Prediction of wetland soil carbon storage based on near infrared hyperspectral imaging and deep learning

Abstract

To enhance the accuracy of predicting wetland soil carbon storage and address the issues associated with traditional time-consuming, labor-intensive, and costly detection methods, we adopted a novel approach integrating near-infrared hyperspectral imaging technology with deep learning. Initially, near-infrared hyperspectral imaging was utilized to capture spectral images, from which soil reflectance features were extracted. Subsequently, the data processing phase involved extracting average spectral band reflectance values, serving as inputs for the model with carbon storage values, detected through chemical methods, used as outputs. The continuous projection algorithm (SPA) was applied to eliminate redundant band information, thereby retaining only the bands correlated with carbon storage. In the model establishment phase, four different preprocessing methods were applied to build the models, followed by the utilization of an improved particle swarm optimization (PSO) algorithm to optimize the long-short term memory (LSTM) neural network. It was the improved PSO-LSTM model that yielded significant predictive performance, achieving a coefficient of determination (R2) of 0.95, a root mean square error (RMSE) of 1.23%, and a ratio of performance to deviation (RPD) of 4.47. This method provides crucial support for the rapid, accurate, and non-destructive detection of soil carbon storage in modern wetlands, thereby contributing to the global carbon cycling and climate change studies and supporting future agricultural production efforts.