Automated mapping of the mean particle diameter characteristics from UAV-imagery using the CNN-based GRAINet model
Abstract
This study uses the GRAINet convolutional neural networks (CNN) approach on unmanned aerial vehicles (UAVs) optical aerial imagery to analyze and predict grain size characteristics, specifically mean diameter (dm), along a gravel river point bar in Sumava National Park, Czechia. By employing a digital line sampling technique and manual annotations as ground truth, GRAINet offers an innovative solution for particle size analysis.
Eight UAV overflights were conducted between 2014 and 2022 to monitor changes in grain size dm across the river point bar. The resulting dm prediction maps showed reasonably accurate results, with mean absolute error (MAE) values ranging from 1.9 to 4.4 cm in 10-fold cross-validations.
Mean squared error (MSE) and root-mean-square error (RMSE) values varied from 7.13 to 27.24 cm and 2.49 to 4.07 cm, respectively. Most models underestimated grain size, with around 68.5% falling within 1s and 90.75% falling within 2s of the predicted GRAINet mean dm.
However, deviations from actual grain sizes were observed, particularly for grains smaller than 5 cm. The study highlights the importance of a large manually labeled training dataset for the GRAINet approach, eliminating the need for user-parameter tuning and improving its suitability for large-scale applications.