Integration of Landsat-8 Thermal and Visible-Short Wave Infrared Data for Improving Prediction Accuracy of Forest Leaf Area Index

Round 1

Reviewer 1 Report

Dear authors,

I checked your revised manuscript. Regarding to my previous comments, I request the following correction.

1. Description about PV

 Regarding the following sentence in line 283, there are several published research papers about PV (forest cover) estimation. Therefore you must cite some of them. I recommend to discuss about accuracy of the current PV estimation which is enough for your work or not. The accuracy of LAI estimation using parameters from remote sensing alone in your model is under question. I feel that the author should make it clear to the readers.

“Hence, accurate estimation of the PV by means of remote sensing data”

Thank you very much for your effort.

END

Author Response

Regarding the following sentence in line 283, there are several published research papers about PV (forest cover) estimation. Therefore, you must cite some of them. I recommend to discuss about accuracy of the current PV estimation which is enough for your work or not. The accuracy of LAI estimation using parameters from remote sensing alone in your model is under question. I feel that the author should make it clear to the readers.

Answer: Thank you so much for the comment. It should be noted that the ‘Proportion of the vegetation cover,’ which is used in this study, was measured during the fieldwork for corresponding plots and is in situ measured data, and not estimated data. The in situ data usually considers as reference data for validation in the literature.

Meanwhile, as requested we have added a few references regarding the estimation of the Pv by means of spectral vegetation indices and machine learning approaches according to the literature.

Please see Page 15 Line 309-312

Hence, an accurate estimation of the P_V by means of remote sensing data for calculating LSE using NDVI^THM should be taken into account in future studies as review of the literature have shown that P_V could be estimated with different accuracy by means of vegetation indices [87] and machine learning approach [88] over different ecosystems.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

I appreciate your answers to all my questions and remarks. Considering your modifications and additions, the paper can be published in its present form.  Readers can benefit from your suggestion to combine thermal and optical Landsat data to obtain improved estimations of LAI of forests, and your comments clearly indicate the crucial point of the method, the need of precise direct or indirect measurements of proportion of vegetation cover.

Author Response

I appreciate your answers to all my questions and remarks. Considering your modifications and additions, the paper can be published in its present form.  Readers can benefit from your suggestion to combine thermal and optical Landsat data to obtain improved estimations of LAI of forests, and your comments clearly indicate the crucial point of the method, the need of precise direct or indirect measurements of proportion of vegetation cover.

 Answer: Thank you so much for your kind comment. We appreciate it for the comments provided.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors, 

My comments for your ms to be accepted having corrected these points: 

Line-44: Due to the dominant control of LAI over primary production (e.g., photosynthesis), transpiration, energy exchange, as well as other physiological characteristics pertinent to the wide range of ecosystem processes, the accurate prediction of LAI has been a concern for a broad spectrum of studies [4]. 

Comment: Though LAI maybe an important vegetation property, there are lots of knowledge gaps as to how LAI is affecting the basic processes such as Photosynthesis and Respiration at leaf/ canopy to ecosystem scales. Here I am referring the ongoing open scientific debate (from lab to Earth System Models) focusing on 'electron transfer and Rubisco' at the leaf level (V_cmax). Recent papers highlighted- vegetation photosynthesis/ respiration processes are mainly controlled by climatic drivers supplemented with soil nutrients in a fabourable conditions (enough chlorophyll, light available) based on optimal theory and least-costs cordination. 

Please provide more important information in your paper with recent references in support of your Line 44 statement. 

Cheers

Author Response

Reviewer 3

My comments for your ms to be accepted having corrected these points:

Line-44: Due to the dominant control of LAI over primary production (e.g., photosynthesis), transpiration, energy exchange, as well as other physiological characteristics pertinent to the wide range of ecosystem processes, the accurate prediction of LAI has been a concern for a broad spectrum of studies [4].

Comment: Though LAI maybe an important vegetation property, there are lots of knowledge gaps as to how LAI is affecting the basic processes such as Photosynthesis and Respiration at leaf/ canopy to ecosystem scales. Here I am referring the ongoing open scientific debate (from lab to Earth System Models) focusing on 'electron transfer and Rubisco' at the leaf level (Vcmax). Recent papers highlighted- vegetation photosynthesis/ respiration processes are mainly controlled by climatic drivers supplemented with soil nutrients in a favorable conditions (enough chlorophyll, light available) based on optimal theory and least-costs cordination.

Please provide more important information in your paper with recent references in support of your Line 44 statement.

Answer: Thank you for the comment. We acknowledge the respected reviewer comment. However, we believe that the variability in LAI depends on climatic and growing conditions and change in canopy structure (e.g., LAI) can control the relationship between GPP and sun-induced fluorescence. In addition, the LAI variations significantly change radiation regime within the canopy microclimate as well as sink/source distributions of CO₂ and H₂O. In this respect, we added more references to the manuscript to support our statement regarding Line 44.

Gower, S.T., C.J. Kucharik, and J.M. Norman, Direct and indirect estimation of leaf area index, f APAR, and net primary production of terrestrial ecosystems. Remote sensing of environment, 1999. 70(1): p. 29-51.

Simic, A., R. Fernandes, and S. Wang, Assessing the impact of leaf area index on evapotranspiration and groundwater recharge across a shallow water region for diverse land cover and soil properties. J. Water Resour. Hydraul. Eng, 2014. 3: p. 60-73.

Hesketh, J., Predicting canopy photosynthesis from gas exchange studies in controlled environments, in Predicting photosynthesis for ecosystem models. 2017, CRC Press. p. 37-50.

Zhang, Y., et al., Energy exchange and evapotranspiration over irrigated seed maize agroecosystems in a desert-oasis region, northwest China. Agricultural and forest meteorology, 2016. 223: p. 48-59.

Launiainen, S., et al., Do the energy fluxes and surface conductance of boreal coniferous forests in Europe scale with leaf area? Global change biology, 2016. 22(12): p. 4096-4113.

Gondim, P.S.d.S., et al., Environmental control on water vapour and energy exchanges over grasslands in semiarid region of Brazil. Revista Brasileira de Engenharia Agrícola e Ambiental, 2015. 19(1): p. 3-8.

Taugourdeau, S., et al., Leaf area index as an indicator of ecosystem services and management practices: an application for coffee agroforestry. Agriculture, ecosystems & environment, 2014. 192: p. 19-37.

Author Response File: Author Response.pdf

Reviewer 4 Report

Brief summary

In their study, the authors explored the use of thermal infrared data for LAI retrieval.

Several tests were carried out exploiting filed campaign data and Landsat 8 satellite data.   In particular, different estimation techniques have been tested, including artificial neural networks.

General comments

In my opinion, the work is interesting since it seeks new ways to estimate LAI from satellite data, also using thermal data. However, before recommending its publication, there is the need for some improvements. Please, find below some suggestions:

Abstract:

lines 30-33. Could you use only one specific set of prediction indicators for every showed results?

Introduction:

the introduction should put your work in a precise framework. Considering that there are many ways to estimate LAI from remote sensing data, the introduction section should at least include a brief state of the art in this regard. For example, you cloud find an exhaustive classification of LAI retrieval methods from satellite data in:

Verrelst, J., Rivera, J. P., Veroustraete, F., Muñoz-Marí, J., Clevers, J. G., Camps-Valls, G., & Moreno, J. (2015). Experimental Sentinel-2 LAI estimation using parametric, non-parametric and physical retrieval methods–A comparison. ISPRS Journal of Photogrammetry and Remote Sensing, 108, 260-272.

Novelli, A., Tarantino, E., Fratino, U., Iacobellis, V., Romano, G., & Gentile, F. (2016). A data fusion algorithm based on the Kalman filter to estimate leaf area index evolution in durum wheat by using field measurements and MODIS surface reflectance data. Remote Sensing Letters, 7(5), 476-484.

3. section 2.3:

please add further details for the satellite data (e.g. scene ID).

Have you tested the Landsat level 2 products? They include also surface reflectance data. Could you explain, in the manuscript, why you did not use level 2 data?

  4. section 2.5.2:

this section needs to be extended. In particular, details on the ANN typology and on the ANN  training phase are missing

Author Response

Reviewer 4

In my opinion, the work is interesting since it seeks new ways to estimate LAI from satellite data, also using thermal data. However, before recommending its publication, there is the need for some improvements. Please, find below some suggestions:

Abstract:

Lines 30-33. Could you use only one specific set of prediction indicators for every showed results?

Answer: Thank you for the comment. As requested the R²_CV and RMSE_CV were selected as a prediction indicator for the results.

Introduction:

The introduction should put your work in a precise framework. Considering that there are many ways to estimate LAI from remote sensing data, the introduction section should at least include a brief state of the art in this regard. For example, you cloud find an exhaustive classification of LAI retrieval methods from satellite data in:

Verrelst, J., Rivera, J. P., Veroustraete, F., Muñoz-Marí, J., Clevers, J. G., Camps-Valls, G., & Moreno, J. (2015). Experimental Sentinel-2 LAI estimation using parametric, non-parametric and physical retrieval methods–A comparison. ISPRS Journal of Photogrammetry and Remote Sensing, 108, 260-272.

Novelli, A., Tarantino, E., Fratino, U., Iacobellis, V., Romano, G., & Gentile, F. (2016). A data fusion algorithm based on the Kalman filter to estimate leaf area index evolution in durum wheat by using field measurements and MODIS surface reflectance data. Remote Sensing Letters, 7(5), 476-484.

Answer: Thank you for the comment. We have added an explanation into the ‘Introduction’ section as requested. However, it should be noted that the main aim of this study is showing the potential of integration of the TIR and VNIR/SWIR data to improve the prediction accuracy for the LAI.

Please see Page 3 Lines 69 to 74. 

In this respect, Verrelst et al. [17] were evaluated the all possible band combination for two and tree-bands indices as well as different machine learning approaches using Sentinel-2 data for LAI retrieval and revealed that machine learning approaches performed with greater accuracy. Also, Neinavaz et al. [18] demonstrated that multivariate methods (e.g., artificial neural network) are the most promising approach in comparison with the univariate approaches (e.g., vegetation indices) for prediction of the LAI using hyperspectral thermal infrared (TIR, 8–14 µm) data.

3. Section 2.3:

Please add further details for the satellite data (e.g. scene ID).

Answer: thank you for the comment. The image’s ID is added to the manuscript. LANDSAT_SCENE_ID = "LC81920262015221LGN01"

Have you tested the Landsat level 2 products? They include also surface reflectance data. Could you explain, in the manuscript, why you did not use level 2 data?

Answer: Thank you for the comment. Landsat Level-2 Surface Reflectance data have atmospheric corrections applied. We have preferred to do atmospheric correction manually by our self.  As the FLAASH properties consider water vapor, distribution of aerosols as well as scene visibility for the atmospheric correction.

Please see Page 6 and Lines 156-159.

For the OLI images, the conversions of radiance to reflectance and atmospheric correction have been done using the FLAASH module. As the FLAASH properties consider water vapor, distribution of aerosols as well as scene visibility for the atmospheric correction.

 4. Section 2.5.2:

This section needs to be extended. In particular, details on the ANN typology and on the ANN training phase are missing

Answer: Thank you for the comment. The text modified. Please see Page 9 Lines 214-224.

 The ANN consists of different layers including inputs, hidden layers as well as outputs. In this study, three scenarios were considered as input layers to the artificial neural networks (ANNs) for LAI estimation. These included the reflectance of bands 1 to 7 from the OLI sensor (VNIR/SWIR), the combination of reflectance and LSE as well as the combination of reflectance and LST (Table 3). For network training, the Levenberg-Marquardt algorithm was used as the common training algorithm in backpropagation networks to develop models for LAI prediction. (Atkinson, 1997 #313@@author-year) suggested that by raising the number of hidden layers, the network could tackle more complex dataset. However, still, there is no specific rule for defining the optimal number of hidden layer. Since the prediction accuracy of ANN is related to the number of neurons in the hidden layer, the ideal ANN size was determined by examining various numbers of the neurons. In this respect, the early stopping approach was used to avoid over-fitting. In this approach, training network pauses as soon as performance on the validation dataset begins to deteriorate (Nowlan, 1992 #314).

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Dear authors,

1) What is the variation of LAI measurements of different forest structures (conifer, broadleaf and mixed) from the 37 field plots. How do you relate field LAI variations against your estimations using the proposed method (reflectance+ thermal)? Should the LAI range predictions be similar for all forest types?

2) Other VIs such as standard NDVI, NDVI green and EVI derieved from OLI data should also be tested for this proposed method. 

3) Replace Figure 3 with a 1:1 line scatterplot for LAI validation.

4) Provide P-value for Figure 3.

Author Response

#Reviewer 1

What is the variation of LAI measurements of different forest structures (conifer, broadleaf and mixed) from the 37 field plots. How do you relate field LAI variations against your estimations using the proposed method (reflectance+ thermal)? Should the LAI range predictions be similar for all forest types?

Answer: Thank you for the questions. The variation of LAI regarding different forest structures are added to the manuscript. Please see Page 9 Lines 220 to 223.

Recently we have proved that by increasing the LAI value of various species the emissivity increase under the laboratory condition and LAI can be estimated with acceptable accuracy using thermal hyperspectral data. We would expect the same result by means of airborne or space born TIR hyperspectral data. Unfortunately, in this study, it is not possible to verify the effect of forest stand on LAI prediction accuracy as the number of available plots for broad-leaf (n=4), and mixed forest (n=7) are not sufficient for proper calibration and validation.

Other VIs such as standard NDVI, NDVI green, and EVI derived from OLI data should also be tested for this proposed method.

Answer: Thank you for the suggestion. The NDVI and EVI have been added to the analysis.

Replace Figure 3 with a 1:1 line scatterplot for LAI validation.

Answer: Thank you so much for the suggestion. It has been corrected

Provide P-value for Figure 3.

Answer: thank you for the comment. We estimated P-value for each scattered plots and mentioned them only here. As we believe that, the RMSE is much more important than P-value. P- Values for LAI measured versus LAI estimated are presented in the table below.

Author Response File: Author Response.pdf

Reviewer 2 Report

see attached file

Comments for author File: Comments.pdf

Author Response

#Reviewer 2

This paper is a contribution to the widely-studied subject of empirical estimation of LAI of forests from remotely-sensed data, based on some results obtained in the Ph. D. thesis of the first author (Neinavaz et al. 2017). The title suggests an improvement of estimation accuracy from the integration of thermal and optical data issued from Landsat images. However, the demonstration is not totally convincing. The best estimation is obtained from emissivity - LAI relationship, and emissivity parameter is inferred from NDVI and proportion of vegetation cover (not from thermal data). One may also question if an approach of LAI calculation from a robust machine learning algorithm like ANN or SVM ( Omer et al., 2016), using as inputs a combination of vegetation indices, could not have given better results than optical-thermal integration (taking also into account that the original spatial resolution of thermal data is 100m). For example, Bajwa et al. (2008) obtained very good accuracies when estimating LAI of crops with an ANN approach using only NDVI and SIPI as inputs (r2 = 0.91). Therefore, it may be of interest to add to the paper an attempt of estimating LAI from the ANN approach using appropriate vegetation indices as inputs.

Answer: Thank you for the comments. Here, emissivity is calculated using NDVI threshold method which is the most practical and well-known approach among thermal infrared remote sensing experts, which calculates land surface emissivity according to the relationship between NDVI and the emissivity of the soil and vegetation values derived from the thermal infrared region.  In addition, we have used emissivity to correct the LST value Eq.4 in the manuscript.

Regarding the given example by the reviewer (Bajwa et al., 2008), we doubt applying VIs as an input layer for ANN; as the chance of overfitting when only one matrix column for input layer is available would be significantly high.

Choice of vegetation indices (VIs): Only 4 vegetation indices have been considered, and curiously, the most commonly-used NDVI has been ommited. One may suggest the addition of other indices like SIPI (Bajwa et al., 2008)  and water sensitive indices like Reduced Simple Ratio RSR  (Sternberg et al., 2004), or ISR (Eigemeier et al., 2012).

Answer: Thank you so much for the suggestion. We have added NDVI, EVI, and RSR in the analysis as requested. However, since, in this manuscript, the commonly used vegetation indices for retrieving LAI have been considered, we did not apply SIPI, as SIPI is more reliable for quantifying Chlorophylls and Carotenoids contents (Blackburn 1998).

It should be noted that despite Stenberg et al. (2004) revealed that RSR responds dynamically to LAI particularly over homogeneous forest stand, Zhu et al. (2010) demonstrated that RSR is more sensitive to the influence of topography. Therefore, the topographic correction is necessary for the prediction of the LAI derived from RSR in mountainous areas as LAI is underestimated by RSR on slopes.

LAI estimations from VIs: Results indicate rather low accuracies, compared to results of many publications about remote sensing of LAI of forests. For example, Sternberg et al. (2004) used RSR for estimating LAI of Pine and Spruce stands in Finland and obtained good accuracy (r² = 0.75 and SEE = 0.33 for homogeneous stands). It is suggested to select only homogeneous stands and to give accuracies respectively for needle-leaf, broad-leaf, mixed and all forest stands.

Answer: Thank you for the comment and suggestion. We understand your point and are aware that LAI can be estimated with a varying degree of success over the different ecosystems. However, in this work, it is not possible to consider plots separately according to the forest stands, as the number of the sample (plots) for broad-leaf (n=4) and mixed forest (n=7) are not sufficient for model calibration/validation.

Emissivity: equations 1a, b, c: « f or NDVI » f is not defined and not needed, mention only NDVI You must indicate how is computed Pv (ANN trained with ground measurements)

Answer: Thank you so much for the comment. Regarding equations 1a, b, c it is for not f or. For is removed as respected reviewer requested. In this work, we used in situ P_V measurements collected from the fieldwork for 37 plots for calculating LSE. We did not estimate P_V using the ANN approach. Please see section 2.1.

Scatter plots: The scatter plots displayed in this paper are different from similar plots in Neinavaz (2017), although they are supposed to represent the same data set of 37 plots. Example: Figure 5.7 (Neinavaz et al. 2017) and Figure 2a (this paper)

Answer: Thank you for the comment. We note that the scatter plot, which the reviewer is referring to (Figure 5.7 (Neinavaz et al. 2017)) is different from the scatter plot presented in this manuscript. As, for calculating LSE for this work, we used in situ P_V measurements, while in Neinavaz 2017, LSE has been computed with the P_V being estimated by means of reflectance spectra as an input layer for the artificial neural network.

Blackburn, G.A. (1998). Quantifying chlorophylls and caroteniods at leaf and canopy scales: An evaluation of some hyperspectral approaches. Remote Sensing of Environment, 66, 273-285

Neinavaz, E., Skidmore, A.K., Darvishzadeh, R., & Groen, T.A. (2017). Retrieving vegetation canopy water content from hyperspectral thermal measurements. Agricultural and Forest Meteorology, 247, 365-375

Stenberg, P., Rautiainen, M., Manninen, T., Voipio, P., & Smolander, H. (2004). Reduced simple ratio better than NDVI for estimating LAI in Finnish pine and spruce stands

Zhu, G., Ju, W., Chen, J.M., Zhou, Y., Li, X., & Xu, X. (2010). Comparison of forest Leaf Area Index retrieval based on simple ratio and reduced simple ratio. In, Geoinformatics, 2010 18th International Conference on (pp. 1-4): IEEE

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors,

I enjoyed your paper and I feel potential of thermal data applications. Contents are clear, however, I recommend following revise on your paper.

1. Missing Figure

220 The relationship 220 between the field measured LAI and PV is presented in Figure 2.

              Fig. 2 doesn’t show the relationship between LAI and Pv. A figure should be added.

2. Estimation of Pv

Field Pv is used in this paper for calculation of LSE and LST. Therefore the LAI estimation models using optical thermal data are not complete remote sensing parameter model. The authors must remind to the readers, establishing accurate Pv estimation using satellite data is essential for practical application of thermal remote sensing data for LAI estimation. 

Author Response

#Reviewer 3

1. Missing Figure

The relationship 220 between the field measured LAI and PV is presented in Figure 2. Fig. 2 does not show the relationship between LAI and Pv. A figure should be added.

Answer: Thank you for the comment. The text is modified.

2. Estimation of Pv

Field Pv is used in this paper for calculation of LSE and LST. Therefore, the LAI estimation models using optical thermal data are not complete remote sensing parameter model. The authors must remind to the readers, establishing accurate Pv estimation using satellite data is essential for practical application of thermal remote sensing data for LAI estimation.

Answer: Thank you for the valuable comment. The request of a respected reviewer is considered and added into the discussion section, as follows;

‘It should be highlighted that in this study, the in situ PV measurements were used for computing LSE. Hence, accurate estimation of the PV by means of remote sensing data for calculating LSE is essential and should be taken into account in future studies., Please see Page14, Lines 291-293

Author Response File: Author Response.pdf