A System for Sensing Human Sentiments to Augment a Model for Predicting Rare Lake Events

Abstract

Fish kill events (FKE) in the caldera lake of Taal occur rarely (only 0.5\% in the last 10 years) but each event has a long-term effect on the environmental health of the lake ecosystem, as well as a devastating effect on the financial and emotional aspects of the residents whose livelihood rely on aquaculture farming. Predicting with high accuracy when within seven days and where on the vast expanse of the lake will FKEs strike will be a very important early warning tool for the lake's aquaculture industry. Mathematical models to predict the occurrences of FKEs developed by several studies done in the past use as predictors the physico-chemical characteristics of the lake water, as well as the meteorological parameters above it. Some of the models, however, did not provide acceptable predictive accuracy and enough early warning because they were developed with unbalanced binary data set, i.e., characterized by dense negative examples (no FKE) and highly sparse positive examples (with FKE). Other models require setting up an expensive sensor network to measure the water parameters not only at the surface but also at several depths. Presented in this paper is a system for capturing, measuring, and visualizing the contextual sentiment polarity (CSP) of dated and geolocated social media microposts of residents within 10km radius of the Taal Volcano crater ($14^\circ$N, $121^\circ$E). High frequency negative CSP co-occur with FKE for two occasions making human expressions a viable non-physical sensors for impending FKE to augment existing mathematical models.