A Mariculture Fish Mortality Prediction Using Machine Learning Based Analysis of Water Quality Monitoring

The unstable production due to mass mortality of grouper is a big issue for fish farmers in Gondol, the northern coast of West Bali, Indonesia. This study aimed to develop fish mortality risk prediction for the next day using machine-learning from a water quality data sensor network and daily fish mortality record input by the fish farmers. This study uses seawater temperature, current, salinity, conductivity, chlorophyll, turbidity and DO (dissolved oxygen) as water quality attributes. This study used Fast Fourier Transform (FFT) to detect periodic fluctuation of seawater temperature and found that 12-day, 7-day and 3-day are important periodic indicators. Yet, when and how water quality affects fish mortality is unknown, hence, this study used the moving average (MA) based on FFT results to capture the trend in the data. The independent variables for the prediction were MA of sea water temperature, current, salinity, chlorophyll, turbidity and DO. Conductivity was omitted due to the high correlation with salinity. The dependent variable was daily five levels of fish mortality based on the number of fish mortality characteristics to make it easier for providing the prediction result to the fish farmer. This study used python 3 and Random Forest (RF) algorithm in the Scikit-learn library to predict the daily fish mortality. This study found that the accuracy of the proposed prediction worked best using the 3-day moving average data set. In particular, the average accuracy of 10-time running the prediction was 74.92%. RF feature importance indicated that sea water temperature, salinity and turbidity are important variables to predict fish mortality in the target area.