Fossil charcoal particle identification and classification by two convolutional neural networks
Introduction
Fire has existed on Earth for over 400 million years (Bowman et al., 2009:481) and is an important environmental process interconnected to climate, vegetation structure, and carbon cycling (Beringer et al., 2015; Bowman et al., 2009; Veenendaal et al., 2017). Fire occurrence is also linked to people; humans have had a long and complex history with fire (Bowman et al., 2011; Scott et al., 2016), including the widespread use of fire as a landscape management tool (e.g. Anderson, 1994; Archibald et al., 2012; Montiel and Kraus, 2010; Rolland, 2004; Rull et al., 2015). Moss and Kershaw (2000), for example, suggest the impacts of Australian indigenous use of fire in the landscape can be seen over 38,000 years ago in the palaeoenvironmental record. Both natural and anthropogenic fire are connected to significant issues for asset management, conservation and cultural practice.
Fossil charcoal is an important palaeofire proxy. It has high preservation potential (Conedera et al., 2009; Mooney and Tinner, 2011; Whitlock and Larsen, 2001), and charcoal records are available worldwide (see Global Paleofire Working Group 2017 as well as Power et al., 2010, for depictions of the spatial and temporal scope of global charcoal records). Analysis of fossil charcoal, including identification of the type of vegetation that burned to create it, allows for the creation of long term fire records contextualised by fuel type (e.g. Aleman et al., 2013; Crawford and Belcher, 2014; Jensen et al., 2007). Such information allows for a greater understanding of fire and vegetation dynamics across time and space. However, traditional (optical) charcoal analysis is a time-intensive process; fossil charcoal is commonly quantified on pollen slides (particles <125 μm diameter) or wet sieved and suspended in water, with charcoal abundance measurements taken via microscope as either particle counts or area measurements (see Mooney and Tinner, 2011; Stevenson and Haberle, 2005). Increasing the speed of charcoal analysis will enable researchers to process a larger volume of samples in a given time frame, allowing for higher resolution records.
Recent developments in artificial neural networks have led to their successful application to problems across a diverse range of disciplines (e.g. review of developments and applications including finance, bioinformatics and environmental risk, Bassis et al., 2014; engineering, Mehrjoo et al., 2008; medical imaging, Wu et al., 2017). Drawing from these developments and using them as a framework for this study, the identification and classification process of fossil charcoal particles is an ideal candidate for automation using neural networks.
Section snippets
Background and related work
While volumes of preserved charcoal (number of fossil charcoal particles) help indicate the amount of fire in past landscape, the fuel source (vegetation type) of a fossil charcoal particle is significant as this reflects the composition of the surrounding environment. The aspect ratio of a macroscopic (>125 μm) charcoal particle provides this data, with more elongated particles identified as grass-derived and blockier particles as wood- or leaf-derived (e.g. Umbanhowar and McGrath, 1998;
Sample Collection and Preparation
Samples were taken from Holocene sediment cores collected from three wetlands in tropical northern Australia: Sanamere Lagoon (11.117°S, 142.35°E), Big Willum Swamp (12.657°S, 141.998°E) and Marura Sinkhole (13.409°S, 135.774°E). Sediment samples were prepared for fossil charcoal analysis following the method outlined by Stevenson and Haberle (2005); samples were placed in mid-strength (∼5% concentration) bleach for 72 h before wet sieving to isolate the >63 μm fraction. Samples were suspended
Results
The trained U-Net network achieved 96.06% accuracy, while the trained VGG network achieved 75.15% accuracy (Fig. 2).
The U-Net network results were sliced by using a connected components algorithm (Grana et al., 2010) to isolate individual particles, shown in Fig. 2 as a red bounding box. Identified particles measuring less than 63 × 63 μm were discarded, as samples were processed to only contain particles >63 μm (described in Sample Collection and Preparation above).
Discussion and areas for future work
This study is a proof of concept for the application of neural networks to charcoal particle analysis. Our initial results demonstrate the feasibility of this methodology, with high accuracy achieved by U-Net for charcoal identification and VGG for morphotype classification. In combination with an appropriate mechanical apparatus for particle photography such as an automated stage, this methodology has the potential to significantly accelerate particle analysis workflows, reducing the number of
Conclusion
The automated classification presented in this experimental study provides a fast and flexible method of charcoal analysis. This two-stage pre-trained network utilizing a broadly applicable morphotype classification system can be applied to samples from any site without requiring the creation of additional training datasets. Alternatively, these neural networks can be trained using any morphotype system, including classifications that may be more region- or biome-specific.
Automated charcoal
Data availability
The training dataset of processed charcoal particle images used in this paper is available via the following:
Rehn, E.; Rehn, A. (2019): Fossil charcoal particle training data for neural networks. James Cook University. (dataset). http://doi.org/10.25903/5d006c1494cf9.
CRediT authorship contribution statement
E. Rehn: Conceptualization, Investigation, Writing - original draft, Writing - review & editing. A. Rehn: Methodology, Software, Writing - original draft, Writing - review & editing. A. Possemiers: Software, Writing - original draft.
Acknowledgements
This study forms one part of author ER's PhD degree within the College of Science and Engineering, James Cook University. Sample collection was undertaken as part of an Australian Research Council Laureate Fellowship ID:FL140100044 (CI: M. Bird). ER acknowledges financial support from the Australian Institute of Nuclear Science and Engineering (Postgraduate Research Award 12143) and an Australian Government Research Training Program Scholarship. The authors would like to thank Cassandra Rowe
References (46)
Prehistoric anthropogenic wildland burning by hunter-gatherer societies in the temperate regions: a net source, sink, or neutral to the global carbon budget?
Chemosphere
(1994)- et al.
A 25,000-year record of environmental change from Welsby Lagoon, North Stradbroke Island, in the Australian subtropics
Quat. Int.
(2017) - et al.
Reconstructing past fire regimes: methods, applications, and relevance to fire management and conservation
Quat. Sci. Rev.
(2009) - et al.
Tracking recorded fires using charcoal morphology from the sedimentary sequence of Prosser Lake, British Columbia (Canada)
Quat. Res.
(2006) - et al.
A new approach to automated pollen analysis
Quat. Sci. Rev.
(2000) - et al.
Identifying past fire regimes throughout the Holocene in Ireland using new and established methods of charcoal analysis
Quat. Sci. Rev.
(2016) - et al.
Automatic classification of native wood charcoal
Ecol. Inf.
(2018) - et al.
Damage detection of truss bridge joints using artificial neural networks
Expert Syst. Appl.
(2008) - et al.
The last glacial cycle from the humid tropics of northeastern Australia: comparison of a terrestrial and a marine record
Palaeogeogr. Palaeoclimatol. Palaeoecol.
(2000) - et al.
Fire history and the Global Charcoal Database: a new tool for hypothesis testing and data exploration
Palaeogeogr. Palaeoclimatol. Palaeoecol.
(2010)
New insights on palaeofires and savannisation in northern South America
Quat. Sci. Rev.
TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems
Image processing with ImageJ
Biophot. Int.
Tracking land-cover changes with sedimentary charcoal in the Afrotropics
Holocene
Evolution of human-driven fire regimes in Africa
PNAS
Recent advances of neural network models and applications
Fire in Australian savannas: from leaf to landscape
Glob. Chang. Biol.
Fire in the Earth system
Science
The human dimension of fire regimes on Earth
J. Biogeogr
Charcoal morphometry for paleoecological analysis: the effects of fuel type and transportation on morphological parameters
Appl. Plant Sci.
Area-volume relationships for fossil charcoal and their relevance for fire history reconstruction
Holocene
The ImageJ User Guide 1.44
Global Charcoal Database
Cited by (11)
Recognition of a quasi-static region in a granular bed impacted with a sphere
2022, Powder TechnologyCitation Excerpt :Guo et al. used a U-Net network to accurately extract the liquid phase region in gas–liquid and solid–liquid separation, solving the problem of large errors in the grayscale thresholding method [17]. For the first time, Rehn et al. [18] used a neural network algorithm to recognize and extract fossil carbon particles from images, using the U-Net model. Because the pyrolytic carbon and buffer layers of melamine-coated film have different feature information, Zhu et al. [19] used the U-Net model to solve the problem of the difficulty of segmenting the thresholding method and extracted the region of these two layers.
Toward a standardized procedure for charcoal analysis
2021, Quaternary Research (United States)Species-Level Microfossil Prediction for Globotruncana genus Using Machine Learning Models
2023, Arabian Journal for Science and Engineering