Climate change is creating dramatic and unprecedented changes in stream flow, resulting in hazardous conditions due to overbank flooding, flash flooding, debris flows, and landslides. Satellite based weather forecasting along with traditional monitoring can predict hazardous stream conditions in many cases, but often falls short due to microclimate variations and the lack of fine-grained stream monitoring.

Here we describe the work of Kemeny and Kim (2022) and Kemeny and Kim (2023).  It involved developing a novel ai-based remote sensing approach to monitoring stream conditions and forecasting hazardous conditions. This work is innovative and only used sound to monitor stream flow. We are all familiar with the sound of water flowing, whether it is a river nearby, the crashing waves along a beach, or a dripping faucet or toilet. Human beings have the ability not only to recognize water sounds but also to isolate the sound of water amongst the noise of cars, children playing, and building construction. This work shows that a trained neural net model can also accurately discriminate the sound of water in a noisy environment. We are using the concepts of Edge AI and tinyML, training on-sensor ai models to continually inference stream conditions and transmit only small packets of inferenced results through low bandwidth protocols. Our sensors can be utilized in both urban and natural environments, and installation can be as easy as hanging the device from a tree.

For ai training, we collected some Proof of Concept data using an iPhone from 1) a small stream in Colorado (Slate River) that sees flow changes due to spring snowmelt and summer monsoon rainfall, and 2) the Rillito River in Tucson that saw very high monsoon rainfall in summer 2022. Our stream classification ai model utilizes signal processing and deep convolutional neural network (CNN) modeling of the collected data, and the trained model inferences changes in stream conditions in discrete classifications. The Rillito River data was collected under the Campbell Ave. bridge along the Rillito River in Tucson, Arizona during a series of large rainfall events. Based on the amount of flow detected at a nearby USGS station on the Rillito River, our collected data was labeled no flow, low/med flow, high flow, and very high flow. Hundreds of labeled sound samples were collected, with 68% of the data used for training and 32% was used for testing. Using a fairly simple CNN model with two convolutional layers, training achieved a 95.3% accuracy and testing (data it was not trained with) achieved a 92.9% accuracy. This site was very noisy with cars continually passing along Campbell Ave., and also the water sounds are very complex derived from flow past the bridge pillars and rocks in the river. The figure below shows pictures of the four flow cases, as well as a dimensionally reduced plot of the embedding vectors from the second CNN layer. The embedding vector plot indicates how well the trained model is at differentiating sounds from the different flow levels, showing fairly good separation between the no flow (green), low/med flow (orange), and high flow (blue) with some mixing between the high and very high flows (red). Overall our results support the use of sound to monitor flow. In the future we also plan to evaluate the use of computer vision to assist the inferencing at high flow levels. In this example the very high flow flooded the bike path and could be readily inferenced by camera images.

Figure showing No flow, Med/High flow, High flow, Very High flow, t-SNE dimension reduction of embedding vectors of trained CNN model.

If you are interested in this topic and want to work with us on further developing it (as research or a potential commercial product) please contact and let’s discuss.

Kemeny, J. and K. Kim. 2022. In-Situ Sensing of Stream Flow using Deep Learning Sound Classification with Tagged Images, Oral presentation at the 2022 Fall Meeting of the American Geophysical Union, Chicago, IL.

Kemeny, J. and K. Kim. 2023. Simple Sensor Solutions (With the Help of AI) for Geologic and Hydrologic Hazards Associated with Climate Change, Poster presentation at the 2023 Fall Meeting of the American Geophysical Union, San Francisco, CA.