Authors / CoAuthors
Du, Z. | Wilford, J. | Roberts, D. | Czarnota, K.
Abstract
<div><strong>Output Type: </strong>Exploring for the Future Extended Abstract</div><div><br></div><div><strong>Short Abstract: </strong>Most geological mapping either over-estimates the amount of bedrock exposed at the surface or can miss local bedrock exposures in geological units describing cover materials (i.e. alluvium and colluvium). A machine learning, Convolutional Neural Network (CNN) has been applied to detect outcrops (exposed bedrock at the earth’s surface) and areas of very shallow cover over bedrock (i.e. sub-crop) at one meter resolution. We used a multi-feature training dataset consisting of sites associated with urban areas, roads, outcrops, waterbodies, soil (includes bare soil and soil covered by green and dry vegetation), trees, and shadows. Even though we were only interested in mapping outcrop, a multi-criteria label set significantly improved overall accuracy of the model. The explanatory variables or covariates included high-resolution satellite imagery, Sentinel-2 imagery, and terrain derivatives. The modelling approach was tested over an area in central West NSW, Australia. Labels were split into 80% for training and 20% for out-of-sample validation. Spatial K-groups were used in the training set to minimize auto-spatial correlation between neighbouring points and reduce the potential for overfitting. Two CNN model architectures were evaluated: Simple-Net and UNet. The Simple-Net structure consists of 2D Convolution layer and flatten layer, whereas the UNet architecture includes a mixture of 2D convolution layer, max pooling, up sampling and flatten layer. These models were tested with and without the use of high-resolution imagery. The UNet model incorporating high resolution imagery gave the best results (accuracy of 0.841 and an F1 score of 0.814), compared with Simple-Net (accuracy of 0.823 and an F1 score of 0.786). However, the Simple-Net’s model without the incorporation of high-resolution imagery was a slight improvement over the UNet architecture and due to the lack of national coverage for high-resolution imagery, the Simple-Net model offers better scalability. The detection of outcrop/sub-crop has broad application in improving the spatial explicitness of existing geological maps, improving sample detection and interpretation of litho-stratigraphy and geochemistry. High spatial resolution of outcrop/sub-crop also has implications in the agricultural and civil engineering sectors, ecology and in understanding surface and near-surface hydrological systems. After this proof-of-concept phase we plan to up-scale the approach nationally using a more representative labels and national covariates.</div><div><br></div><div><strong>Citation: </strong>Du, Z., Wilford, J. & Roberts, D., 2024. Find a rock or a rock nearby using Convolution Neural Networks. In: Czarnota, K. (ed.) Exploring for the Future: Extended Abstracts. Geoscience Australia, Canberra. https://doi.org/10.26186/149631</div>
Product Type
document
eCat Id
149631
Contact for the resource
Resource provider
Point of contact
Cnr Jerrabomberra Ave and Hindmarsh Dr GPO Box 378
Canberra
ACT
2601
Australia
Point of contact
- Contact instructions
- MEG
Digital Object Identifier
Keywords
- ( Project )
-
- EFTF – Exploring for the Future
- ( Project )
-
- Australia’s Resources Framework
-
- geological mapping
-
- machine learning
-
- outcrop
- theme.ANZRC Fields of Research.rdf
-
- Geomorphology and earth surface processes
-
- Published_External
Publication Date
2024-08-11T22:56:48
Creation Date
2024-05-22T22:00:00
Security Constraints
Legal Constraints
Status
completed
Purpose
Purpose 1. Making geological maps more spatially explicit and accurate. 2. Targeting rock exposes for geological interpretation and sampling (e.g. geochemistry)3. Improving our understanding of surface and near-surface hydrology4. Improve agricultural assessment associated with surface fertility, drainage, and land suitability classification.5. Ecological modelling – identifying flora and fauna habitat and refuge. 6. Inputs into civil engineering planning – road construction, quarrying and excavation.
Maintenance Information
notPlanned
Topic Category
geoscientificInformation
Series Information
Lineage
<div>This work is based on a machine learning (ML) workflow using Convolution Neural Networks (CNNs) to generate high resolution outcrop predictions. The ML workflow uses 'Landshark' which is a python workflow to streamline CNN processing for large models. Landshark which was developed through a collaborative between Geoscience Australia and CSIRO. Landshark uses the TensorFlow platform (Abadi et al., 2016).</div><div><br></div><div>ABADI, M., et al. TensorFlow: a system for Large-Scale machine learning. 12th USENIX symposium on operating systems design and implementation (OSDI 16), 2016. 265-283.</div><div><br></div><div>Using ML to detect rocky outcrops has tried previously. Similar work was undertaken by Petliak et al. (2019) where a CNN approach was used to separate exposed bare rock from soil or regolith. In this study we have used a combination of high resolution (1m) 3-band imagery together with a customized set of covariates including derivatives from a 30m terrain model and temporal-spectral covariates generated from time series and bare earth (Roberts et al. 2019) Sentinel-2 satellite observations. The CNN model was trained on seven key surface features including urban areas, roads, outcrops, waterbodies, soil (includes bare soil and soil covered by green and dry vegetation), trees, and shadows. We select 1000+ sample for each class. We used 80% of the sites for training and 20% for out-of-sample. Results are found in the extended abstract.</div><div><br></div><div>Petliak, H.; Cerovski-Darriau, C.; Zaliva, V.; Stock, J. Where’s the Rock: Using Convolutional Neural Networks to Improve Land Cover Classification. <em>Remote Sens.</em> <strong>2019</strong>, <em>11</em>, 2211. </div><div><br></div><div>Roberts D., Wilford J. & Ghattas O., 2019. Exposed soil and mineral map of the Australian continent revealing the land at its barest. <em>Nature Communications </em>10:5297</div><div><br></div><div><br></div>
Parent Information
Extents
[-38, -28, 140, 154]
Reference System
Spatial Resolution
Service Information
Associations
Source Information