A machine learning approach to predict the pink shrimp harvest in the Patos Lagoon estuary

Abstract

This paper presents a novel methodology to predict the natural behavior of pink shrimp (Farfantepenaeus paulensis) harvest, in the Patos Lagoon Estuary (PLE) by using supervised machine learning. This prediction is a critical task due to its environmental, economic and social impact. Supervised machine learning algorithms such as Support Vector Machines (SVM), decision trees and rules learning were combined with meta-learning techniques to perform the discrete prediction of the harvest. Performance of several classifiers is evaluated by a set of metrics, especially by a specific metric to deal with the inherent relation of order between the classes. The official harvest data, provided by government agencies, may be affected by random and systemic errors caused mainly by illegal fishing and lack of efficient landing control. These errors, together with the lack of knowledge of the fishing effort employed, increase the difficulty of the prediction task. Results obtained using meta-learning techniques combined with classic algorithms reached an accuracy of 91% for the pink shrimp harvest prediction.