loading page

Plant Structure and Carbon Storage Assessment Utilizing Drone-Borne Lidar and Deep Learning Technologies in a Danish Agricultural Expanse.
  • +3
  • Katerina Trepekli,
  • Jaime Caballer Revenga,
  • Stefan Oehmcke,
  • Fabian Gieseke,
  • Rasmus Jensen,
  • Thomas Friborg
Katerina Trepekli
University of Copenhagen

Corresponding Author:atr@ign.ku.dk

Author Profile
Jaime Caballer Revenga
University of Copenhagen
Author Profile
Stefan Oehmcke
University of Copenhagen
Author Profile
Fabian Gieseke
Münster University
Author Profile
Rasmus Jensen
University of Copenhagen
Author Profile
Thomas Friborg
University of Copenhagen
Author Profile

Abstract

The increase of vegetation greenness in the Northern latitudes suggests a rise in the fixation of CO2 by photosynthesis, but the observed upward trends in respiration could compensate for elevated uptake by photosynthesis, necessitating the monitoring of variation in vegetation structure and carbon (C) storage at very high spatio-temporal resolution. Compared to passive optical remote sensing, Light Detection and Ranging (Lidar) scanners may improve the quantification of C sink by providing 3D information of plant structures without apparent sign of saturation of spectral response over dense canopies. We evaluate a novel approach to precisely map C sequestration and key metrics describing the 3D canopy structure of a temperate agricultural expanse by implementing drone-borne Lidar scanner technology and deep learning (DL) architectures potentially capable of detecting individual plants and associated geometrical properties while deriving their above ground biomass (AGB) from point cloud datasets originating from the scanner. An intensive aerial and field campaign was carried out over an Integrated Carbon Observation System (ICOS) class 1 station site (60 ha) in Denmark to remotely measure the horizontal and vertical canopy structure at 15-day intervals during the vegetation growing period, and to collect ground truth data of crop growth in terms of height, density, AGB and green area index of more than 1200 plants. The point cloud data are processed using pattern recognition tools to remove noise and classify them to ground and non-ground points. Two DL models specifically designed to handle the irregular structure of raw point clouds are trained to extract features of vegetation by labeling the processed point cloud data; DL’s suitability for assigning semantic information on 3D data representing cropland is assessed by validating them with the field-based observations. In combination with tower-based flux data, the application of Lidar and DL technologies appear to offer a characterization of the dynamic interaction between climatic conditions, vegetation growth, C sink, water and CO2 fluxes suitable to the challenge of assessing the rapidly changing northern landscapes.