Review Reports

Comments 1: [I noticed some grammatical errors in writing, therefore, I suggest the revision of English by a native speaker.]

Response 1: We thank the reviewer for pointing out the need for language improvement. The manuscript has been professionally edited by American Journal Experts (AJE), and a certificate of editing is provided with this submission. Furthermore, we have conducted an additional round of careful proofreading to correct any remaining grammatical errors. We believe the manuscript now meets the high standards of academic English.

Comments 2: [Authors must reformulate the abstract. Note that you are presenting an abstract with 347 words, and Horticulturae limits an abstract to 200 words. I also highlight that authors must follow the premise of presenting the highlights of the results in the abstract, something that I did not observe in this summary.]

Response 2: We thank the reviewer for the constructive feedback regarding the abstract. We have thoroughly revised it to strictly adhere to the journal's 200-word limit. Additionally, we have restructured the content to prominently highlight the key results of our study, ensuring that the main findings are clearly and concisely presented. The revised abstract now accurately reflects the highlights and significance of our work.

Comments 3: [The introduction cites limitations of the direct use of spectral indices and canopy structural variation, but it does not clearly differentiate to what extent the adjustment of the Average Leaf Angle (ALA) represents an advance over previous studies.]

Response 3: We thank the reviewer for this insightful comment. We have revised the Introduction (lines 76-82) to explicitly clarify how the adjustment of the Average Leaf Angle (ALA) represents a specific advancement over the conventional use of PROSAIL parameters. The added text clearly differentiates our methodological contribution and highlights its significance in addressing the limitations of prior approaches.

Comments 4: [The literature review is extensive, but it could include recent works (2023–2025) that explore PROSAIL adjustments for perennial fruit crops.]

Response 4: We thank the reviewer for this suggestion. We agree that including the most recent literature would strengthen the introduction. Our review indeed confirmed that the application of PROSAIL, particularly with parameter adjustments, remains predominantly focused on annual crops like wheat and maize, with very limited studies on perennial fruit crops. This gap effectively highlights the novelty and contribution of our work. In response, we have revised the introduction to explicitly frame this context (lines 85-89). We believe this now provides a clearer and more compelling rationale for our study.

Comments 5: [Elements of direct agronomic impact are lacking, such as implications for nitrogen management or precision irrigation.]

Response 5: We thank the reviewer for this insightful comment regarding the agronomic impact of our work. In direct response, we have added a discussion on the practical implications of our findings for nitrogen management and precision irrigation in the revised manuscript (lines 39-42). This addition explicitly links our methodological advancements to their direct applications in orchard management.

Comments 6: [Although flight dates and conditions are described, information on spatial variability is missing (how many hectares of the orchard were covered, representativeness of the selected trees).]

Response 6: We thank the reviewer for this comment. The requested information regarding the spatial extent and representativeness of the sampled trees has been added to the manuscript (lines 126-129).

Comments 7: [The use of Pix4D, ENVI, and ArcGIS is mentioned, but there is no detail on how anisotropy effects, shadows, or atmospheric corrections were handled.]

Response 7: We thank the reviewer for raising this important methodological point. We have thoroughly revised the manuscript to provide the requested details. Specifically:

Atmospheric and Anisotropic Correction: The specific workflows for atmospheric correction (QUAC) and anisotropy correction (BRDF) have been clearly detailed in the revised manuscript (lines 150-153).

Shadow Mitigation: To address potential shadow effects, we have clarified that all UAV flights were conducted around solar noon (line 144). Furthermore, given that the experimental area is flat and open, and that leaf sampling was intentionally restricted to sun-exposed top-layer canopies, the influence of shadows was considered minimal and thus not a primary focus of our correction procedures.

We believe these additions provide the necessary technical clarity regarding our image processing chain.

Comments 8: [The Cerovic model (R² = 0.94) was adopted directly, but there was no local validation. This may compromise accuracy, since regional calibrations are sensitive to species, leaf age, and edaphoclimatic conditions.]

Response 8: We thank the reviewer for raising this important point. Our decision to directly apply the Cerovic et al. model was based on its reported high accuracy (R² = 0.94) and noted applicability across a diverse range of plant species, with particularly strong performance in dicotyledons, as established by its developers. We have clarified in the text that this approach was chosen specifically because it aligns with our objective of developing a rapid, non-destructive method. We acknowledge that localized calibration could further optimize accuracy and have framed this as an important direction for future work. This is particularly relevant as our team currently lacks the facilities for destructive chlorophyll analysis, making local validation beyond the scope of this initial study. Acquiring this capability is a defined objective for our subsequent research.

Comments 9: [The process for adjusting ALA and fusing it with PROSAIL needs to be described more transparently (parameters adjusted, bounds considered, how many iterations were required).]

Response 9: We thank the reviewer for this suggestion. The requested details regarding the ALA adjustment process have been added to the manuscript for greater transparency (lines 254-259).

Comments 10: [The choice of metrics (R², RMSE, MAE) is appropriate, but there is no indication of how confidence intervals were calculated.]

Response 10: We thank the reviewer for this insightful comment. In response, we have now explicitly stated that the 95% confidence intervals for R² and RMSE were calculated using the bootstrap resampling method with 1,000 replicates, as described in the revised manuscript (lines 304-306). This approach was implemented to quantify the uncertainty associated with the model performance estimates.

Comments 11: [ALA is optimized, as the authors show the performance gain after adjusting the mean leaf angle, but this gain is not quantified in percentage terms of improvement, which would be crucial to demonstrate practical relevance.]

Response 11: We thank the reviewer for this valuable suggestion to better quantify the practical relevance of our findings. In direct response, we have now incorporated the percentage improvement terms in both the abstract (line 28) and the conclusion section (line 605) to explicitly highlight the performance gain achieved through ALA adjustment.

Comments 12: [Discussion is superficial in some points. As an example, the variability of SPAD (38–74) and LCC (32–99 μg/cm²) is presented without discussing whether such ranges reflect different managements or only natural variation.]

Response 12: We thank the reviewer for this insightful comment regarding the depth of our discussion. We fully agree that exploring the underlying causes of the observed variability is crucial. In response, we have substantially revised the discussion section (lines 549-557) to provide a reasoned interpretation of the SPAD and LCC ranges. We now explicitly address this point, arguing that the observed variability primarily reflects natural heterogeneity within the orchard rather than differential management practices, and we provide the supporting rationale for this interpretation.

Comments 13: [There is no exploration of how this model could be transferred to other citrus species or to different phenological stages.]

Response 13: We thank the reviewer for this insightful comment regarding the model's transferability. We agree that exploring generalizability across species and phenological stages is an important direction for future research. In our study, we made a strategic focus on a single key phenological stage to establish the fundamental viability of our hybrid inversion approach. Preliminary analysis indicated that spectral differences at the canopy scale across major citrus species are minimal compared to the interspecies variability, which is why cross-species transferability was not the primary focus. However, we fully acknowledge that phenological stage is a critical factor, as both the PROSAIL simulations and measured data would differ substantially across developmental phases. To address this point directly, we have revised the discussion to explicitly state these considerations and to frame the investigation of model transferability to different phenological stages as a clear and prioritized objective for our future work.(lines 583-588)

Comments: [As a minor and main note, I highlight:

1 – Use the Mendeley Reference Manager for references as well as citations, as both Horticulturae standards are not standardized in the body of every manuscript.]

Response: We thank the reviewer for this important note. We have now thoroughly reformatted the entire manuscript, including all in-text citations and the reference list, to strictly adhere to the specific formatting standards of Horticulturae using Mendeley. We confirm that the references now fully comply with the journal's style guidelines.

Comments 1: [The most critical issue is possible data leakage in your experiment. You estimate ALAadj from an RFR model with measured spectra, LAI, and LCC as inputs (ALAadj = RFR(SPEC, LAI, LCC)) and utilize this ALAadj to create spectra which you blend with measured spectra for training your inversion models. I need clear guarantee that the samples used for ALA derivation were entirely independent of those used to train and test the hybrid models. If independence is not imposed, you need to re-run experiments with strict partitioning of data.]

Response 1: We thank the reviewer for raising this critical point regarding data leakage, which we agree is paramount for validating our methodology. We can provide a clear guarantee of complete independence between the datasets. The samples used for ALA derivation came exclusively from a separately generated set of 10,000 PROSAIL simulations. In this process, PROSAIL was run with a broad range of ALA inputs to create a standalone spectral library. This library was used only for deriving the ALAadj value and was never blended with or exposed to the measured spectra used for training and testing the final hybrid inversion models. Therefore, the data partition is inherently strict, and there is no risk of data leakage as the two datasets were generated and used for entirely separate purposes.

Comments 2: [With your optimum 1:4 ratio, your previous models are trained on 80% synthetic data. This risks having algorithms learn PROSAIL rules idealized and minimal real world spectral complexity. I suggest making this limitation clear and reporting the ratio as a pragmatic accommodation for this study specifically and not an across-the-board optimum.]

Response 2: We thank the reviewer for this insightful comment regarding the potential limitations of a data blend dominated by simulated data. We agree that this is an important consideration. In response, we have now explicitly acknowledged in the study that a high proportion of simulated data may risk learning idealized PROSAIL rules, and we have framed the 1:4 ratio as a pragmatic accommodation for this specific study context, rather than a universal optimum.(lines 21-23, 399, 598)

Comments 3: [You chose the 1:4 mixing ratio from the PLSR experiments only. As best blending depends on the algorithm, I require robustness evidence across all four algorithms. Either provide performance vs. mixing ratio plots for all algorithms or use a nested cross-validation procedure to choose the ratio in an algorithm-independent way.]

Response 3: We thank the reviewer for highlighting the need for algorithm-independent robustness evidence. We have now provided performance-versus-mixing-ratio plots for all four machine learning algorithms (see new Figure 6) to substantiate our choice. 

Comments 4: [Supply: (1) Cross-validation schemes and hyperparameter grids for ML models;]

Response 4: We thank the reviewer for this suggestion. As requested, we have now provided a detailed description of the cross-validation schemes and hyperparameter grids for all four machine learning models in the Methods section (lines 362-371). This includes the use of 5-fold cross-validation for model evaluation and the complete set of hyperparameters explored during grid search for each algorithm(Table 4).

Comments 5: [Supply:(2) Detailed train/test sampling procedures (random seeds, stratification);]

Response 5: We thank the reviewer for this comment.Part of the program is as follows：

def complete_data_pipeline(X, y, random_state=42):

    np.random.seed(random_state)

    n_simulated = 400  

    n_measured = 100   

    train_simulated = 350  

    train_measured = 50    

    data_source = np.array(['simulated'] * n_simulated + ['measured'] * n_measured)

    simulated_indices = np.where(data_source == 'simulated')[0]  

    measured_indices = np.where(data_source == 'measured')[0]    

    train_simulated_idx = np.random.choice(simulated_indices, size=train_simulated, replace=False)

    train_measured_idx = np.random.choice(measured_indices, size=train_measured, replace=False)

    train_idx = np.concatenate([train_simulated_idx, train_measured_idx])

    test_idx = np.setdiff1d(np.arange(len(X)), train_idx)

    train_weights = np.zeros(len(train_idx))

    train_simulated_count = sum(train_idx < n_simulated)

    train_measured_count = sum(train_idx >= n_simulated)

    train_weights[train_idx < n_simulated] = 0.5 / train_simulated_count  

    train_weights[train_idx >= n_simulated] = 0.5 / train_measured_count  

    return X.iloc[train_idx], X.iloc[test_idx], y[train_idx], y[test_idx], train_weights, train_idx, test_idx

Comments 6: [Supply:(3) Correct record of the 350 simulated + 50 measured training samples selection]

Response 6: We thank the reviewer for emphasizing the need for precise documentation. We have now provided a correct and detailed record of the training sample selection in the manuscript (lines 440-449). This includes explicitly stating the composition of the training set (350 simulated and 50 measured samples) and describing the reproducible, stratified random sampling procedure used to select them.

Comments 7: [Add the following reference for Hybrid Inversion Model (Lines 97):

Improved PROSAIL Inversion via Auto Differentiation for Estimating Leaf Area Index and Canopy Chlorophyll Content

Exploring the Impact of Noise on Hybrid Inversion of PROSAIL RTM on Sentinel-2 Data]

Response 7: We thank the reviewer for the suggestion. The recommended reference for the Hybrid Inversion Model has been added to the manuscript (line 94).