A novel neural network model to achieve generality for diverse morphologies and crop science interpretability in rice biomass estimation
Tomoaki Yamaguchi, Keisuke Katsura
Abstract
Accuracy of crop biomass estimation based on images collected by remote sensing has been improved by the use of machine learning and image analysis techniques. However, the generality of the biomass estimation model for various plant morphologies and model interpretability have been challenges. The authors proposed a multi-traits-output-chained neural network (MTOC-NN) for biomass estimation that takes into account sub-target traits other than the main-target traits estimated by itself. MTOC-NN was expected to achieve high accuracy for various morphologies of rice plants by considering various aspects of the plant body to be estimated, and to be useful in terms of crop science interpretability by analyzing the importance of variables including sub-traits. First, it was investigated whether MTOC-NN could achieve higher prediction accuracy than multi-output neural networks, single-output neural networks, or other conventional machine learning models. The results showed that MTOC-NN had the lowest root mean squared error among all algorithms, indicating its high performance. Furthermore, the importance of each sub-target trait in biomass estimation was quantified by calculating the permutation importance. The results showed that the importance of the sub-target traits to leaf or stem biomass varied greatly among the varieties. From the above, we conclude that MTOC-NN is a model with great potential to provide a crop science interpretation while achieving high accuracy.