Estimation of soil properties from the EU spectral library using long short-term memory networks

• Published in: Geoderma Regional Vol. 18; p. e00233
• Main Authors: Singh, Simranjit, Kasana, Singara Singh
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2019
• Subjects: Acrisols; Andosols; Calcisols; data collection; Deep learning; European Union; Gypsisols; Hyperspectral data; LSTMs; neural networks; organic carbon; PCA; polymerase chain reaction; principal component analysis; Quantification; soil properties; Deep learning; Quantification; Hyperspectral data; LSTMs; Acrisols; Gypsisols; Andosols; Calcisols; PCA
• ISSN: 2352-0094; 2352-0094
• DOI: 10.1016/j.geodrs.2019.e00233

Recent advances in Deep learning altered the traditional way of data analysis in various domains. This study focuses primarily on the quantification of soil properties from the hyperspectral data obtained from LUCAS using Long Short-Term Memory Networks based deep learning model. Hyperspectral data is sequential in nature due to which LSTMs are preferred as these are predominantly used for sequential problems. In addition to this, Principal Component Analysis is also performed on the LUCAS dataset to reduce the dimensions for efficient calibration of LSTMs. Better performance is achieved using the proposed framework with best R2 of 0.94 in case of Organic Carbon and the effectiveness is shown by comparing it with existing models such as PLSR, SVR, PCR, MLR and SWR. •The proposed framework is based on PCA and Deep learning based Long Short-Term Networks.•Due to high dimensions, PCA is used to reduce the dimensions of hyperspectral data.•The proposed method outperformed the popular regression models.