Intelligent Prediction of the Daily Irrigation Water Requirements for Citrus Orchards Using FFNN with Levenberg Marquardt Algorithm

Abstract

This paper describes an approach to predict the citrus water requirements using a Multi Layer Feed Forward Neural Network MLFFNN and the Levenberg-Marquardt Algorithm (LMA). Literature relates that the LMA is faster than the SONN and FFBP-NN algorithms. So the aim of this paper is focused on which best data base conception and the most performant MLFFNN architecture using LMA to predict water requierement for citrus plants in natural grow conditions. The drip irrigated citrus orchards are grown at the Technical Centre of Citrus (CTA) of Nabeul, at Northeast of Tunisia (36.5ºN, 10.2ºE), where trees are cultivated under standard conditions (disease-free, well-fertilized and not short of water). The daily climatic data (minimum and maximum air temperatures and relative humidity, wind speed, precipitation and solar radiation) recorded at the site of the experiment were used as inputs to the MLFFNN. Two secondary databases (DB) were used, both taken from the CTA's original database, to test the network, which has, for the case study, a single output: the amount of irrigation water needed by the citrus groves. Several architectures of the MLFFNN have been tested. The selected network, named A1-MLFFNN is constituted by two hidden layers, each comprising 5 neurons. With the LMA, the implementation of the A1-MLFFNN provides rapid and accurate results. The LMA showed high performance expressed in terms of the convergence speed and the reliability of the MLFFNN and particularly when using the DB2 (original database private of the unavailable water days) which is the most robust, giving the lowest RMSE and the highest coefficient of correlation (r = 0.799), and thus, allowing a better estimation of the amounts of irrigation water.