Data Driven Sensor Modeling for Algal Blooms Monitoring

Currently, salvage is considered as an effective way for protecting ecosystems of inland water from toxin-producing algal blooms. Yet, the magnitude of algal blooms, which is the essential information required for dispatching salvage boats, cannot be estimated accurately with low cost in real time. In this paper, a data-driven soft sensor is proposed for algal blooms monitoring, which estimates the magnitude of algal blooms using data collected by inexpensive water quality sensors as input. The modeling of the soft sensor consists of two steps: 1) magnitude calculation and 2) regression model training. In the first step, we propose an active learning strategy to construct high-accuracy image classification model with ~50 % less labeled data.

Based on this model, we design a new algorithm that recognizes algal blooms and calculates the magnitude using water surface pictures. In the second step, we propose to use Gaussian process to train the regression model that maps the multiparameter water quality sensor data to the calculated magnitude of algal blooms and learn the parameters of the model automatically from the training data. We conduct extensive experiments to evaluate our modeling method, AlgaeSense, based on over 200 000 heterogeneous sensor data records collected in four months from our field-deployed sensor system. The results indicate that the soft sensor can accurately estimate the magnitude of algal blooms in real time using data collected by just three kinds of inexpensive water quality sensors.