Thread by @WardHydro

🚨 New Pub 🚨
How can we design robust sensor networks? How do we assess the future value of data BEFORE we invest in several years of observations? Tyler Balson & I explore this in a new pub @HydroProcesses. a 🧵

#MachineLearning #Hydrology

doi.org/10.1002/hyp.14…

Motivation: To improve a forecast (e.g., 1-day forecast of in-stream #nitrate in the Wabash River Basin), where do we put a sensor? To benefit whom?

Does a sensor to improve city A forecast also help City B? Can we leverage for network-scale benefits? Is there an optimal?

First problem: How to assess the value of data you don't have? We used Agro-IBIS to generate a synthetic data set. We proceed to analyze these model outputs at locations of current and potential sensors.
H/T to @Chris_Kucharik my AgroIBIS mentor

@Chris_Kucharik With (synthetic) timeseries in-hand to match the existing gauge network, we used SVMRs to make forecasts at the 12 largest cities in the basin. Overall the SVMRs performed pretty well. Put plainly, we can make reasonable forecasts with our existing network.

@Chris_Kucharik Now what if you force the SVMRs to add nitrate data from another site? We took adding a nitrate sensor at the USGS gauges that didn't already have one as our population of potential sites?
(a) most sites yield a benefit for a given city (tall bars w/ small RMSE benefit)

@Chris_Kucharik (b) a few sites yield an anomalously high benefit (bars on the far right)
(c) a few sites actually degrade the prediction (bars on far left, meaning you would simply ignore them in practice, not actually degrade your model by forcing them in)

@Chris_Kucharik (a) through (c) above are generally true for all of the locations were we tried forecasting. Columbus forecast was notably difficult - many more sites that were not useful. Hypothesize b/c of unique geology in basin vs. other sites.

@Chris_Kucharik If one city (row) acts in its self interest and installs their optimal sensor, what happens to forecasts in other cities (columns)? Usually minor benefit.

But - benefits are not reciprocal. Champaign benefits from adding Kokomo's optimal sensor, but K does not benefit from C's

@Chris_Kucharik What if you want to balance benefit to the network instead of individual cities (e.g., as a state or federal agency)? Top-right of (b) shows maximum RMSE improvement AND number of locations improved.

@Chris_Kucharik Take-home messages:
(1) We present a framework for using synthetic data as ML as an efficient way to evaluate the future value of sensor deployments.

@Chris_Kucharik (2) We introduce trade-offs between selfish action by individual cities vs. benevolent action with a network-scale benefit, which may spur discussion when coordinating on sensor network design and objectives.

@Chris_Kucharik (3) Tons of room for expansion here. It is easy to imagine an objective function that includes more types of forecasts (1- & 7-day), allows for multiple sensors or parameters, etc.

@Chris_Kucharik Sensor deployments are seldom informed by unbiased evaluation of value to forecasts, instead networks organically result from a series of local and project-based decisions. As hydrologists engage in forecasting, tools to assess the future value of data to forecasts are needed.

@Chris_Kucharik Support from: @doescience @NSF_geo @USUKFulbright @bIGIdeas_UoB @IULuddy @IUONeillSchool @CASResearch_OSU @OSUAgSci @EngineeringOSU @LES_UniBham

Data on @CUAHSI HydroShare

@threadreaderapp unroll