Forest Type Differentiation Using GLAD Phenology Metrics, Land Surface Parameters, and Machine Learning
Round 1
Reviewer 1 Report
The authors presented the Forest Type Differentiation using GLAD Phenology Metrics, Terrain Variables, and Machine Learning.
It is a challenge to classify different forest types using remote sensing data, and it affects by distribution features, study area, available field plots and methodology. The research results in the paper have illustrated the strengths and limitations of the proposed method used to classify different forest types using remote sensing data. The result of this study is reliable, the style, structure, and writing the paper writing is fine to me.
I have suggestion/comments as below:
- The cited references relevant to the research should be chosen to be more concise and succinct, considering omitting old references.
- In section 2.2, it would be nice to have more information on the phenology of each forest type /the growth characteristics of the species which will determine the EOS and SOS times of each season, and the correlation of the phenology of each forest type with GLAD phenology metrics parameters.
- Line 205: the start of the growing season (SOS)
- Line 206: the end of the growing season (EOS)
Author Response
Please see attached.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Summary and contribution
The authors use Landsat variables, including phenology variables, topographic variables, and an extensive existing ground observation network to predict and map 6-7 different forest types across the state of West Virginia. They achieved overall accuracies of 54-76%, and found that topographic variables improved model performance, and that phenology metrics were useful, although not as useful as harmonic regression coefficient variables. The authors also illustrated how including prediction probabilities might be warranted, as in reality forest type classes in their forest types are likely fuzzy.
In the Background section, the authors provide a nice summary of how their work fits within previous work classifying forest types. One of the main motivations to this work was to test how well Landsat-based globally-consistent GLAD Phenology Metrics could be used to classify forest type. Although the authors find that harmonic regression coefficients, which have been demonstrated previously to be useful predictors, outperformed GLAS Phenology Metrics, the authors’ investigation is well done and a worthwhile addition to forest type mapping with Landsat data. The paper seems like a fine fit for MDPI Geographies, and would also fit well in MDPI Remote Sensing.
Major comments
The manuscript is interesting, well written, well organized and therefore easy to read. The random forest modeling analysis is rigorous and well described. The figures and tables are well made and aid in understanding the authors’ analysis. I could find no major issues with the manuscript. I have only minor recommendations for clarifying methods and making the manuscript more understandable to readers.
Minor comments
Line 14: Consider changing ‘is’ to ‘was’ here to consistently use past tense throughout the abstract.
Lines 97-107: Same comment as above concerning verb tense. The authors switch between verb tenses here. For the most part this paragraph is in present tense, except for the sentence beginning on Line 102. I recommend writing the entire paragraph in either present or past tense.
Lines 259-279: I recommend elaborating more on the WVDNR field datasets here. The authors provide reference #73, which likely has all the necessary details, but it’d be helpful to readers to provide more information relevant to this study here. For instance, since this is a Landsat-based analysis, I wonder about the area of the ground observations relative to a 30-m Landsat pixel. Are they comparable? Also, what years did the plots span? I realize this might be more relevant to western North America, but was there a percent canopy cover threshold for excluding plots as non-forested?
Lines 292-298: The time period of the GLAD Phenology metrics is unclear to me. Either here, or in Section 2.2, I recommend stating the time period that the GLAD metrics span, i.e. are the GLAD Phenology metrics normals spanning 1984-present, annual, or user defined? If they are normals, how are disturbed areas where forest transitions to non forest, dealt with? Again, I realize this information is likely in reference #56, but it’d be helpful to readers to summarize here. Related to this, I don’t understand what is meant by the metrics being calculated relative to the year 2019, both how that is different than the other GLAD metrics and why it was done? I wonder if the poorer performance of the GLAD metrics relative to harmonic regression coefficients could be due to the time period of the harmonic regression coefficients not being an optimal fit for the data? Likely not, just speculation.
Figure 3: Perhaps this was described in the methods, but it looks like non forest, including water, is masked in white. How was non forest masked?
Line 579: I recommend redefining metric set acronyms in the Table caption, or below the table, so readers don’t have to scroll between Tables 10 and 4 to interpret Table 10.
Lines 652-662: Was it possible that field observation locations were disturbed between the time of field observation and imagery dates? This would reduce classification accuracy.
Lines 672-629: This is a good recommendation and I don’t recommend removing it from the text. However, like my comment about my confusion over the time period associated with the GLAD phenology metrics, I wonder about what time period would be most appropriate for generating ARD harmonic regression coefficients, and can see an argument for leaving definition of the time period to the user, given the likelihood of forest being disturbed between 1984 to present (at least in western North America). In many cases, a user might wish to define their years of interest. If I were to use harmonic regression coefficients, I imagine I might need the ability to define the time period myself, preferably with a routine within Google Earth Engine (GEE), similar to the functions available within the Landtrendr implementation in GEE (https://emapr.github.io/LT-GEE/) or the CCDC implementation in GEE (https://gee-ccdc-tools.readthedocs.io/en/latest/background.html).
Lines 668-674: One way to make generation of topographic variables easier/faster would be to do so within GEE. It looks like a 1/3 arc second DEM of the United States is available within GEE (https://developers.google.com/earth-engine/datasets/catalog/USGS_3DEP_10m). Then it’d just be a matter of coding terrain functions within GEE. Related to this, was any sort of tiling scheme used for processing of such a large study area, or was one grid produced for the entire state of West Virginia?
Author Response
Please see attached.
Author Response File: Author Response.pdf
