Assessment and transferability analysis of soil total nitrogen with different particle sizes based on proximal hyperspectral imaging

Abstract

Hyperspectral imaging serves as a powerful method for conducting efficient and non-invasive detection of total nitrogen in soil. Nevertheless, its complete potential remains underutilized due to the significant need for annotated samples and the influence of variations in soil spectral properties associated with different particle sizes on the generalization capacity of the model. Therefore, this paper proposes a Transfer Component Analysis Adaptive Enhanced Convolutional Neural Network (TACNN), enabling the transfer of soil total nitrogen assessment models between different soil particle size datasets. Six transferability strategies were developed to address diverse application scenarios, and the performance of TACNN was evaluated against TASVR, TAElman, as well as classical ensemble learning models AdaBoost-CNN, AdaBoost-SVR, and AdaBoost-Elman across six soil particle size datasets. The results indicate that classical ensemble learning models achieve satisfactory estimation of soil total nitrogen within the same particle size soil dataset, but fail to transfer across different particle size soil datasets. The combination of TACNN integration with model update demonstrates enhanced capability in estimating soil total nitrogen content across diverse particle size datasets. This research highlights the potential of transfer learning to reduce the dependence of soil total nitrogen assessment models on extensive sample datasets, thereby improving their generalization performance.