Backpropagation Neural Network for Soil Moisture Retrieval Using NAFE’05 Data: A Comparison of Different Training Algorithms
S. Chai, B. Veenendaal, G. West and J.P. Walker
Artificial Neural Networks have been widely applied to solving real world problems of different complexity. The Backpropagation Neural Network (BPPNN) is one of the most common neural network structures, as it is simple and effective. It has been used extensively for soil moisture retrieval problems. Examples include Atluri et al. (1999); Yuei-An et al. (2001); Angiulli et al. (2004). Although plausible soil moisture results have been obtained, it is difficult to compare results and methds because different training algorithms for the BPPNN have been use for many different problem domains in soil moisture retrieval ranging from smooth soil, bare soil to rough and dense soil. For a specific soil moisture retrieval problem domain, there has been no discussion on the selection of BPPNN training algorithms. Training algorithms have been chosen randomly without justification of the reasons and benefits of the chosen training algorithm over other training algorithms. In addition, much use has been made of simulated data for training and testing purpose with noise added to show robustness and to verify the accuracy of the testing results for the BPPNN. This paper will address the question of the selection of a particular training algorithm for the BPPNN by comparing the accuracy obtained for different types of BPPNN training algorithms using real ground truthed data from the National Airborne Field Experiment 2005 (NAFE’05). The H-polarized brightness temperature and the ground average temperature data are used as input to train the different BPPNN algorithms. Results obtained are between the range of 3.93% to 5.77% of root mean square mean error (RMSE) of soil moisture retrieval.