Attention-Based CNN Ensemble for Soil Organic Carbon Content Estimation With Spectral Data

At present, deep learning method that relies on its strong feature extraction ability has been successfully applied to the estimation of soil organic carbon (SOC) content with hyperspectral data. However, due to the high dimensionality of hyperspectral data and equal treatment of all bands, the performance of these methods is hampered by learning features from useless bands. To address this issue, in this letter, attention mechanism is combined with convolutional neural network (CNN) to assign different weights to different bands of the hyperspectral data. This method constructs a three-layer CNN with a squeeze-and-excitation module at the front of it. Then, five attention-based CNNs are combined to establish an ensemble regression system with diversity. The inputs of each branch in this system are the original hyperspectral data and its transformed data. Moreover, an improved label distribution smoothing (ILDS) technique is proposed to address the problem of imbalanced samples. The experimental results on three soil datasets, Land Use/Land Cover Area Frame Survey (LUCAS) 2009, LUCAS2015, and Africa Soil Information Service (AfSIS), show that this method obtains good estimation performance compared with several state-of-the-art methods, especially in the areas with high SOC content which has small sample sizes.