RockMapper not processing braided prairie rivers well #6
Description
I attempted to process recordings from the Arkansas River north of Kaw Lake Oklahoma (in Kansas) using RockMapper, which is a river that was not used when training the model. This river is extremely wide and shallow, and nearly 100 percent sand. While the area recorded on sonar is quite wide, only the middle 30m had substrate that was recognizable on SSS due to the wide and shallow nature of the river. For this reason, manual substrate delineation of this river consisted of only this middle 30m (see picture 4). When processing the recording with RockMapper, the model predicted the middle 30m of the river mostly correctly, however, outside of this legible range, the RockMapper predictions were extremely unreliable and created a noisy patchwork of boulder/cobble, shadow, and bedrock classes (all of which were not actually there). This noisy patchwork also created problems when removing polygons smaller than the minimum mapping unit of 5m. To remove small polygons with the current workflow, it increased the processing time for the same set of recordings from 5 hours to 36 hours+.
