Author(s): Kayla Andersen; Branko Kerkez

Linked Author(s):

Keywords: Soft sensing; Ecological monitoring; Gaussian processing; Machine learning

Abstract: Low-cost wireless sensor networks are valuable tools for environmental modeling, yet many ecologically significant variables remain difficult and expensive to estimate in real-time [1]. This study contributes a soft-sensing methodology which converts low-cost water level sensor data into difficult-to-measure ecological variables. We present a case study of the Huron River Watershed, Michigan where we develop a machine learning (ML) toolchain to map sensor data to spatially distributed velocity profiles to predict smallmouth bass nesting habitat quality; a key local indicator species [2]. We leverage an IoT network of over 30 sensors across 100 km2 and a comprehensive fieldwork campaign from the 2025 storm season. We trained a Gaussian process (GP) model with composite Radial Basis Function and Rational Quadratic kernels on relating water levels to in-stream velocity profiles and fish nest locations. The model successfully predicted local variations in streambed velocities and identified areas of suitable habitat. While the model currently lacks high-discharge training data, which will be addressed in the 2026 fieldwork season, the proposed ML-toolchain is a replicable method for predicting ecological variables from increasingly available sensor networks, applicable to any river system where such manual measurements are possible.

DOI:

Year: 2026