using UAV imagery and deep learning

Zhang, Ce and Atkinson, Peter and George, Charles and Wen, Zhaofei and Diazgranados, Mauricio and Gerard, France (2020) Identifying and mapping individual plants in a highly diverse high-elevation ecosystem using UAV imagery and deep learning. ISPRS Journal of Photogrammetry and Remote Sensing, 169. pp. 280-291. ISSN 0924-2716

Text (ISPRS_paramo_paper_20200930_accepted)
ISPRS_paramo_paper_20200930_accepted.docx - Accepted Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (8MB)

Abstract

The identification and counting of plant individuals is essential for environmental monitoring. UAV based imagery offer ultra-fine spatial resolution and flexibility in data acquisition, and so provide a great opportunity to enhance current plant and in-situ field surveying. However, accurate mapping of individual plants from UAV imagery remains challenging, given the great variation in the sizes and geometries of individual plants and in their distribution. This is true even for deep learning based semantic segmentation and classification methods. In this research, a novel Scale Sequence Residual U-Net (SS Res U-Net) deep learning method was proposed, which integrates a set of Residual U-Nets with a sequence of input scales that can be derived automatically. The SS Res U-Net classifies individual plants by continuously increasing the patch scale, with features learned at small scales passing gradually to larger scales, thus, achieving multi-scale information fusion while retaining fine spatial details of interest. The SS Res U-Net was tested to identify and map frailejones (all plant species of the subtribe Espeletiinae), the dominant plants in one of the world’s most biodiverse high-elevation ecosystems (i.e. the páramos) from UAV imagery. Results demonstrate that the SS Res U-Net has the ability to self-adapt to variation in objects, and consistently achieved the highest classification accuracy (91.67% on average) compared with four state-of-the-art benchmark approaches. In addition, SS Res U-Net produced the best performances in terms of both robustness to training sample size reduction and computational efficiency compared with the benchmarks. Thus, SS Res U-Net shows great promise for solving remotely sensed semantic segmentation and classification tasks, and more general machine intelligence. The prospective implementation of this method to identify and map frailejones in the páramos will benefit immensely the monitoring of their populations for conservation assessments and management, among many other applications.