Convolutional neural networks for image-based sediment detection applied to a large terrestrial and airborne dataset

[摘要] Image-based grain sizing has been used to measure grain size moreefficiently compared with traditional methods (e.g., sieving and Wolman pebblecount). However, current methods to automatically detect individual grainsare largely based on detecting grain interstices from image intensity whichnot only require a significant level of expertise for parameter tuning butalso underperform when they are applied to suboptimal environments (e.g.,dense organic debris, various sediment lithology). We proposed a model(GrainID) based on convolutional neural networks to measure grain size in adiverse range of fluvial environments. A dataset of more than 125 000grains from flume and field measurements were compiled to develop GrainID.Tests were performed to compare the predictive ability of GrainID withsieving, manual labeling, Wolman pebble counts (Wolman, 1954) andBASEGRAIN (Detert and Weitbrecht, 2012). When compared with the sievingresults for a sandy-gravel bed, GrainID yielded high predictive accuracy(comparable to the performance of manual labeling) and outperformedBASEGRAIN and Wolman pebble counts (especially for small grains). For theentire evaluation dataset, GrainID once again showed fewer predictive errorsand significantly lower variation in results in comparison with BASEGRAIN andWolman pebble counts and maintained this advantage even in uncalibratedrivers with drone images. Moreover, the existence of vegetation and noisehave little influence on the performance of GrainID. Analysis indicated thatGrainID performed optimally when the image resolution is higher than 1.8 mm pixel −1 , the image tile size is 512×512  pixels and the grain areatruncation values (the area of smallest detectable grains) were equal to 18–25 pixels.

[学科分类] 土壤学