Estimation of Soil Property Content With Vis–NIR Spectra by Multitask Deep Learning Based on Attention Mechanism and Loss-Weight Balancing

Currently, deep learning methods have been successfully applied to soil property content estimation from soil spectra due to their powerful feature extraction capability. In practical production, it is necessary to estimate the contents of multiple soil properties simultaneously. The accuracy of such estimation heavily depends on the ability of the algorithm to balance multiple estimation tasks. In this letter, a multi-task learning network combining attention mechanism and loss-weight balancing approach based on feature correlation is proposed. First, a parameter-sharing module of a three-layer convolutional neural network (CNN) is constructed. Second, an independent channel importance recalculation module is constructed for each estimation task, which consists of an efficient channel attention (ECA) module. Finally, the features extracted from these two modules are concatenated, and a two-layer CNN is constructed to further extract features for estimating each soil component. Moreover, an improved loss-weight uncertainty technique based on the correlation between soil spectra and property contents is proposed to reconcile the learning effects of multiple estimation tasks. The experimental results on two soil datasets, LUCAS (Land Use/Land Cover Area Frame Survey) 2009 and AfSIS (Africa Soil Information Service), show that this method provides competitive accuracy compared with several state-of-art methods.