Rapid Prediction of Soil Available Potassium Using Hyperspectral Data and Convolutional Neural Networks

Rapid and accurate acquisition of soil composition information is essential for efficient land resource management and precision forestry. This study proposes a convolutional neural network (CNN) model to predict soil available potassium content using hyperspectral data. The goal is to establish a low‐cost, low‐risk, and environmentally friendly method for soil nutrient estimation. The main contributions of this work are as follows: (1) Three CNN models with different network structures and loss functions are developed and compared in terms of predictive performance. (2) A principal component regression (PCR) model is constructed as a baseline for comparison, highlighting the improvements achieved by deep learning methods. (3) To enhance the generalization ability of the CNN, a regularization term based on the L2 norm of the difference between adjacent weights is added to the mean square error loss function. This promotes smoother transitions between neighboring spectral bands and maintains spectral continuity. Experimental results on Eucalyptus plantation soil from the Huangmian Forest Farm show that the enhanced CNN model significantly outperforms the PCR model, achieving a root mean square error (RMSE) that meets practical application requirements. This research demonstrates the potential of CNN‐based approaches for rapid, accurate, and cost‐effective prediction of soil nutrient content. Moreover, the model’s performance was found to be competitive when compared with literature‐reported approaches, highlighting the robustness of the proposed method.