Artificial neural networks as support for leaf area modelling in crop canopies

This paper investigates the feed forward neural networks utilization in improving some crop growth parameters (e.g. Leaf Area Index) forecasting based on cumulated Photosynthetically Active Radiation (PAR) time series. PAR data recorded by the DAQ system were used as input in a nonlinear sigmoid function with 3 parameters. LAI series of the red clover canopies represented the outputs of the nonlinear identified model. These outputs (N=257 for each set) were used to train different network topologies of feed-forward ANN using Quickprop and Rprop learning algorithms. The ANN output represented the one LAI ahead forecasted value. Best fitting results were obtained using QuickProp algorithm with 6 units in the input layer, 4 or 8 neurons in the hidden layer and one output neuron for unfertilized variants, and 10-4-1 and 6-8-1 network topologies for foliar fertilized variants. Advantages of neural computing techniques relied on faster computation, learning ability and noise rejection.