Management of U.S. Agricultural Lands Differentially Affects Avian Habitat Connectivity

Round 1

Reviewer 1 Report

File provided

Comments for author File: Comments.pdf

Author Response

The authors used new and leading edge methods to examine the influence of agriculture on habitat connectivity of birds through the application of citizen science data. Overall, I found the manuscript was well informed by the most recent work in the field and well prepared by the authors. I primarily provide major feedback in this review and have not focused less on specific word changes or grammar.

The method section of the paper is the bulk of the manuscript coming in at about 10 of 25 pages of the text. I was able to follow all the methods, and the authors described in detail the steps that were completed in the analysis, though I did refer back to the papers they referenced especially in regards to curating eBird data and creating the agricultural intensity score. What I think would be very helpful is a process figure that guides a reader through the data curation steps, habitat layer development, habitat suitability model development, and connectivity model development. Thus, you could illustrate how all of the data interact in your models. I found myself sketching out what you did at each step and then referring back to my sketch as I continued reading. I don't think a figure would allow you to reduce the text in the methods, but it would provide a guide for readers to work through the different steps of the analysis process.

RESPONSE: Thank you for this suggestion. We have now added a process figure as Figure 1, outlining each step of the analysis, including data preparation/processing and modeling.

A lot of my focus is on sage-grouse as it is a species on which I work, it is of high conservation importance as you stated, and there are many habitat suitability models available including the new sagebrush ecological integrity model (Doherty et al. 2022). However many of these comments could be more broadly applied to American black duck and bobolink.

I've collected eBird data and have been a birder for 25 years, but I never considered the data I collected to inform an explicit measure of habitat suitability. I spent some time exploring the idea around encounter rate as measured through eBird data and its application as habitat suitability and have a few thoughts to offer. Your paper is robust with citations, but on L305, you do not provide a citation for assuming encounter rate directly translates to habitat suitability. A paper you rely on (Robinson et al. 2017) provides the initial methods for using eBird data, but they refer to the y-axis in the partial dependence plots as ‘probability of detection’ and use those data to map a distribution for probability of detection (see below).

I think one of the challenges in using the encounter rate data as a measure of habitat suitability for sage-grouse is that you haven't accounted for possible changes in detection probability among habitat types. I think this is less of an issue for black duck and bobolink because you have the focal periods and I don't expect that detectability of a bobolink during the breeding season would be influenced by habitat type. But I would expect that sage-grouse during the incubation period are very difficult to detect but are more detectable during the post-breeding period when they seek out mesic meadows for forbs. 

RESPONSE: Thanks very much for pointing this out. We agree that the quantity being directly modeled by our random forest models (i.e., encounter probability) is an imperfect proxy for habitat suitability. However there is indeed precedent in the literature for thinking of encounter rate (including those derived from community science data like eBird) in the context of habitat suitability, particularly when a datasource can provide both detection and non-detection data, as is the case here. The assumption in this case is that, once factors affecting detection probability (e.g., survey effort) are accounted for, encounter rate will be higher where abundance/density is higher, and density will be higher where habitat is more suitable. We have added additional citations to the manuscript to justify this usage of encounter rate as a proxy for habitat suitability (e.g., on lines 289 and 326). However, we have also taken better care throughout the manuscript to note that encounter probability is indeed a proxy for (and not a direct estimate of) habitat suitability, and explicitly note (lines 291-297), as the reviewer suggests below, that encounter rate could vary by habitat type in ways that are not strictly tied to density, e.g., if the species is more cryptic in some habitats than others. 

We also note that the previous version of our manuscript included a typo, stating that “we considered detection probability to be a proxy for habitat suitability.” This is not the case. We considered the quantity modeled by the RF model (i.e., encounter rate) as a proxy for habitat suitability, and used covariates in the model to account for spatially explicit differences in detection probability. This misstatement may have caused additional confusion and has now been corrected (Lines 289-290).

Regarding sage-grouse in particular, it was beyond the scope of this analysis to dig into fine-scale differences in detection probability between habitat types. But we note that our results largely correspond to those of previous studies in emphasizing the importance of relatively natural landscapes (and in particular sagebrush/shrub cover) over cultivated agricultural lands or developed areas. We have also included a new analysis (described in greater detail below) comparing predicted sage-grouse encounter rate/habitat suitability between seasons, which suggested very limited spatial differences in relative habitat suitability between seasons based on modeled differences in detection probability.

And I think the likely difference in detection probability among habitat types (which is likely correlated with season) might be influencing the very narrow range of encounter rate values that you show in Figure 1. And I wonder if detectability by habitat type and its influence on encounter rate could be influencing your patterns in Figure 1. I appreciate you put a caveats the text about being cautious about interpreting the high L values from sage-grouse, but the plot that you show in Figure 1 is opposite of what we know about sage-grouse and agriculture, and I expected the pattern to more closely align with what you showed for black duck than what is shown.

RESPONSE: The previous Figure 1 used partial dependence (PD) plots to examine the effect of agricultural land management intensity (L) on encounter rate when all other covariates are held at their means. However, these plots can be quite difficult to interpret given that PD plots incorporate complex interactions with other environmental and detection-related covariates in the random forest model and may do a poor job of reflecting data availability across the range of a covariate. For this reason, we have now moved the PD plots to supplementary material and replaced them with what is in our view a much more interpretable/straightforward visualization of the effect of L on encounter rate for each species. This new figure (now Figure 2, given addition of new process figure as Fig 1) uses predicted values from the RF model at many thousands of random locations across each species range to estimate the median encounter rate value at several levels of agricultural L. This more intuitive look at the effect of L on encounter rate shows that sage-grouse median encounter rate is highest at low, but non-zero, L values (corresponding in our analysis to those values associated with rangeland habitats), but that median encounter rate decreases quickly with increasing agricultural management intensity. This pattern (which was always reflected in our habitat suitability model, though poorly represented by the prior PD plot) matches that described in the literature for sage-grouse regarding low suitability in heavily managed agricultural lands. We describe this new approach to understanding the effect of L on encounter rate in updated Methods (Lines 338-352) and Results (Lines 443-458) sections. 

You noted on L306-L310 that you addressed detection probability when building the habitat suitability surfaces by making predictions for a “standard” checklist and setting the Julian date and checklist start times to those with the highest detection probability for each species (sage-grouse: 18 March at 05:02, black duck: 7 December at 05:05). I looked over Strimas-Mackey et al. (2020) for more info on the standardizing method and they extracted wood thrush records in June (Section 2.2). They then made predictions for a standard eBird checklist that was in June, the focal period of their data (Section 4.6). For your analysis, your standardized date is outside of the focal window for American black duck and is restricted to the lekking period for sage-grouse. If you made habitat suitability predictions from a “standard” checklist on 18 March for sage-grouse, how does Figure 2 show habitat suitability for sage-grouse in all seasons? Would Figure 2 not show suitable lekking habitat based on the standardized date? Similar for black duck; Figure 2 references fall and spring migration, but your standardized date is in winter. Bobolink is appropriately matched. If I’m incorrect in my interpretation, additional text is warranted L306-L310.

RESPONSE: The reviewer’s point regarding the potential effects of seasonal differences in detection probability on RF model predictions is an important one and we have now added a new set of analyses for the two species (sage-grouse and black duck) whose study period spanned multiple seasons comparing the effect of season on encounter rate model results (Lines 331-352). For the sage-grouse, we predicted range-wide habitat suitability based on the date with the highest detection probability (18 March - lekking season) and a comparison date during the post-breeding season (15 July). For the black duck, we predicted range-wide habitat suitability on 7 December (the date with the highest detection probability, which we note is indeed within the date range of the fall migration period, though towards the end), as well as a comparison date during the spring migration period (19 March). We provide predicted habitat suitability maps for each of these seasonal model predictions in the supplementary material (Fig. S2) for visual comparison. We also quantitatively compared the differences between seasons in relative habitat suitability across each species’ range by extracting the encounter rate value predicted by each model at thousands of random locations across each species range and calculating the correlation between values predicted by the two seasonal models for a given species. As noted in the results (lines 458-462), relative habitat suitability was highly correlated between seasons for each species, suggesting that, while exact spatial patterns do vary to some degree between seasons (Fig. S2), our overall model results regarding relative suitability across the landscape are largely insensitive to changes in detection probability among seasons.

As someone who would be looking to apply these results to conservation or to develop impact assessments, I’m uncertain how the L values in Figure 1 are aligned with your four different agriculture categories. I read your referenced 2023 paper where you provide baseline values for the four agriculture types. If someone is looking to manage, conserve, or develop energy within the range of these species, how do they interpret Figure 1 in the context of an area in which they are looking to work?

RESPONSE: We anticipate that the new version of the figure referenced above (now Figure 2) will prove substantially more interpretable in the context of species management, showing that, for the sage-grouse, encounter rates decline quickly as management intensity (i.e., intensity of mowing, harvest, fertilizer use) increases. However, we also encourage the reader to consider Figure 5 (connectivity value as a function of agricultural and other land cover types) in the context of management, which provides a sense of the relative importance of each land cover/use category to species movement and which mirrors to a large extent the relationship between land cover type and habitat suitability.

I liked the connectivity model, but there are a few results that are counter to our understanding of species ecology I’d like you to consider. First, I think the movement distance you used for sage-grouse is too large. The 120 km movement distance is uncommon and is known from one population (Newton et al. 2017). The average movement distance from nest to winter site in WY for 607 individuals was 17.3 km and did not exceed 85 km (see Figure 2 below from Fedy et al. 2012). How would you expect your analysis to change if you used a smaller distance that is likely better representation of typical sage-grouse movement patterns? 

RESPONSE: Thanks for pointing this out. It is worth noting here that we’re not actually modeling ‘average’ or ‘typical’ movements by sage grouse but rather setting the upper limit of what the connectivity algorithm could possibly connect. For that reason, we have elected to parameterize the model using the maximum observed movement distance as our moving window radius in an effort to account for all possible movements, not just typical movement patterns. We note that previous work directly comparing models run with varying moving window radii (e.g., Suraci et al. 2023. Biol Cons. 109896) has shown that larger moving window sizes tend to perform well in highlighting important connectivity areas also identified with smaller window sizes, while also capturing the potential for rare long-distance movements. Thus we do not anticipate that our sage-grouse model results would change substantially if a smaller radius (e.g., 85 km) were used, but expect that the current model may better capture areas of potential connectivity between populations, which could be relevant to management under changing environmental conditions. We now make this point clear in the Discussion section (lines 587-592)

Figure 3 shows high current flow for sage-grouse through woodland, but studies have shown that juniper encroachment results avoidance by sage-grouse (Coates et al. 2017, https://doi.org/10.1016/j.rama.2016.09.001) and higher mortality risk (Prochazka et al. 2017, https://doi.org/10.1016/j.rama.2016.07.004). I’m uncertain how to interpret the current flow through woodland, which is almost as high as pasture, in the context of known sage-grouse response patterns.

RESPONSE: Thanks for pointing out this source of confusion. “Woodland” in this case is a very specific, ag-related cover type characterized as “natural or planted forest cover that is part of a functioning farm unit and no further than 160 m from cropland or pasture” (Lines 225-226). So this would tend not to encompass natural woodlands but rather remnant or planted patches of tree cover on functioning farms. The reason for relatively high predicted current flow through this cover type is likely that, relative to neighboring, more intensively managed agricultural areas (e.g., cultivated crops), these patches of more natural cover represent lower resistance to movement and are thus considered by the model as being locally important for connectivity. We now remind the reader of the agriculture-based definition of “woodland” in the results section to avoid further confusion (Lines 494, 518-520). 

Results for developed land for black duck and bobolink are presented in detail (L483-496). However, there is no discussion of developed land and what this means for black duck and bobolink movement. These species could be using developed areas but could this be one or two birds on a checklist vs larger groups in less developed areas?

RESPONSE: We have now added a paragraph on black duck and bobolink connectivity in developed areas to the discussion and acknowledge the difficulty of treating all checklists similarly despite differences in number of individuals detected (Lines 646-663). However, given the potential for errors in eBird count data (e.g., for most observers, particularly non-professionals, identifying that a species was present is much less error prone than reporting an exact number of individuals), we believe this is a reasonable trade-off and unlikely to affect our overall connectivity results given that these species do at least occasionally move through developed areas.

What could be driving the difference among models? L530-531.

RESPONSE: We have now substantially revised the Discussion paragraph (lines 559-599) describing the comparison between our model and previous sage-grouse connectivity models to highlight the important areas of similarity and dissimilarity among these analyses. We have also added a new set of maps to the supplementary material (Fig S3) to allow for direct visual comparison by the reader

You haven’t measured habitat quality in this paper L543-545.

RESPONSE: We have removed reference to habitat quality here.

Overall, I found the paper very interesting, but the long methods section would be enhanced by a figure to help a reader follow all of the steps in the paper. Most of my issues arise from the inclusion of sage-grouse, which as you note, has been studied extensively. Recently, Doherty et al. (2022), which you cite, developed an ecological integrity score for sagebrush habitat, which incorporates human modification into the integrity score. It seems that much work has been done on understanding habitat selection and movement of sage-grouse, and I had difficulty understanding how your sage-grouse analysis offered something new.

RESPONSE: As noted above, we have now added the suggested process figure (new figure 1) to provide an overview of the modeling workflow and have revised the discussion to better highlight the value of our sage-grouse analysis in the context of existing work.

One aspect I kept coming back to is using encounter rate as a measure of habitat suitability, as you do not define suitability. On L304-L306 you make this assumption, but do not cite any studies that demonstrate this relationship. A limitation of the analysis is that 100 bobolink or 1 bobolink are treated the same in the analysis – it is a checklist with a bobolink present. How would individual count affect the assumption regarding habitat suitability? One bobolink in a downtown park near Chicago is a much different observation than 100 bobolink in a grassland. Further, the core papers you cite (e.g., Robinson et al. 2017) use the term ‘encounter rate’, and I don’t think your paper would lose much if you moved away from habitat suitability to encounter rate.

RESPONSE: As described in detail above, we have made substantial revisions throughout the manuscript (including adding additional references) to justify our use of “habitat suitability” and put it in the context of previous studies using detection/non-detection data. We have also addressed the limitations and trade-offs with data quality of treating all detection checklists similarly regardless of number of individuals detected. Please see our prior responses for details.

Reviewer 2 Report

The manuscript highlights the important point that not all agricultural habitat is equal with respect to its suitability for different species, including non-cropland species, and that agricultural land can serve some needs of some species (e.g. food or nesting sites, corridors). I also like the explicit focus on management implications for conserving the three species.

The methods employed are well explained with a wealth of detail given. I liked the approach followed regarding data selection. The figures are very well made, and manage to make rather complex relationships accessible to the reader.

I only have very minor comments:

L507: maybe rephrase to make clear that this is only true for certain agricultural land types (especially) for sage grouse.

L528 ff: The differences between the high connectivity areas predicted by the author’s model and the two other studies should be discussed a bit more in the discussion section. Especially, I would like to see the high connectivity areas discussed, which their model did not find but which were reported in the other studies.

Author Response

The manuscript highlights the important point that not all agricultural habitat is equal with respect to its suitability for different species, including non-cropland species, and that agricultural land can serve some needs of some species (e.g. food or nesting sites, corridors). I also like the explicit focus on management implications for conserving the three species.

The methods employed are well explained with a wealth of detail given. I liked the approach followed regarding data selection. The figures are very well made, and manage to make rather complex relationships accessible to the reader.

 I only have very minor comments:

L507: maybe rephrase to make clear that this is only true for certain agricultural land types (especially) for sage grouse.

RESPONSE: Following the reviewer’s suggestion, this sentence has been rephrased to “Our results suggest that agricultural landscapes can in some cases provide valuable movement habitat for sage-grouse and Bobolinks, depending on the type and intensity of agriculture practiced, while providing limited value for migratory black ducks.”

 L528 ff: The differences between the high connectivity areas predicted by the author’s model and the two other studies should be discussed a bit more in the discussion section. Especially, I would like to see the high connectivity areas discussed, which their model did not find but which were reported in the other studies.

RESPONSE: We have now substantially revised the Discussion paragraph (lines 559-599) describing the comparison between our model and previous sage-grouse connectivity models to highlight the important areas of similarity and dissimilarity among these analyses. We have also added a new set of maps to the supplementary material (Fig S3) to allow for direct visual comparison by the reader

Round 2

Reviewer 1 Report

I appreciate the responses and revisions you have completed based on my review. I like Figure 1 as I can follow the progression of the data through the analysis.  In reading the papers you cited, your manuscript represents an advancement in the field of using eBird data to understand where species occur and how they move across the landscape. I like that you moved the partial dependence plots to the supplementary, and appreciate the effort to complete a new analysis to produce the revised Figure 2. I appreciate you creating two seasons for modeling detection probability for sage-grouse and black duck, and although the predictions were not sensitive to season, the value of demonstrating this method in your manuscript will likely have value for other researchers that use similar methods. I liked the discussion you added on the sage-grouse model comparison and that on boblink and black duck current flow through developed areas. To me, the value added here is very high. Finally, the information you added about detection probability and habitat suitability provided exactly the context I was looking for in the revision.

Overall, I really enjoyed reading your manuscript, and I’ve already talked with my colleagues about enhancing our capabilities of working with eBird data. And I have a new appreciation for the data I contribute to eBird!