Soil Organic Carbon Prediction Using an Efficient Channel Attention-Enhanced CNN-LSTM Model With LUCAS Spectral Library

Abstract

Visible near-infrared reflectance spectroscopy (Vis–NIR) has been widely used in soil organic carbon (SOC) prediction due to its rapid, cost-effective, and non-destructive characteristics. Numerous soil spectral libraries have been used for SOC prediction. However, the growing volume of Vis–NIR spectral data has amplified its complexity, high dimensionality, and nonlinearity, creating significant challenges for traditional analytical models, particularly in terms of feature extraction, prediction accuracy, and generalisation capacity. To address these limitations, we developed a novel hybrid deep learning model that synergistically combines an enhanced convolutional neural network (CNN), a long short-term memory (LSTM) network, and an efficient channel attention (ECA) mechanism, termed the CNN-LSTM-ECA model. The CNN-LSTM-ECA model was evaluated using the LUCAS spectral library. Additionally, the SOC prediction performance of the CNN-LSTM-ECA model was compared against that of the CNN and CNN-LSTM models. To further assess the predictive performance of the model, spectral data specific to France were extracted from the library for validation. The results show that the CNN-LSTM-ECA model significantly outperforms the CNN and CNN-LSTM models in SOC content prediction. Specifically, the proposed model achieved remarkable prediction accuracy with an R² of 0.92 and an RMSE of 25.07 g kg⁻¹ on the validation, representing significant improvements of 10.72% and 7.15% in RMSE compared to the CNN (RMSE = 28.08 g kg⁻¹) and CNN-LSTM (RMSE = 27.00 g kg⁻¹) models, respectively. The model's generalisation capability was further confirmed through additional testing on the French dataset, where it maintained consistent predictive performance (R² = 0.93, RMSE = 24.83 g kg⁻¹). These findings underscore the model's high prediction accuracy and robust generalisation across diverse datasets. This study illustrates that the CNN-LSTM-ECA model significantly improves both accuracy and generalisation in SOC prediction, thereby providing a promising approach for spectral data analysis.