Integrating Three‐Channel Spectral Stacking and Mixup Augmentation Into Improved EfficientNet for Vis‐NIR Determination of Total Nitrogen and Total Phosphorus in Water

Accurate quantitative detection of total nitrogen (TN) and total phosphorus (TP) in water bodies is crucial for water quality assessment, yet extracting effective information from reflectance spectra remains challenging due to weak spectral signals and susceptibility to interference from complex aquatic backgrounds. To address the interference between spectral bands and overfitting in small samples, this study used 64 water samples collected from the estuary of the Moshui River in the Huaihe River Basin, Shandong Province. For each water sample, we measured 300–420 spectra, giving a total of 20,588 spectra, and utilized the 380–900 nm spectral band. This study proposes a deep learning spectral analysis method based on an enhanced EfficientNet (EfficientNet‐b0‐spec). The method utilizes a three‐channel spectral stacking strategy that integrates raw, Savitzky–Golay smoothed, and SNV‐normalized spectra into a 2D feature map to capture diverse characteristics. Additionally, a triplet attention mechanism is introduced to capture cross‐dimensional features, while the L‐SHADE algorithm optimizes a Huber–Welsch weighted combined loss function to enhance robustness against outliers. Furthermore, using Mixup data augmentation and ImageNet transfer learning alleviates the overfitting issue in small‐sample scenarios and accelerates model convergence. Experimental results show that the model performs well on both of the TN and TP prediction tasks. For TN, the model achieves an R 2 of 0.9693, with an RMSE of 0.4856 mg/L and an MRE of 10.33%. For TP, it reaches an R 2 of 0.9600, along with an RMSE of 0.0168 mg/L and an MRE of 10.94%. Compared to traditional machine learning and standard deep learning benchmarks, the proposed strategy demonstrates superior prediction accuracy and generalization capability.