COMBINING GENERATIVE MODEL AND RANDOM FOREST TO PREDICT SHRIMP DISEASE OCCURRENCE

Abstract

Penaeus vannamei is one of the most cultured species. The global production of  Penaeus (Litopenaeus) vannamei reached 5.8 million tonnes in 2020, contributing to 51.7% of total shrimp production. However, despite its high production, there are still many issues in this industry. One of those is the disease. The disease threatens shrimp farming, such as slowing shrimp growth rate and even mortality. To help the farmers in mitigating the impact of disease we tried to develop a predictive model that is able to give early warning of disease occurrence. We focused on predicting acute hepatopancreatic necrosis disease (AHPND), white feces disease (WFD), infectious myonecrosis virus (IMNV), and white spot disease (WS). We used data from 1839 cultivation cycles. The cycles are managed by 383 Farms. The data covered 4 physical parameters measured twice daily (in the morning and evening). Those parameters are water temperature, dissolved oxygen, salinity, and pH. The data also cover disease tests. We conducted several processes to develop the predictive model. First, we improve the data quality using the Generative Adversarial Network model (GAN). The improved data is then used for feature engineering and model training. We used the Random Forest Model as the predictor to the data we managed to achieve an average F1 score of 0.91 for the four diseases. The model achieved an F1 score of 0.91 for AHPND, 0.89 accuracy for WFD, 0.93 accuracy for IMNV, and 0.9 accuracy for WS. Those results indicate a good possibility to predict the disease occurrence based on water quality data. Hence the method can be used as an early warning system to help the farmer in mitigating disease occurrence.