Systematic evaluation of hyperspectral imaging workflows for predicting pigment and nutrient traits in tomato leaves under varying nitrogen supply levels

Abstract

Abstract Purpose Rapid and non-destructive detection of pigment and nutrient traits in tomato ( Solanum lycopersicum L.) leaves is essential for precision fertilization and greenhouse management. However, most existing studies focus on individual traits (e.g., chlorophyll or nitrogen) with isolated models, limiting the establishment of robust analytical workflows across growth stages and cultivation conditions. This study systematically evaluated hyperspectral imaging workflows for estimating pigment and nutrient traits in tomato leaves. Methods Hyperspectral images were collected at the stages of flowering-fruiting, ripening and harvest from tomato plants supplied with nitrogen at 0, 210, 300 and 390 kg N ha⁻¹. Total contents of chlorophyll, total nitrogen and nitrate were measured by standard biochemical assays for model calibration and validation. Result After sample partition with four strategies, seven spectral preprocessing methods were evaluated, namely moving average (MA), Savitzky-Golay smoothing (SG), Gaussian filtering (GF), median filtering (MF), normalization, baseline correction and standard normal variate (SNV), with normalization, MA and SNV yielded the best predictive performance for chlorophyll, total nitrogen and nitrate, respectively. For feature wavelength selection, competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) were used with CARS yielding the best performance for total chlorophyll and nitrate prediction, while SPA was optimal for total nitrogen prediction. By application of the above optimal methods, random forest (RF), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost) and convolutional neural network (CNN) models were developed to predict pigment and nutrient indicators. Conclusion The SVM performed best for chlorophyll ( R c ²=0.823, R p ²=0.431) prediction, while the CNN achieved higher accuracy for total nitrogen ( R c ²=0.826, R p ²=0.780) and nitrate ( R c ²=0.851, R p ²=0.753) prediction. Overall, leaf nitrogen-related traits were predicted more reliably than total chlorophyll, for which validation performance remained limited. Impact These findings demonstrate that sample partitioning, spectral preprocessing, wavelength selection and model selection should be optimized for each target trait rather than applied uniformly. This study provides a methodological basis for non-destructive assessment of tomato leaf N status and precision fertilization management.

Plant phenotyping relevance

トマト葉の色素・養分形質を対象に、ハイパースペクトル画像処理、波長選択、機械学習モデルを体系的に比較・検証しており、植物形質取得手法が研究の中心である。

Code and data availability

The paper describes hyperspectral imaging of tomato leaves and trait prediction models, but no public dataset, code, or model repository is provided. The Data availability statement says data are available only on request, so no public paper-specific asset qualifies.

No evidence-backed public reproduction asset is currently recorded.