Mapping Invasive Herbaceous Plant Species with Sentinel-2 Satellite Imagery: Echium plantagineum in a Mediterranean Shrubland as a Case Study

Round 1

Reviewer 1 Report

Dear authors, 

I got some questions about your manuscript, that I will share in the following:

- I saw the area delimitation, but I can not find what is the total study area? Also, I would like to know what is the training dataset's area?

- You mention that EP could coexists with shrublands/agriculture, do you consider this could affect your results significantly?

What would be the ideal patch size to improve the EP's detection?

How could you to reduce the comission error? (35.29%) related confusion of EP with agriculture/shrubland.

Why would be the reason you got similar results using Sentinel-2, compared with another study using hyperspectral imagery EO-1 Hyperion?

Do you think that using another species with high sinergy with EP, would help to improve the classification results?

I found your research very interesting, because it targets how to detect small plants (IAP) through remote sensing, which could make a big impact on detection, management and mitigation of their impacts on important niches or the environment in general.

Best regards, 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

 

Plant invasions are one of the main drivers of ecosystem degradation, and Mediterranean ecosystems such as the South African fynbos are severely affected. Remote sensing tools, especially freely available satellite data such as Sentinel 2 imagery as well as processing platforms such as GEE, offer new possibilities for monitoring and management.

Remote sensing of herbaceous invasive species in Mediterranean ecosystems using freely available satellite is understudied. In this study, a workflow is presented to map the invasive Echium plantagineum in a Mediterranean shrubland in the South African Fynbos. Using the Random Forest algorithm, E. plantagineum can be mapped with good accuracy. The limitations of this approach concerning the pixel size of the satellite data and the patch size of the invader are discussed. This is of high importance, but not addressed in the majority of such studies. Overall, the workflow is clearly understandable and the manuscript is well written.

 

Here are my detailed comments:

L36: The statement is about the natural environment and agriculture, but the citation refers to agriculture. Rephrase or add another reference.

L113-121: This detailed description of model accuracies of different studies reads a bit like a section in the discussion. Either move, or summarize with focus on research question and/or gap.

L128: Some of the studies mentioned in the citation are about mapping Echium plantagineum. The reader does not necessarily know this, but it could be pointed out.

L135-139: Maybe examples of remote sensing studies of plant invasions in diverse Mediterranean shrublands that are relevant for your study could be added here. For example, invasive acacia was mapped in diverse Mediterranean dune shrublands and linked to anthropogenic disturbance and water availability (Große-Stoltenberg et al. 2018). Further, Andrew & Ustin (2008) address the challenge of the environmental context in mapping an herbaceous invasive species using hyperspectral data. de Sa et al. (2017) use freely available Landsat time series data to map invasive acacias in Mediterranean shrublands. Kattenborn et al. 2019 used combined Sentinel 1 and 2 data to map invasive woody species in Mediterranean Chile. Some of those topics (challenge of transferability, the environmental context, time series data, sensor fusion) could be picked up in the discussion section again as an outlook for future studies. Consider adding more references in this section and discuss the findings in more depth in the discussion, as mapping invasive plant species can be challenging in diverse ecosystems such as Mediterranean shrublands.

 

L179-190: So if the flowering period covers several months and if the study site is quite heterogeneous (e.g. regarding disturbance and water availability) which might affect the phenological status of Echium patches, would it make sense to use more than one satellite scene especially when they are freely available? See e.g. the study by Pastick et al. (2020) about assessing grass invasion in drylands using Landsat and Sentinel 2 time series. Please discuss in the Discussion section.

L257ff: Would it be possible to assess the probability for each classified pixel using Random Forest as a measure of model uncertainty or to improve accuracy (see e.g. Belgiu & Drăguţ 2016. Sage et al. 2020)?

 

L257ff: The importance of parameter tuning is emphasized. For which metric were the parameters optimized? Overall accuracy, out-of-bag error, PA/UA of Echium, or another accuracy metric? See e.g. Große-Stoltenberg et al. (2018), Belgiu & Drăguţ (2016).

 

L284: Did you calculate the correlation of the predictors, and did you remove highly correlated predictors applying a certain threshold (see e.g. Dormann et al. 2013)?

 

L318: Introduce MDG in the Methods section.

L345 (Figure 2):

-        -Is “Overall accuracy” displayed on the Y-axis?

-        -Instead of “Analysis A, B, C” you could also use Bands + VIs”, “VIs” and “Bands”. This would be shorter and easier to understand.

-        -I find it hard to distinguish the points in Fig.2. Maybe there is another way of colorization and visualization (smaller dots, other colours, other symbols)?

L350 (Figure 3): Could you show the variable importance for all models? And would it make sense to order the variables by importance?

L379: This statement would be clearer if Fig.3 would be ordered by importance score, and if Fig. 3 included some information about the covered wavelength regions (e.g. VIS, RE, NIR, SWIR) for each predictor.

L379ff: Important variables were identified based on the Random Forest variable importance, and recommendations were derived from this analysis. However, variable importance might vary between classifiers, there are further approaches for variable selection, and UA/PA accuracies might differ, too (see e.g. Große-Stoltenberg et al. 2016 as example for a study in a Mediterranean ecosystem). Could you elaborate on this in the Discussion?

L410ff: The challenge of small invader patches compared to the pixel size is addressed, which is highly important for monitoring and management, but rarely addressed or quantified in remote sensing studies. One potential solution in this context could be downscaling spatial resolution of satellite data using deep learning (see Chen et al. 2020). Could you elaborate on this in the discussion section as it might be a relevant future research direction regarding the application of remote sensing in plant invasion management?

References:

Andrew, M. E., & Ustin, S. L. (2008). The role of environmental context in mapping invasive plants with hyperspectral image data. Remote Sensing of Environment, 112(12), 4301-4317.

Belgiu, M., & Drăguţ, L. (2016). Random forest in remote sensing: A review of applications and future directions. ISPRS journal of photogrammetry and remote sensing, 114, 24-31.

Chen, M., Ke, Y., Bai, J., Li, P., Lyu, M., Gong, Z., & Zhou, D. (2020). Monitoring early stage invasion of exotic Spartina alterniflora using deep-learning super-resolution techniques based on multisource high-resolution satellite imagery: A case study in the Yellow River Delta, China. International Journal of Applied Earth Observation and Geoinformation, 92, 102180. de Sa, N. C.,

Carvalho, S., Castro, P., Marchante, E., & Marchante, H. (2017). Using landsat time series to understand how management and disturbances influence the expansion of an invasive tree. IEEE Journal of selected topics in applied earth observations and remote sensing, 10(7), 3243-3253.

Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., ... & Lautenbach, S. (2013). Collinearity: a review of methods to deal with it and a simulation study evaluating their performance. Ecography, 36(1), 27-46.

Große-Stoltenberg, A., Hellmann, C., Werner, C., Oldeland, J., & Thiele, J. (2016). Evaluation of continuous VNIR-SWIR spectra versus narrowband hyperspectral indices to discriminate the invasive Acacia longifolia within a Mediterranean dune ecosystem. Remote Sensing, 8, 334.

Große-Stoltenberg, A., Hellmann, C., Thiele, J., Werner, C., & Oldeland, J. (2018). Early detection of GPP-related regime shifts after plant invasion by integrating imaging spectroscopy with airborne LiDAR. Remote Sensing of Environment, 209, 780-792.

Kattenborn, T., Lopatin, J., Förster, M., Braun, A. C., & Fassnacht, F. E. (2019). UAV data as alternative to field sampling to map woody invasive species based on combined Sentinel-1 and Sentinel-2 data. Remote sensing of environment, 227, 61-73.

Pastick, N. J., Dahal, D., Wylie, B. K., Parajuli, S., Boyte, S. P., & Wu, Z. (2020). Characterizing land surface phenology and exotic annual grasses in dryland ecosystems using Landsat and Sentinel-2 data in harmony. Remote Sensing, 12(4), 725.

Sage, A. J., Genschel, U., & Nettleton, D. (2020). Tree aggregation for random forest class probability estimation. Statistical Analysis and Data Mining: The ASA Data Science Journal, 13(2), 134-150.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The article is quite well written and smooth to read.

However, the number of samples for training and testing are really too small, in computer vision you need to add at least +10k samples, with various temporal and spatial resolution. The experiment SHOULD have been splitted into training, validation and test dataset. While the validation dataset can be used to finetune the model (hyper parameter selection)

The choice and use of vegetation indices is questionable when more accurate methods can learn specific indices for detecting specific specices (deep learning like DeepIndices, Genetic Algorithm, Symbolic Regression and other regression advanced machine learning) see wikipediahttps://en.wikipedia.org/wiki/Vegetation_index  . As we can see in the results, vegetation indices are less important than pure spectral data, meaning that the RandomForest (or any other machine learning) can get more realiable features from the input data than from empirical vegetation indices.

I suggest the author to increase the size of the dataset, and investigate on other methods, such as genetic algorithm and deep learning to realy finetune the model. The best thing would be to see if the model/algorithm is stable over time (summer / winter / ... years).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf