GEOSCAN Search Results: Fastlink

GEOSCAN Menu


TitleEvaluating ground-truth data collection methods for earth observation of lichen woodlands: error and error propagation when upscaling to higher-level assessments
 
AuthorLovitt, JORCID logo; Chen, WORCID logo; Wowk, W; Leblanc, S GORCID logo; He, L; Schmelzer, I; Arsenault, A; Fraser, RORCID logo; Prévost, C; White, H PORCID logo; Pouliot, D
SourceCanadian Remote Sensing Society, Program, 41st Canadian Symposium on Remote Sensing: landscapes of change; remote sensing for a sustainable future/Société canadienne de télédétection, programme, 41e Symposium canadien de télédétection : paysages du changement ; télédétection pour un avenir durable; 2020 p. 88
Image
Year2020
Alt SeriesNatural Resources Canada, Contribution Series 20200432
PublisherCanadian Remote Sensing Society
Meeting41st Canadian Symposium on Remote Sensing / 41e Symposium canadien de télédétection; CA; July 13-16 juillet, 2020
Documentbook
Lang.English
Mediaon-line; digital
File formatpdf
ProvinceQuebec; Newfoundland and Labrador
AreaLabrador
Subjectsgeophysics; environmental geology; remote sensing; photography; field methods; software; Lichen; Biomass; Forests; Methodology; Data processing; Classification
Released2020 07 13
AbstractConventional methods of estimating ground lichen abundance (percent cover) include ocular estimates and point-intercept frames. The first can be prone to human bias depending on the experience and training of the data collector, while the second can be time-consuming and/or introduce significant error depending on the point density used. A third option is to collect digital plot photographs and perform binary image classification (presence/absence) after returning from the field. Using the three methods described above, we assessed lichen abundance of 93 plots collected in 2019 from 10 study sites across northern Québec and Labrador. Prior to performing the digital photo classifications (DPC), we segregated plots into three image quality groups (A, B, C). We then extracted cover class samples from Group A photos (highest quality) to train a convolutional neural network in Trimble eCognition software, and used a combined CNN and OBIA approach to produce a refined binary classification for each photo. We compared the resulting lichen abundance estimates with those derived from the ocular method and multiple densities of digital photo point intercepts (DPIs). We also performed a theoretical upscaling of each plot to stand-level lichen biomass (kg/ha).
Our initial results (n=53) indicate the greatest discrepancies between methods occurred in plots with more than 25% lichen cover and a heterogeneous lichen distribution. We found good agreement between DPC and DPI output with high DPI point densities (1000 points; RMSE: 0.09). Ocular estimates of lichen cover were consistently higher, predicting 10-15% more lichen cover on average than DPI and DPC methods. This translated to higher lichen biomass predictions at the stand-level (+27% mean error). Unsurprisingly, we noted the DPC method is sensitive to image quality. When applied to Group C images the DPC method produced lichen abundance estimates that were significantly different from ocular estimates of the same plots (-15% mean, RMSE: 0.21). Based on these preliminary results, we consider the DPC method to be especially useful in estimating lichen abundance for plots with heterogeneous lichen distributions. However, special care must be taken to ensure images are collected with sufficient quality to produce reliable results.
GEOSCAN ID327276

 
Date modified: