Open AccessBAYESIAN REGULARIZED NEURAL NETWORKS APPROACH AND UNCERTAINTY ANALYSIS FOR REFERENCE EVAPOTRANSPIRATION MODELING ON SEMIARID AGROECOSYSTEMS
Author(s) -
C. de O. F. Silva,
A. H. de C. Teixeira,
Rodrigo Lilla Manzione
Publication year - 2020
Publication title -
revista brasileira de engenharia de biossistemas
Language(s) - English
Resource type - Journals
eISSN - 2359-6724
pISSN - 1981-7061
DOI - 10.18011/bioeng2020v14n1p73-84
Subject(s) - evapotranspiration , mean squared error , wind speed , bayesian probability , bayesian network , artificial neural network , precipitation , variable (mathematics) , mathematics , statistics , scale (ratio) , environmental science , computer science , meteorology , machine learning , ecology , geography , cartography , mathematical analysis , biology
The Penman–Monteith equation (PM) is widely recommended by The Food and Agriculture Organization (FAO) as the method to calculate reference evapotranspiration (ET0). However, the detailed climatological data required by the PM are not often available. The present study aimed to develop bayesian regularized neural networks (BRNN)-based ET0 models and compare its results with the PM approach. Forteen weather stations were selected for this study,located in Juazeiro (BA) and Petrolina (PE) counties, Brazil. BRNN were trained with different parameters choices and obtained R² between 0.96 and 0.99 during training and between 0.95 and 0.98 with validation dataset. Root mean squared error (RMSE) less than 0.10 mm.day-1 for BRNN when compared to PM denoted the good performance of the network using only air temperature, solar radiation and wind speed at average daily scale as input variable. Epistemic and random uncertainties were evaluated and precipitation was identified as the variable with the greatest uncertainty, being therefore discarded for modeling.