Laboratory channel widening quantification using deep learning

Linear erosion channel (LEC) devastates arable land and significantly contributes to soil loss in agricultural watersheds. In the presence of a less- or non-erodible layer, channel widening governs the erosion process once the channel bed incises to this layer, accompanied by failure block generation and transport. Current knowledge on channel widening, however, is limited due to the lack of robust and efficient methods to capture the rapid sidewall expansion process. Laboratory experiments were designed to simulate the channel widening process with an initial channel width of 10 cm. Two packed soil beds with a non-erodible layer and two slope gradients (5 % and 11 %) were subjected to the inflow rate of 0.67 L/s. Images were captured by mounted digital cameras and automatically transformed into orthophotos. Channel edges and failure blocks were automatically detected by deep learning algorithm in a newly developed Channel-DeepLab network model based upon DeepLabv3+ platform. The procedure includes learning samples labelling, data augmentation, model construction, training, and validation. Sediment discharge and changes in channel width, geometry of channel edges, and failure blocks were measured. The results indicate that initial period is critical for erosion prediction and remediation due to its small sidewall failure interval, high channel expansion rate and sediment discharge. Channel surface area has great potential on accumulated sediment discharge prediction. The slope section that witnessed the fastest channel widening rate migrated downwards when slope gradient increased from 5 % to 11 %. The total number and area of the failure blocks increased with time, while the collapse frequency of the sidewalls decreased. Upstream reach experienced the highest sidewall collapse frequency and rate of disaggregation and transport, while the downstream reach experienced the highest total number of failure blocks. A time lag was found between sidewall collapse and sediment discharge, which increased as time progressed, attributing to decreased runoff erosivity as the flow velocity decreased. Results of this study will provide methodological support for channel sidewall and streambank retreat monitoring, realizing the automatic detection of channel edges and efficient output of rapid sidewall expansion process with high temporal and spatial precision. Future work can be focused on broadening the applicability of the Channel-DeepLab network model and quantifying the delayed response process between sidewall failure and sediment discharge.