Detection and mapping of Antarctic lichen using drones, multispectral cameras, and supervised deep learning

Abstract

The difficulty of accurately detecting lichens in Antarctic landscapes, due to their fine-scale spatial patterns and low spectral contrast, drives the need for high-resolution drone-based remote sensing imagery to develop and validate robust mapping methods. Few studies have explored the use of remote sensing and deep learning (DL) techniques for mapping and monitoring lichen density in Antarctic regions. This study aims to fill this gap by using multispectral (MS) cameras onboard uncrewed aerial vehicles (UAVs) and DL to detect and map Antarctic lichen through a workflow that enhances detection using a semi-automatic labelling technique based on vegetation indices (VIs). This methodology was validated through a data collection campaign at Robinson Ridge, Windmill Islands, Antarctica in January 2023. Two DL methods were evaluated to classify and map Usnea spp., Umbilicaria and Pseudephebe species (black lichen), moss and non-vegetation: method (1) standalone DL model fitting, namely fully convolutional network (FCN), U-Net, and Deeplabv3+, with semi-automatic labelling thresholding using VIs; and method (2) ensemble stacking by using eXtreme gradient boosting (XGBoost) as the input model, whose predictions are used as features for training a U-Net model. In Method 1, U-Net exhibited the best performance over the other models. Specifically, for Usnea spp., the results demonstrate an intersection over union (IoU) of 84%. Also, the black lichen class obtained an IoU of 86%. In contrast, Method 2, which employed the ensemble stacking technique, demonstrates an IoU of 71% for Usnea spp. and IoU of 75% for black lichen. This study provides promising evidence that using MS cameras on UAVs combined with DL models is an effective approach for detecting and mapping lichen density in Antarctica, though further exploration across diverse regions is recommended to validate its scalability and adaptability.