Content uploaded by Chulantha Jayawardena
Author content
All content in this area was uploaded by Chulantha Jayawardena on Dec 13, 2023
Content may be subject to copyright.
Proceedings of SLRMES International Conference 2023
ISRM specialized SLRMES conference on Rock Mechanics for Infrastructure and Geo-Resources Development
53
Weathered Rock Surface Classification with Unpiloted
Aerial Vehicle Imagery and Machine Learning
CL Jayawardena*, K Brinthan, K Gamsavi, KGAU Samarakoon
and TMB Senarathna
Department of Earth Resources Engineering, University of Moratuwa, Katubedda, Sri Lanka
*Corresponding author – chulanthaj@uom.lk
Introduction and Motivation for the Research
Weathering of rocks is a critical process that significantly influences the stability and response
of rocks under dynamic conditions. Weathering is the primary process by which the strength
of rocks is adversely affected. Traditional methods of determining the degree of weathering
such as Schmidt rebound, penetration-obstruction, hardness tester, and ultrasonic methods
require more workforce and lead to time consuming and difficult field monitoring tasks.
Frequently, they either give incomplete field data or put the working staff in danger because
of the difficult terrains with potentially hazardous conditions despite the safety measures and
time-consuming efforts. The advancement of technology, tools, data-gathering platforms, and
data-processing techniques provide attractive alternatives for traditional time-consuming
methods while demanding rigorous evaluation prior to usage as acceptable replacements for
the existing. In light of this, aerial or satellite images for such operations are the preferred
choice at present besides, the major drawbacks in spatial and temporal resolution. Unpiloted
aerial vehicles (UAVs) or drones have become a popular and successful remote sensing tool
for many disciplines due to their high spatial range, resolution, and flexibility in handling.
Objectives of the Research and Procedures
Objective: Using the UAVs to assess the possibilities of identifying the rock weathering
patterns aiming to replace the laborious tasks involved in the traditional field works.
Procedures: The constructed benches of an abandoned quarry site having different weathering
grades and abnormally weathered regions were imaged using a DJI Multi-spectral Phantom 4
UAV and processed with DJI GS Pro software. The UAV flew at the height of 120 m with a
front overlap of 80% and a side overlap of 60% while covering an area of 0.33 km2 with a whisk
broom flight path for image acquisition. The captured images were processed using Pix4D
Mapper, and mosaic maps of each band were created and georeferenced using ground-control
points (GCPs). The GCP coordinate measurements were taken from a Stonex DGPS unit and
were used to increase the georeferencing accuracy. The UAV images consisting different
shades of features, including soil, various stages of weathering (complete, moderate, and
slight), fresh rocks, wet rocks, and vegetation were converted into unique classes for
classification purposes using 800 × 500-pixel region and subject to undergo supervised
machine learning techniques. The obtained multispectral images having five spectral bands
were initially prepared with natural colour composites and sixty false-colour composites.
Experts performed ground-truth labelling for the selected study area through on-site
Proceedings of SLRMES International Conference 2023
ISRM specialized SLRMES conference on Rock Mechanics for Infrastructure and Geo-Resources Development
54
observations and high-resolution UAV images. Fifty-eight critical features were extracted
from each false-colour composite using different filters. The performance of the 18 machine
learning algorithms for this specific classification study was assessed using the Pycaret
library’s “compare_models()” function. Algorithms that performed equally well, based on
overall accuracy and F1 score were selected to assess other false-colour composites. A
preliminary assessment was performed to determine the best combination of false-colour
composites and selected ML algorithms. The algorithms were employed with default
parameter settings and their overall accuracy and F1 scores were used to assess their
performance. To further support our findings, we analysed the confusion matrix, receiver
operating characteristic (ROC) curve, area under the curve (AUC), and contrast of spectral
values between the selected bands. Finally, a high-performance model and promising band
combinations were recommended for further analysis.
Results
Depending on the overall accuracy and F1 score of the high-performance ML algorithms
measured for the natural colour composite image, the Random Forest (RF) classifier and
Extreme Gradient Boosting Machine (XGB) performed well, with an overall accuracy of 0.65
and an F1 score of 0.6, respectively. Therefore, RF and XGB were chosen for further evaluation
of the other false colour composite images. XGB performed slightly better than RF in
classifying most classes, such as discoloured rock, completely weathered rock, moderately
weathered rock, fresh rock, wet rock, and vegetation cover. However, both algorithms
performed similarly in classifying soil and slightly weathered rocks. Among these, the
weathered rock class (comprising completely, moderately, and slightly weathered rocks)
achieved an F1 score of 0.88 using the RE-NIR-R (RE- Red-edge, NIR- Near infrared, R- Red)
band combination with a 96% correct prediction rate.
Conclusions
This study highlights the potential of utilising UAVs with multiple ML algorithms to classify
weathered rock surfaces. The RE-NIR-R band combination with the XGB algorithm yielded an
impressive F1 score of 0.88 for classifying weathered rocks in general. However, the
subclassification of different weathering classes yielded a range of F1 scores (i.e., 0.4 to 0.71),
mainly owing to restrictions, such as the spatial resolution, enforced by the multispectral
sensors. However, this combination of bands enabled the identification of a clear distinction
between vegetation and discoloured rocks. Despite these limitations, our study serves as a
promising first step in demonstrating the potential of combining images acquired from UAVs
and ML algorithms to identify and classify rock-weathering patterns on exposed rock surfaces.