Advancing prediction of biophysical and biochemical traits in potatoes using hyperspectral data and artificial intelligence

Abstract

Optimizing nitrogen (N) management is fundamental for enhancing crop productivity and mitigating environmental impacts in potato (Solanum tuberosum L.) cultivation. Traditional approaches for quantifying plant N uptake and biomass are labor-intensive and destructive, necessitating innovative remote sensing techniques. This study integrates hyperspectral sensing with machine learning (ML) and deep learning algorithms to estimate plant N uptake, biomass accumulation, and predict tuber yield. The hyperspectral data (400–2500 nm) was collected at multiple potato growth stages from an N management study conducted over two growing seasons (2023–2024) at two locations. The study compared three spectral preprocessing methods to optimize model performance: raw spectra, Savitzky–Golay filtering, and first derivative (FD) transformation. Six predictive models were evaluated, including support vector regression, partial least squares regression, random forest regression, ridge regression (RR), least absolute shrinkage and selection operator regression, and a one-dimensional convolutional neural network (1D-CNN). FD preprocessing enhanced estimation accuracy, with the 1D-CNN model achieving the highest performance for N uptake (R2 = 0.82) and biomass estimation (R2 = 0.84), outperforming traditional ML models. However, for tuber yield prediction, RR provided the best performance (R2 = 0.67). SHapley Additive exPlanations analysis identified key spectral regions in the spectrum that contributed to model predictions. The study demonstrates that hyperspectral data, coupled with AI-driven predictive modeling, has the potential to improve N-use efficiency and optimize fertilizer applications, thereby enhancing sustainability in potato production.