Advancing remote sensing of biocrusts with drone imagery and machine learning

Biocrusts are a major ground cover type in drylands, driving ecosystem function and contributing to biodiversity at large scales. However, their small size and similar colour to background soils and vegetation make them challenging to monitor with remote sensing. We developed a simple and accurate f...

Full description

Saved in:
Bibliographic Details
Main Authors: Jana Stewart, Roxane J. Francis, David J. Eldridge, Richard T. Kingsford, Nathali Machado de Lima
Format: Article
Language:English
Published: Elsevier 2025-06-01
Series:Geoderma
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S0016706125001533
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850150771814825984
author Jana Stewart
Roxane J. Francis
David J. Eldridge
Richard T. Kingsford
Nathali Machado de Lima
author_facet Jana Stewart
Roxane J. Francis
David J. Eldridge
Richard T. Kingsford
Nathali Machado de Lima
author_sort Jana Stewart
collection DOAJ
description Biocrusts are a major ground cover type in drylands, driving ecosystem function and contributing to biodiversity at large scales. However, their small size and similar colour to background soils and vegetation make them challenging to monitor with remote sensing. We developed a simple and accurate field method for large scale surveys of biocrust, using drone imagery and machine learning, guided by visual ground survey data. We compared the accuracy of three different camera sensors- RGB, multispectral, and thermal. We used XGBoost predictive modelling to classify groundcover into six classes including three biocrust community morphology types (bare ground, cyanobacteria-lichen biocrust, crustose and foliose lichen biocrust, moss biocrust, dead vegetation, live vegetation). Visual ground-based survey data and fine-scale photography were used to ground truth drone imagery to develop training datasets. Modelled outputs demonstrated that Multispectral was the best drone camera sensor type, with the highest accuracy of 97.0 %, with NDVI the most important band for the model. When we applied the model to 50 m2 plots to validate its predictions, we had similar results to visual classification from field surveys and fine-scale photographs, successfully separating biocrust from bare ground. Our relatively simple method can be applied to biocrusts using readily available, low-cost technology. Considerable opportunities exist for using this approach to provide landscape-level biocrust assessment, using remote sensing, leading to improved restoration and management of drylands for conservation.
format Article
id doaj-art-cfb21863a7ea44a4bb661a24534f24e3
institution OA Journals
issn 1872-6259
language English
publishDate 2025-06-01
publisher Elsevier
record_format Article
series Geoderma
spelling doaj-art-cfb21863a7ea44a4bb661a24534f24e32025-08-20T02:26:27ZengElsevierGeoderma1872-62592025-06-0145811731510.1016/j.geoderma.2025.117315Advancing remote sensing of biocrusts with drone imagery and machine learningJana Stewart0Roxane J. Francis1David J. Eldridge2Richard T. Kingsford3Nathali Machado de Lima4Centre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, UNSW, Sydney 2052, Australia; Corresponding author.Centre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, UNSW, Sydney 2052, AustraliaCentre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, UNSW, Sydney 2052, AustraliaCentre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, UNSW, Sydney 2052, AustraliaCentre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, UNSW, Sydney 2052, Australia; School of Biotechnology and Biomolecular Sciences, UNSW, Sydney 2052, AustraliaBiocrusts are a major ground cover type in drylands, driving ecosystem function and contributing to biodiversity at large scales. However, their small size and similar colour to background soils and vegetation make them challenging to monitor with remote sensing. We developed a simple and accurate field method for large scale surveys of biocrust, using drone imagery and machine learning, guided by visual ground survey data. We compared the accuracy of three different camera sensors- RGB, multispectral, and thermal. We used XGBoost predictive modelling to classify groundcover into six classes including three biocrust community morphology types (bare ground, cyanobacteria-lichen biocrust, crustose and foliose lichen biocrust, moss biocrust, dead vegetation, live vegetation). Visual ground-based survey data and fine-scale photography were used to ground truth drone imagery to develop training datasets. Modelled outputs demonstrated that Multispectral was the best drone camera sensor type, with the highest accuracy of 97.0 %, with NDVI the most important band for the model. When we applied the model to 50 m2 plots to validate its predictions, we had similar results to visual classification from field surveys and fine-scale photographs, successfully separating biocrust from bare ground. Our relatively simple method can be applied to biocrusts using readily available, low-cost technology. Considerable opportunities exist for using this approach to provide landscape-level biocrust assessment, using remote sensing, leading to improved restoration and management of drylands for conservation.http://www.sciencedirect.com/science/article/pii/S0016706125001533Biological soil crustsUAVXGBoostMultispectralDrylandsArid ecology
spellingShingle Jana Stewart
Roxane J. Francis
David J. Eldridge
Richard T. Kingsford
Nathali Machado de Lima
Advancing remote sensing of biocrusts with drone imagery and machine learning
Geoderma
Biological soil crusts
UAV
XGBoost
Multispectral
Drylands
Arid ecology
title Advancing remote sensing of biocrusts with drone imagery and machine learning
title_full Advancing remote sensing of biocrusts with drone imagery and machine learning
title_fullStr Advancing remote sensing of biocrusts with drone imagery and machine learning
title_full_unstemmed Advancing remote sensing of biocrusts with drone imagery and machine learning
title_short Advancing remote sensing of biocrusts with drone imagery and machine learning
title_sort advancing remote sensing of biocrusts with drone imagery and machine learning
topic Biological soil crusts
UAV
XGBoost
Multispectral
Drylands
Arid ecology
url http://www.sciencedirect.com/science/article/pii/S0016706125001533
work_keys_str_mv AT janastewart advancingremotesensingofbiocrustswithdroneimageryandmachinelearning
AT roxanejfrancis advancingremotesensingofbiocrustswithdroneimageryandmachinelearning
AT davidjeldridge advancingremotesensingofbiocrustswithdroneimageryandmachinelearning
AT richardtkingsford advancingremotesensingofbiocrustswithdroneimageryandmachinelearning
AT nathalimachadodelima advancingremotesensingofbiocrustswithdroneimageryandmachinelearning