CAT-MTLNet: A Novel Deep Multitarget Framework for Hyperspectral Soil Properties Estimation Using Concrete Autoencoder and Transformer

Due to their advanced feature extraction capabilities, deep learning (DL) models have shown significant success in estimating soil properties from hyperspectral data. However, challenges persist in accurately estimating multiple soil properties simultaneously while handling high-dimensional spectral inputs. Hence, this letter proposes a novel multitarget learning framework—concrete autoencoder and transformer-based multitarget learning network (CAT-MTLNet) to address these issues. It integrates a concrete autoencoder (CAE) for optimal spectral band selection with a transformer-multilayer perceptron (MLP) model for robust multisoil properties estimation. The framework jointly estimates soil organic carbon (SOC), phosphorus (P), pH, and calcium (Ca). Experimental evaluations on the Africa soil information service (AfSIS) and LUCAS-2009 datasets demonstrate the effectiveness of CAT-MTLNet. For AfSIS, $R {^{{2}}}$ scores range from 0.453 to 0.953 with root mean square error (RMSE) values between 0.289 and 0.390. For LUCAS-2009, $R {^{{2}}}$ scores range from 0.209 to 0.931, and RMSE varies from 26.094 to 39.006. CAT-MTLNet successfully balances interpretability, dimensionality reduction, and predictive performance, offering a lightweight and scalable solution for estimating hyperspectral soil properties.