Multiscale feature enhancement and lightweight ensemble modeling for hyperspectral chlorophyll inversion in greenhouse tomato

Chlorophyll content measured by a Soil and Plant Analyzer Development (SPAD) meter is a key indicator of nitrogen status and photosynthetic capacity in greenhouse-grown tomatoes. However, hyperspectral data collected under greenhouse conditions are strongly affected by leaf posture, illumination variability, high-dimensional redundancy, and multicollinearity, which make small-sample modeling unstable To address these challenges, this study proposes an advanced and lightweight inversion framework integrating multiscale spectral enhancement, deep latent compression, ensemble modeling, and output calibration. A total of 240 leaf spectra (450–950 nm) were processed using Savitzky-Golay (SG) smoothing, fractional-order differentiation (FOD), and Morlet-L7 continuous wavelet transform (CWT) to enhance chlorophyll-sensitive structural features. A convolutional autoencoder (CAE) was used to extract 64-dimensional latent representations, which were fused with red-edge parameters, vegetation indices, and wavelet statistics to form a multi-source feature set. Support vector regression (SVR), gradient boosting regression tree (GBRT), kernel ridge regression (KRR), partial least squares regression (PLSR), and a lightweight Lightformer model were trained, and their out-of-fold (OOF) predictions were integrated through Ridge Stacking, followed by linear calibration. The proposed “Stacking + LinearCal” framework achieved R² = 0.782, RMSE = 1.451, and RPD = 2.156 on the independent test set (n = 72), outperforming all single models. SHAP analysis showed that CAE features, red-edge slope, red-edge inflection point (REIP), and near-infrared tail statistics within 940–950 nm contributed most to prediction. The framework demonstrates high accuracy, stability and interpretability, providing a practical basis for nutrient monitoring in greenhouse tomato production.