This study aims to predict the nutritional composition of pine needles using a dataset comprised of canopy-level needle images from 282 trees of 1-year-old Pinus elliottii × Pinus caribaea hybrids. The experiment was conducted at the Hongling Seed Orchard Nursery in Taishan, Guangdong Province, China (22.164769°N, 112.822761°E). The seedlings were selected and potted by experts in soil composed of loess and a lightweight substrate in a 6:4 ratio, with a pot height of 16 cm. Among the 282 samples, 162 were used for hyperspectral reconstruction, while the remaining 120 were reserved for final nutrient prediction. Unlike many previous studies that rely on detached pine needles collected under laboratory conditions, this study preserved the structural integrity of the needles by conducting data collection in situ without any physical disruption (see Figure 1 ). Although the imaging was performed in a laboratory setting, the process maintained the original morphology of the samples. This approach enhances the ecological validity and generalizability of the resulting models. Consequently, models trained on this in situ dataset are more likely to accurately reflect real-world field conditions and maintain strong predictive performance in practical applications.

In this study, a visible and near-infrared hyperspectral camera (Gaifields Pro-V10, Dualix Spectrum Imaging Co.) was used to capture hyperspectral images of pine needle canopies under controlled laboratory conditions. The spectral range spanned 400–1000 nm and comprised 176 spectral bands. During image acquisition, illumination was provided by a halogen ring light with a photosynthetic photon flux density (PPFD) ranging from 88 to 95 µmol/(m²·s) and a lighting uniformity of 7.6%.

To simulate natural daylight and enhance the model’s generalizability to field conditions, RGB images were generated from hyperspectral data using the CIE 1931 color matching functions and the CIE standard illuminant D65. D65, with a correlated color temperature of approximately 6500 K, closely approximates the spectral distribution of natural midday outdoor light, encompassing both direct sunlight and diffuse skylight components. These RGB images were then used as input features for the spectral reconstruction model. Although the data were collected indoors, applying the D65-based RGB transformation allowed us to approximate realistic field illumination conditions, thereby improving the robustness and applicability of the trained model in outdoor scenarios.

During training, hyperspectral images were evenly divided into two subsets: 88 spectral bands in the visible range (400–700 nm) and the rest in the near-infrared range (700–1000 nm), which were used as training targets. This division was motivated by two main factors. First, spectral reflectance patterns in the visible range exhibit greater variability than those in the near-infrared range; training the model separately on these two segments can help improve reconstruction accuracy. Second, hyperspectral images contain large amounts of spectral information, which can lead to high memory consumption and slow GPU computation during training. By splitting the data into two parts, we enable parallel training, thereby improving computational efficiency and accelerating the training process.

In this study, the processing of pine needle hyperspectral images first involved the extraction of regions of interest (ROIs) to suppress interference from the black background board and retain only the effective area of the pine needle. To achieve this, a threshold was applied to generate a mask, which was then used to extract spectral data for subsequent analysis. The average spectral reflectance of each channel was computed, yielding a one-dimensional array of 176 values. To minimize the effects of external conditions and instrument variability on spectral data, a thorough preprocessing pipeline was implemented. This pipeline included denoising, wavelength calibration, and spectral calibration to help ensure data accuracy and reliability and to enhance the precision of subsequent machine learning regression. Specifically, Multiple Scatter Correction (MSC) and First Order Differentiation (D1) were applied to improve spectral data quality and regression accuracy. To further reduce data dimensions and optimize model performance, the Competitive Adaptive Reweighted Sampling (CARS) algorithm was used to select representative spectral bands. CARS performs iterative sampling and PLS regression calculations to identify the most informative spectral variables, thereby improving the efficiency and robustness of data processing and model development.

In the regression stage of this study, samples were divided into training and testing sets at a ratio of 8:2, with the testing set reserved for final model evaluation. To predict nutrient content, we employed Partial Least Squares Regression (PLSR), a classical regression method that combines the dimensional reduction capabilities of Principal Component Analysis (PCA) with multivariate regression. PLSR identifies new components that maximize the covariance between independent and dependent variables, establishing a linear relationship between them. It is widely used in spectral analysis and nutrient modeling due to its ability to handle high-dimensional data and multicollinearity.

During training, ten-fold cross-validation was employed to enhance the model’s robustness and generalization capability. To comprehensively evaluate predictive performance, three key metrics were utilized: the prediction coefficient of determination (R²p), the root mean square error of prediction (RMSEP), and the relative prediction deviation (RPD), as defined in Equations 4-6. In these equations, S S r e s i d u a l denotes the sum of squared residuals—the squared differences between predicted and observed values—while S S t o t a l represents the total sum of squares, indicating the variability of the observed values around their mean. y ^ i and y ^ i represent the predicted and actual values, respectively. These metrics provide a direct assessment of the model’s generalization and predictive accuracy on unseen data. Specifically, a higher R²p, lower RMSEP, and larger RPD reflect improved predictive performance and greater practical utility. Therefore, these three metrics are used as the primary indicators for evaluating the final effectiveness of the model in this study.

The computations in this study were conducted using the cloud computing platform Autodl. The remote server was supported by a 15v CPU Intel(R) Xeon(R) Platinum 8474C, 80GB of memory, and an RTX 4090D (24GB). During the training process, the number of training epochs was set to 200, with each epoch containing 1000 iterations, the batch size of the training set was 2, the learning rate was 1e-4, and the dimensions of the dataset used for training were 768×768. The patch size was set to 128 with a stride of 8. For adjusting the learning rate, the Adam optimizer was employed with parameters β₁ = 0.9 and β₂ = 0.999. The analysis was conducted using Python 3.8.19. The reconstruction model was trained using open-source Python packages, including Scikit-learn, PyTorch, and OpenCV-Python.

The initial step in training the reconstruction model involved processing the input RGB images and preparing the corresponding hyperspectral data as ground truth. As detailed in Section 2.1, two segments of hyperspectral data were used as the reference, with the matching RGB images serving as inputs to the model. The training dataset comprised 140 images, while the validation dataset included 22 images. Training continued until the loss convergence criterion was satisfied, after which the best-performing model was selected for further evaluation. Table 1 summarizes the parameter count, computational cost, and validation performance of the four models across the 400–700 nm and 700–1000 nm spectral ranges.

In our study, the MIRNet model demonstrated strong performance in hyperspectral reconstruction, particularly across both the 400–700 nm and 700–1000 nm ranges. For the 400–700 nm range, MIRNet achieved a minimum mean relative absolute error (MRAE) of 0.4424, a root mean square error (RMSE) of 0.0349, and a peak signal-to-noise ratio (PSNR) of 29.22 dB. For the 700–1000 nm range, the corresponding values were 0.3233, 0.0531, and 25.65 dB, respectively. Additionally, MIRNet exhibited the lowest computational cost among the tested models, requiring only 58.15 gigaflops and featuring the smallest model size of 20.18 million parameters. These attributes significantly improved the algorithm’s efficiency and supported the deployment of future embedded applications.

To more clearly demonstrate the superiority of our proposed model, this section reviews and compares the performance of several representative methods in hyperspectral reconstruction tasks in recent years. These methods, including HSCNN-D, HRNet, and MPRNet, serve as intuitive and compelling benchmarks against which we can evaluate and highlight the advantages of our approach. In a study by Md. Toukir Ahmed et al. investigating agricultural product quality through hyperspectral image reconstruction, the HSCNN-D model was used to reconstruct 204 spectral bands (400–1000nm) from RGB images of sweet potato. The model achieved a mean relative absolute error (MRAE) of 0.8601, a root mean square error (RMSE) of 0.0545, and a peak signal-to-noise ratio (PSNR) of 26.86 dB on the validation set. On the test set, the corresponding metrics were 0.7914, 0.0566, and 25.91 dB, respectively (Ahmed et al., 2024b). Furthermore, in another study by the same team focusing on chicken embryo mortality, the HRNet model demonstrated improved reconstruction performance, yielding an MRAE of 0.0936, RMSE of 0.0151, and PSNR of 36.79 on the validation set (Ahmed et al., 2024a). However, this study reconstructed only 10 spectral bands within the 520–903nm range, which likely contributed to its higher reconstruction accuracy due to reduced spectral complexity.

Notably, in a study by Weiguo Yi et al. on geographical origin identification of beef, the MPRNet model was applied to reconstruct hyperspectral images (400–1000nm, 125 bands) from RGB inputs. The model performed exceptionally well, yielding an MRAE of 0.177, RMSE of 0.0165, and PSNR of 35.76 (Yi et al., 2024). The high reconstruction accuracy in this study may be due to the large proportion of beef regions within each image and their relatively flat surface characteristics, which minimize depth of field effects. Additionally, the small number of spectral bands to be reconstructed further facilitated accurate reconstruction. Overall, existing methods all demonstrate a certain ability to reconstruct hyperspectral images and are applicable to a range of datasets and application scenarios. However, they struggle to balance reconstruction accuracy with spectral resolution. Our approach effectively overcomes this limitation, offering a more robust solution that achieves both high reconstruction accuracy and strong spectral fidelity.

In the final prediction of nutrient components, spectral data within the 400–450 nm and 950–1000 nm ranges were deliberately excluded. This decision was based on two primary considerations: the high noise levels observed in these spectral regions and their limited relevance to the absorption peaks of the three essential nutrients—nitrogen, phosphorus, and potassium. Including such noisy and non-informative wavelengths could compromise prediction accuracy. To enhance both model efficiency and predictive performance, the spectral range of 450–950 nm was selected for nutrient content estimation. Following hyperspectral image reconstruction, spectral data within this range were extracted for subsequent regression analysis, and the quality of the reconstructed spectra was rigorously evaluated.

To assess model generalization, eight samples were randomly selected from the 120 test set samples, as shown in Figure 6 . The results demonstrate that MIRNet and MPRNet generally provided the best spectral alignment with the ground truth, whereas Restormer and HRNet exhibited more pronounced reconstruction errors across specific spectral bands.

Specifically, MIRNet consistently showed superior spectral fitting across most samples, particularly in wavelengths beyond 700 nm. In contrast, MPRNet exhibited notable fluctuations in this range, especially in samples c and h. HRNet performed exceptionally well on sample c; however, for other samples—particularly f and g—it displayed considerable error fluctuations beyond 700 nm, indicating variable reconstruction stability. Restormer exhibited inconsistent performance across all eight samples, with marked discrepancies observed in samples c, d, and g.

It is also worth noting that in samples a, c, h, and h, MIRNet slightly overestimated the 550 nm peak relative to the ground truth. Nonetheless, in the context of full-spectrum reconstruction, MIRNet demonstrated the most stable and accurate performance, while utilizing the fewest parameters and the least computational resources among all models evaluated. This strong performance may be attributed to MIRNet’s selective kernel fusion mechanism, which effectively integrates features from different receptive fields and enhances the model’s ability to manage spatial and channel resolution challenges.

To more intuitively illustrate the performance of hyperspectral reconstruction, visual analysis was incorporated into the evaluation process. The Mean Relative Absolute Error (MRAE) index was first used to generate error heatmaps, showing the differences between the reconstructed spectra at 700 nm and the ground truth spectra for four representative samples, as shown in Figure 7 . These heatmaps clearly depicted the spatial distribution of reconstruction errors across the pine needle regions for each model.

The visual analysis highlighted the superior performance of MIRNet, which exhibited minimal errors throughout the entire needle area. In comparison, MPRNet showed moderate errors, particularly in the needle regions of samples c and f. HRNet demonstrated more pronounced errors in samples d, f, and g, with a similar distribution pattern to MPRNet across other samples. Restormer performed comparably to MPRNet in sample f but exhibited substantial errors in other samples, especially sample c. Combined with the spectral curve analysis discussed earlier, these results suggest that Restormer struggles with reconstructing fine-grained regions in the sample data.

Additionally, MRAE-based heatmaps were generated to compare the reconstruction accuracy of the best-performing MIRNet model against the ground truth at 500 nm, 700 nm, and 900 nm wavelengths, as illustrated in Figure 8 . These heatmaps further revealed spatial variations in reconstruction quality across the pine needle structures. At 500 nm, significant errors were observed in the needle areas, likely due to surface smoothness and irregularities causing uneven light reflection and refraction, particularly in shadowed regions. At 900 nm, while the outer edges of the needles exhibited accurate reconstructions, the central regions showed larger errors. This is likely attributable to the limited depth of field during image acquisition, compounded by shadowing effects and strong near-infrared responses in dark areas not easily visible in the RGB spectrum. These findings emphasize the challenges of achieving uniform reconstruction across varying spatial and spectral characteristics.

In this section, the MIRNet model, which achieved the highest reconstruction performance, was employed for final regression prediction.

Figure 9 presents the hyperspectral reflectance curves of pine needle canopy samples with varying nitrogen concentrations across the 450–950 nm spectral range. The reflectance curves display distinct patterns corresponding to four nitrogen levels. Specifically, in the 450–600 nm region, reflectance decreases as nitrogen concentration increases. This trend is attributed to the positive correlation between nitrogen content and chlorophyll levels in plants. Chlorophyll strongly absorbs light in the 450–470 nm and 630–670 nm bands, leading to reduced reflectance and broader absorption features, particularly resulting in a diminished green peak around 550.8 nm.

These spectral variations underscore the sensitivity of hyperspectral signals to nutrient content and validate the potential of hyperspectral data for non-destructive estimation of nitrogen levels in pine canopies.

Regions of interest (ROIs) were extracted from the reconstructed spectral data, followed by spectral preprocessing. In this study, three preprocessing methods were applied: Multiplicative Scatter Correction (MSC), first derivative (D1), and a combination of D1 and MSC (D1+MSC). The original spectra and the results after applying these preprocessing techniques are illustrated in Figure 9 .The MSC method enhances spectral bands with greater variability by computing the mean spectrum and correcting the original spectra through offset adjustments. This process effectively reduces spectral differences in the 800–987 nm range, corrects scattering effects, and improves the linearity of chemical information. The D1 method eliminates baseline drift and noise interference while enhancing spectral resolution and detail. The combined D1+MSC approach integrates the advantages of both methods, optimizing the signal-to-noise ratio and feature separability, thereby significantly improving the accuracy and reliability of multi-component detection in complex systems.

Following preprocessing, the Competitive Adaptive Reweighted Sampling (CARS) algorithm was employed to extract characteristic wavelengths, aiming to eliminate redundant bands and enhance predictive accuracy. These selected wavelengths serve as key indicators in spectral analysis. By optimizing the weight of data at these specific wavelengths, both spectral reconstruction and nutrient content prediction performance can be substantially improved. The extracted spectral features were then used to further evaluate the MIRNet model. Additionally, it was observed that the nutrient content values used as labels had a relatively narrow range, particularly for phosphorus, whose absolute values were very small. This limited variation amplifies the impact of noise and may degrade model performance. To mitigate this effect, target variables were standardized during the training of machine learning regression models, thereby improving model robustness. Finally, grid search was employed to optimize model parameters and reduce the risk of overfitting.

The final regression results are presented in Table 2 , highlighting the predictive coefficient of determination (R²p), root mean square error of prediction (RMSEP), and residual predictive deviation (RPD), which collectively evaluate the model’s performance on the independent prediction set. Among the models tested, the PLSR model demonstrated the best overall performance in this study. Using the reconstructed spectral data, the D1 preprocessing method achieved the highest R²p value of 0.8523 for nitrogen prediction. For phosphorus prediction, the MSC preprocessing method yielded the best performance with an R²p of 0.7022. In the case of potassium prediction, the MSC method also achieved the highest R²p value of 0.8087. In this study, PLSR was selected as a primary regression model due to its inherent advantages in handling high-dimensional, collinear spectral data. Compared with Support Vector Regression (SVR) and Random Forest (RF), PLSR exhibits superior robustness and stability under limited sample conditions, while simultaneously integrating dimensionality reduction and regression. Its minimal parameter tuning requirements and strong interpretability—through regression coefficients and VIP scores—further enhance its suitability for hyperspectral analysis. These characteristics make PLSR a particularly effective and widely adopted approach in spectroscopic modeling for nutrient estimation tasks.

In related studies, Li used hyperspectral data in the range of 1,100–2,500 nm with a spectral resolution of 8 nm (totaling 175 wavelengths) to determine nitrogen concentrations in different tissues of slash pine, yielding a coefficient of determination (R²cv) of 0.66 in cross-validation (Li et al., 2022). Shu Meiyan et al. employed spectral decomposition methods to estimate the nitrogen status of maize leaves and identified sensitive wavelengths at 470 nm, 538 nm, 638 nm, 682 nm, 710 nm, 734 nm, and 830 nm. Using a CARS-SVR regression model, they achieved a R² of 0.68 on the test set (Meiyan et al., 2023). In another study, Yakun Zhang conducted in situ, non-destructive detection of nitrogen content in soybean leaves using hyperspectral imaging in the 380–1000 nm range. The model attained a maximum R²p of 0.9428 on the validation set. However, it is noteworthy that the nitrogen content range in that study was relatively wide (8.275–44.724 mg/g), and the hyperspectral data included 707 spectral bands—factors that likely contributed to the model’s high performance (Zhang et al., 2024). Similarly, Di Lin et al. used in situ hyperspectral data (400–900 nm) to predict nitrogen, phosphorus, and potassium content in the maize canopy. Their model achieved validation R² values of 0.821, 0.732, and 0.773 for nitrogen, phosphorus, and potassium, respectively (Lin et al., 2024). These results are comparable to those obtained in the present study. Considering that the hyperspectral data used in this research were reconstructed rather than directly acquired, the findings strongly demonstrate the effectiveness and reliability of the proposed methodology.

Finally, each element was evaluated using its respective optimal model on both ground-truth and reconstructed test sets, and the resulting regression performances are illustrated in Figure 10 . The figure highlights the differences in the R²p values among the three elements. It is noteworthy that in this study, the prediction performance for phosphorus based on reconstructed spectra was slightly superior to that achieved using ground-truth hyperspectral data. This phenomenon may be primarily attributed to the inherently low phosphorus concentration—typically ranging between 0.8 mg/g and 1.6 mg/g—as well as its distribution across both needle and root tissues. Compared to nitrogen and potassium, phosphorus exhibits a more dispersed spatial distribution, which increases the uncertainty in machine learning-based regression. Nevertheless, this outcome indirectly supports the reliability and robustness of our reconstructed spectral data.

As detailed in Section 2.2, the original MIRNet model reconstructed hyperspectral images within the 400–700 nm range, encompassing 31 spectral bands. In this study, we extended the spectral range to 400–1000 nm by utilizing 176 bands, thereby incorporating the near-infrared (NIR) region. While most existing research and publicly available datasets on hyperspectral reconstruction focus primarily on the visible spectrum (<700 nm), many hyperspectral imaging studies have already expanded into the NIR or even infrared regions. As a result, although hyperspectral reconstruction approaches offer cost advantages, their performance often lags behind that of direct hyperspectral imaging, particularly in the extended spectral range. Despite the inherent challenges of reconstructing NIR spectra, our approach achieved low-error reconstruction across the full 400–1000 nm spectrum.

For comparative analysis, we applied the original MIRNet model to RGB images of pine canopy needles to reconstruct hyperspectral images across 31 bands (400–700 nm), as summarized in Table 3 . The reconstructed spectral data were then used for nutrient content prediction to evaluate the model’s effectiveness. Spectral preprocessing and feature band selection methods were kept consistent. The results showed that the highest R² values on the prediction set for nitrogen, phosphorus, and potassium were 0.8159, 0.5040, and 0.7007, respectively—representing decreases of 4.3%, 28.2%, and 13.4% compared to the improved reconstruction model. These findings emphasize the importance of both enhancing the reconstruction model and expanding the spectral range.

In Section 3.3, we observed that nearly all samples exhibited increased reconstruction errors in the central region at the 900 nm band. This phenomenon is likely due to image blurring caused by depth-of-field effects during image acquisition. To test this hypothesis, we applied Gaussian blurring with varying kernel sizes to the central regions of samples b, c, and f, where the issue was most pronounced. These blurred images were then used for hyperspectral reconstruction, and the resulting errors at 900 nm were visualized using heatmaps, as shown in Figure 11 . From left to right, the images represent the original sample and those blurred with Gaussian kernels of sizes 5, 15, 25, and 35, respectively. The heatmaps clearly indicate that as blurring intensity increases, reconstruction errors in the central region become more severe. The average MARE for the three samples increased from 0.0124, 0.0122, and 0.0215 to 0.0154, 0.0141, and 0.0256, respectively, supporting our hypothesis.

The reconstructed spectral reflectance curves in Figure 12 further demonstrate the impact of blurring. For samples b and c, the spectral range from 450 to 750 nm remained largely unaffected. However, beyond 750 nm, sample b exhibited greater sensitivity to blurring than sample c. When the Gaussian kernel size exceeded 5, the reconstructed spectra for sample b consistently fell below the original values, whereas sample c maintained stable spectra across all levels of blurring. This robustness is likely attributed to the smaller surface area of sample c.

In contrast, sample f showed significant sensitivity to blurring in the 500–600 nm and 750–950 nm ranges. Notably, reconstruction errors increased markedly when Gaussian kernels of sizes 5 and 15 were applied. However, with larger kernel sizes, the errors in both spectral ranges gradually stabilized. Importantly, for all three samples, reconstruction errors remained nearly unchanged when Gaussian kernel sizes reached 25 and 35, demonstrating the MIRNet model’s remarkable robustness in handling highly blurred images.

Given the structural characteristics of pine needle canopies, depth-of-field effects during image acquisition are unavoidable. However, this limitation also presents valuable insights for future research. Since the ultimate goal of this study is to deploy the model on external agricultural equipment, real-world image captures will inevitably involve variations in equipment, camera angles, and lighting, leading to varying degrees of image blurring. Therefore, future research will focus on integrating spatial and spectral super-resolution techniques to enhance the quality of original RGB images, thereby improving the overall accuracy of hyperspectral reconstruction.