Topographic Position Index Predicts Within-Field Yield Variation in a Dryland Cereal Production System
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors- Section 1.2.5: The author only describes the soil sampling methods but does not specify the sampling depth and the explicit criteria for selecting sampling points.
- The selection of topographic variables such as TPIand TWI lacks a detailed theoretical explanation.
- Lines 268-272mention that some fields were excluded from the data due to near-zero yields in certain years. However, these zero-yield data may contain valuable information, especially when analyzing yields under drought or abnormal climate conditions.
- The explanatory power of the model remains limited. The spatially-blocked cross-validation showed that the model only explains about a quarter of the yield variation. This suggests that the current model has not captured all influencing factors and may need further optimization.
- Although the study analyzes the factors affecting crop yield, it is unclear how the findings can be translated into specific agricultural management decisions. The paper lacks a discussion on how to formulate concrete management strategies based on these results.
- When discussing the influence of various variables (such as TPI, sand percentage, and soil carbon content) on crop yield, the paper mentions the importance and direction of these variables' effects, but lacks specific charts to visually present these relationships.
- The paper mentions significant spatial variability in crop yield but does not provide specific spatial distribution maps. Although section 3.1briefly discusses yield distribution across fields and crops, it lacks concrete spatial data support and fails to visually demonstrate the differences between fields or crops.
- When discussing the model's prediction performance, although the R²value is mentioned as a measure of model performance, there is no comparison chart between the actual and predicted values.
Author Response
Comment 1: Section 1.2.5: The author only describes the soil sampling methods but does not specify the sampling depth and the explicit criteria for selecting sampling points.
Response 1: Thanks for pointing that out. I've added more details on the soil sampling protocol in the revised manuscript (lines 237-242).
Comment 2: The selection of topographic variables such as TPIand TWI lacks a detailed theoretical explanation.
Response 2: Agreed, a better explanation of which variables were included and why is needed. I have added a new section in the methods describing the process and rationale we used to choose modeling variables (section 2.7). Relevant lines for TPI and TWI specifically are 272-277.
Comment 3: Lines 268-272mention that some fields were excluded from the data due to near-zero yields in certain years. However, these zero-yield data may contain valuable information, especially when analyzing yields under drought or abnormal climate conditions.
Response 3: Yeah, I know what you mean. We definitely debated the inclusion/exclusion of these fields. Ultimately, we chose to exclude them for a few reasons. Our main goal with these analyses was to fit a relationship between each of our spatially varying predictors and yield. But because there is no spatial variability of yield within these fields (every single yield monitor reading is zero), all these fields can really do in the model is flatten the predictor-yield curves, and make the relationships harder to distinguish from the background noise. This wouldn't be the case if the range of predictor values present in different fields were more distinct from each other. If that were the case the models might be able to pick up on some threshold of predictor value beyond which yield drops to zero in, for example, a drought year. But, for the most part, our field have overlapping ranges predictor values (see fig 4). And even for the predictors who's ranges do differ substantial between fields, we didn't see their between-field differences account for the between field differences in yield (i.e. crop failure fields vs non-crop failure fields) in any of the exploratory analyses that we performed before excluding the crop failure fields.
There must be some reason that certain fields failed while other didn't even in the same year, but I think we would need a data set with many failure field and many non-failure field to compare them to, to be able detect a cause. As it stands, in the case of corn for example, we have just two examples of crop failure fields, and no obvious reason why in terms of any of the predictors measured, so its hard to detect any kind of causative factor there.
One possible explanation may be spatial variability in rainfall. It's really surprising how much rain can vary, even over small distances (we discuss this in lines 613-622). It may have been that some areas of the research site received just enough rain to produce a little bit of harvestable grain, while in other areas the plants may have run out of water before being able to do so.
Comment 4: The explanatory power of the model remains limited. The spatially-blocked cross-validation showed that the model only explains about a quarter of the yield variation. This suggests that the current model has not captured all influencing factors and may need further optimization.
Response 4: The limited explanatory power of the models is an important point. Indeed, it suggests to me that there are significant influencing factors that we are unable to capture given our data set. We discuss several factors that likely contribute to the lack of predictive power of our models, and suggest strategies for future research to address these knowledge gaps in lines 612-667.
Regarding further optimization of the models, I don't believe we can significantly improve model performance given the dataset that we have. During the data exploration phase we performed several preliminary analyses which included alternative sets of variables, alternative modeling frame works (XGboost and an artificial nural network), and hyperparameter tuning of the random forest models, but none of these attempts made any substantial difference in terms of predictive performance on hold-out fields.
Comment 5: Although the study analyzes the factors affecting crop yield, it is unclear how the findings can be translated into specific agricultural management decisions. The paper lacks a discussion on how to formulate concrete management strategies based on these results.
Response 5: Some management implications are discussed in lines 526-533 and 578-582. For the most part however, I think specific recommendations are beyond the scope of this work. This study mainly seeks to establish putative drivers of within-field spatial variability in yield. Each of the drivers identified in this study would require it's own follow-up studies to quantify specific ranges of predictor variable values and weather conditions, under which management interventions would be appropriate. To that end, we also discuss strategies for future research, which would represent a step in the direction of developing management strategies, guided by our results in lines 642-654.
Comment 6: When discussing the influence of various variables (such as TPI, sand percentage, and soil carbon content) on crop yield, the paper mentions the importance and direction of these variables' effects, but lacks specific charts to visually present these relationships.
Response 6: Figure 7 shows the direction and magnitude of these relationships for the 4 predictors that tended to be both significant and have large effect sizes across the models.
Comment 7: The paper mentions significant spatial variability in crop yield but does not provide specific spatial distribution maps. Although section 3.1briefly discusses yield distribution across fields and crops, it lacks concrete spatial data support and fails to visually demonstrate the differences between fields or crops.
Response 7: Good point. With 18 field and four years of yield I think it would be a bit much to include all yield maps in the main text, but figure 8 includes yield maps from two example fields, and I have created a supplement with yield maps for all fields and all years.
Comment 8: When discussing the model's prediction performance, although the R²value is mentioned as a measure of model performance, there is no comparison chart between the actual and predicted values.
Response 8: Thanks for the suggestion. I've added a new figure (figure 5 in the revised manuscript) which shows the predicted vs actual yield values. This will provide an intuitive visual representation of the degree of uncertainty, as well as any biases, in the model predictions.
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper is very well written and adds nicely to a large body of work on this topic, not through any groundbreaking methods or findings, but through solid basic analysis of a high-quality dataset. It's good work!
I have a slight quibble with the stated scope of the paper. The abstract and intro state that the dataset was 18 fields over 4 years. I suppose technically this is true, but I find it misleading. The fields are all clustered together, and they seem to be "fields" only in the context of research - in a real-world implementation they are all very small and would not be considered by the average practicioner to be "field" scale plots at all. Arguably it's one field or possibly two functionally contiguous fields. Given the tenor of the introduction about the need for larger datasets I was then surprised to find this dataset much spatially smaller than I was expecting. Compare your 2.6-4.3 ha field size this to the average size of a single (irrigated) field in Colorado of ~50 ha and you see why the abstract and intro had me expecting (and excited for) something very different in terms of scale and thus also topographic variation and diversity.
That also affects some of your dialogue (74-83) around elevation. In your case, the scale is so small that (even though you explain why it's not a great metric) you can use it reasonably. In other studies which more diverse fields we often see relative elevation as a metric (usually relative to the field average). The paragraph at 74-83 seems to be suggesting the it's too local of a metric to use universally, but then you use it without any explanation of why it's appropriate in this case.
In section 2.6, did you consider including plan curvature rather than only profile curvature? Perhaps plan curvature is more influential in wetter regions where this is an indicator of converging vs diverging surface runoff and thus can change the soil moisture patterns within the field (or maybe the TWI captures this - although your later discussion at 423-436, that for flattish areas TWI is probably too exaggerated, suggests otherwise).
Relatedly, section 2.6 would be strengthened by explaining the implication or meaning (what it represents) of each of the metrics rather than only the mathematical definition. Without explaining that, it feels like these metrics were chosen for random reasons.
Minor remarks:
Figure 4 is beautiful, thank you.
Line 339: usually there are hyphens between the words out-of-bag
Author Response
Comment 1: I have a slight quibble with the stated scope of the paper. The abstract and intro state that the dataset was 18 fields over 4 years. I suppose technically this is true, but I find it misleading. The fields are all clustered together, and they seem to be "fields" only in the context of research - in a real-world implementation they are all very small and would not be considered by the average practicioner to be "field" scale plots at all. Arguably it's one field or possibly two functionally contiguous fields. Given the tenor of the introduction about the need for larger datasets I was then surprised to find this dataset much spatially smaller than I was expecting. Compare your 2.6-4.3 ha field size this to the average size of a single (irrigated) field in Colorado of ~50 ha and you see why the abstract and intro had me expecting (and excited for) something very different in terms of scale and thus also topographic variation and diversity.
Response 1: Fair point. I struggled with the terminology while writing the manuscript because the term "plot" doesn't typically encompass the size and heterogeneity of landscape features that our 'fields' do, yet on the other hand our 'fields' are indeed very small. I've compromised in the revised manuscript by using the term "management unit" instead of field.
Comment 2: That also affects some of your dialogue (74-83) around elevation. In your case, the scale is so small that (even though you explain why it's not a great metric) you can use it reasonably. In other studies which more diverse fields we often see relative elevation as a metric (usually relative to the field average). The paragraph at 74-83 seems to be suggesting the it's too local of a metric to use universally, but then you use it without any explanation of why it's appropriate in this case.
Response 2: Yeah, the choice to include elevation is a little thorny given the discussion about how it can be problematic in the intro. One of the reasons I kept elevation in the models is because I wanted to actually SHOW that there isn't necessarily a consistent relationship between elevation and yield. Whereas, if I had exclude elevation from the models I wouldn't have been able to say anything about elevation one way or the other. I hadn't realy made that point in the initial manuscript, but I've added a paragraph in the discussion of the revision (lines 523-532) discussing elevation.
Comment 3: In section 2.6, did you consider including plan curvature rather than only profile curvature? Perhaps plan curvature is more influential in wetter regions where this is an indicator of converging vs diverging surface runoff and thus can change the soil moisture patterns within the field (or maybe the TWI captures this - although your later discussion at 423-436, that for flattish areas TWI is probably too exaggerated, suggests otherwise).
Response 3: We did consider several other topo indices, including plan and total curvature, but I hadn't thought specifically about how plan curvature might indicate converging vs diverging runoff. I think between TPI, roughness, and TWI we've adequately captured areas where water tends to accumulate/converge/pool, but you raise a good point that plan curvature might be better in low flat areas than TWI. Given the short revision timeline of this journal it'd be difficult to add any predictors at this point, but it's a good point to keep in mind for future studies.
Comment 4: Relatedly, section 2.6 would be strengthened by explaining the implication or meaning (what it represents) of each of the metrics rather than only the mathematical definition. Without explaining that, it feels like these metrics were chosen for random reasons.
Response 4: Agreed. I've added a new section in the methods (section 2.7) describing the rationale we used to chose variables for inclusion in our final models. Topo variables specifically are discussed in lines 272-293.
Minor remarks:
Figure 4 is beautiful, thank you.
Thank You!
Comment 5: Line 339: usually there are hyphens between the words out-of-bag
Response 5: Hyphens added
Reviewer 3 Report
Comments and Suggestions for Authors- Justifications of using the inputs is unclear. I suggest using a systematic feature selection approach to better justify this.
- Error distribution analysis related to the model for better presenting the uncertainty will be important.
- Scatterplot relationships between the independent and target variable should be presented and discussed.
- Similarly to Figure7, a map visulization is need for the inputs depicting their variations.
- Discussions of the effect size is good but insufficient. Authors have to extend it.
Author Response
Comment 1: Justifications of using the inputs is unclear. I suggest using a systematic feature selection approach to better justify this.
Response 1: Good point, the rationale we used to choose predictors needs better explanation. I've added a new section in the methods (section 2.7) describing how we arrived at our final set of modeling variables (lines 254-328).
While we didn't use a systematic feature selection process, we did consider many potential variables. Ultimately, we were more interested in exploring the mechanistic drivers of crop yield variability than maximizing percent variance explained, so we didn't feel it was critical to select the very best among groups of highly correlated potential predictors. Rather, we sought to include variables which represented aspects of the landscape that we hypothesized to be important, while maintaining a decent representation of the entire potential predictor space. In other words, in cases where we had groups of several highly correlated potential predictors, we felt that as long as we made sure to include one of them, we wouldn't lose too much information; And by choosing one with a clear mechanistic link to yield, we would make interpretation of the results easier, as well as avoiding 'fishing' for relationships.
Comment 2: Error distribution analysis related to the model for better presenting the uncertainty will be important.
Response 2: Thanks for the suggestion. I've added a new figure (figure 5 in the revised manuscript) which shows the predicted vs actual yield values. This will provide an intuitive visual representation of the degree of uncertainty, as well as any biases, in the model predictions.
Comment 3: Scatterplot relationships between the independent and target variable should be presented and discussed.
Response 3: Visualizing the raw relationships between the predictors and yield could be helpful to get a sense of the direction of relationship, and the level of noise in the system. With 15 predictors and 3 crop types, it wouldn't make sense to include all of them in the main text, so I have created a supplement showing all of these relationships (Supplement 4).
Comment 4: Similarly to Figure7, a map visulization is need for the inputs depicting their variations.
Response 4: Yes, including maps will certainly help the reader visualize the spatial patterns of predictors, as well as assess the similarities and differences among them. Again, with so many fields and predictors I don't think it would make sense to include all these plots in the main text, so I have created another supplement to show maps of the predictors, as well as maps of yield (Supplement 3).
Comment 5: Discussions of the effect size is good but insufficient. Authors have to extend it.
Response 5: Good point. I have expanded on the derivation of the effect size calculation in the methods section (lines 360-371), as well as incorporating discussion of effect sizes into the relevant paragraphs of the discussion section (lines 466-470, 473-475, and 565-567)