Modelling the Canopy Conductance of Cocoa Tree Using a Recurrent Neural Network

Abstract

Direct measurement of crop water use is difficult and labour intensive. In some cases, the technicalities involved can only be exploited by well-trained researchers. Therefore, estimating this important crop parameter from readily available climatic data by way of modelling will ease the burden of direct measurement. The aim of the study is to parameterize models of canopy conductance of rain-fed cocoa tree, suitable for inclusion in physically-based model for predicting water use of cocoa trees. To do this, Sap flow density was monitored in three cocoa trees (Forestaro cultivar group) at the eight (8) year old cocoa plantation of the Federal University of Technology, Akure, Nigeria (7° 18' 15.9"N, 5° 07' 32.3"E), from 8th March 2018 to 7th March 2019, covering the two seasons of the region. Cocoa tree transpiration was determined from the measured sap flow and fitted into a physically based model (PM) to derive canopy conductance used for modelling. To choose the best model that predicts canopy conductance (the stomata control of water transport) in cocoa trees, Vector Autoregressive Models (VAR), a multivariate time series model, and Long Short-Term Memory (LSTM) network, an Artificial Intelligence (AI) model were employed. The prediction power of the VAR model was assessed and visualized using the vars R package, while the LSTM model, a Recurrent Neural Network (RNN) algorithm was implemented using Python programming within Google COLAB jupyter notebook. Before modelling, data were tested for stationarity using the Augmented Dickey-Fuller test. While two-thirds of the data were used to train the models, the remaining one-third of the data were used to test the trained model. As VAR models were evaluated using R-squared and Root Mean Squared Error (RMSE), LSTM was evaluated by comparing the train loss and test loss, and also RMSE. VAR (with Adjusted R-Squared=0.11) is found not to be suitable to model the complex relationship between canopy conductance and climatic variables. Further iteration to exclude insignificant climatic variables from the VAR model did not also improve the model. However, LSTM with RMSE of 0.026 and having the test loss not dropping below the training loss was observed to perform better in modelling the canopy conductance of Cocoa. The result of the research further revealed that temporal dynamics of transpiration is complex and difficult to be defined by traditional regression. LSTM with a prediction accuracy of 97.4% could therefore be used for the prediction of cocoa canopy conductance.