Response to reviewer 1:

Our author responses are in italics.

This is my first review of the manuscript Forecasting Agricultural Drought: The Australian Agriculture Drought Indicators by Andrew Schepen et al. The paper is well-written, clear, and highly relevant to the journal. The topic is timely, as forecasting agricultural droughts is a critical area of research with significant implications.

However, I have a few moderate concerns, outlined below:

1.    The study primarily focuses on the sensitivity of seasonal forecasting system performance in relation to crop production and farm profit, rather than assessing the system’s ability to predict crop growth. This is because the observations used are not real but rather derived from the same system, forced by ground data. While I do not see this as a weakness, I believe the authors should clarify this distinction when presenting the paper and discussing the results. Additionally, if this study is more of a sensitivity analysis of forecast errors, it would be beneficial to explore the relationship between meteorological forecast errors and crop yield errors. This could provide valuable insights into conditions where the system may struggle to predict crop yield accurately.

RESPONSE: The reviewer is correct that our forecasts are evaluated against pseudo-observations, a concept which we introduce in lines 66-69.  We completely agree about the importance of further comparing AADI predictions compare to real-world data. To this end, we have prepared another paper currently under review in NHESS, which compares AADI simulations with ground truth data. We have already referred to this study around lines 180 and 416. As noted in lines 415-418, the end users of the Australian Agricultural Drought Indicators forecasting system are presented with both forecast skill and indicator skill, as both dimensions are crucial for decision making.

2.    The dataset description is not sufficiently detailed. I suggest providing more information to improve clarity and transparency.

RESPONSE:  We appreciate that the descriptions of the datasets are quite brief. Therefore, we propose to provide more information about the input and output datasets in a new table (as suggested by another reviewer). For example, for wheat, we will identify that we are predicting potential yield in tonnes/hectare within the wheat zones.

3.    The AADI system should be described more thoroughly. For example, it is unclear whether irrigation is considered and how water limitations are accounted for. Additionally, what would be the impact of these factors? Given that some crops in Australia are irrigated, a discussion on this aspect would enhance the study’s practical relevance.

RESPONSE: We agree that additional detail about the AADI system can be included. Given the large scale of the simulations, simplified management rules are used in APSIM. Currently, AADI produces water-limited yield, which represents the yield that can be achieved using current best practices, technology and genetics for rainfed crops. We propose to include this additional detail.

4.    Has the system been tested without post-processing? If so, what is the impact? Including this analysis would provide valuable recommendations for the development of simpler systems in other regions.

RESPONSE: Omitting the climate forecast post-processing step is virtually certain to lead to poor performance due to climate model biases. We propose to include an additional comment in the existing discussion, to advise that simpler bias corrections may be applied, however, failure to consider formal calibration of ensemble forecast reliability and restoration of temporal and inter-variable correlation structures may lead to unpredictable results.

Overall, the study is strong, but addressing these points would improve its clarity and impact.

Response to reviewer 2:

Our author responses are in italics.

The manuscript titled "Forecasting Agricultural Drought: the Australian Agriculture Drought Indicators" presents a novel ensemble-based drought forecasting system that integrates post-processed seasonal climate predictions from the ACCESS-S2 model with biophysical and economic models. The authors address a topic of great significance in Australia, where forecasting drought is a primary concern for numerous stakeholders. The topic is presented in an excellent manner, and the manuscript includes a commendable literature review. The authors tackle a critical issue by evaluating forecast performance for four key drought-related indicators: wheat yield, sorghum yield, pasture growth, and farm profit over a historical hindcast period (1990 - 2018). The manuscript is methodologically robust, highly relevant, and well-written. However, some clarifications and refinements are needed. The following are suggestions for improvement:

1.    The Introduction would benefit from an expanded discussion of major historical drought events in Australia, including impacts on agriculture. This sets the context more clearly for the need and relevance of AADI.

RESPONSE: We agree that additional scene-setting regarding major historical drought events will help justify the need and relevance of AADI. At present, we discuss historical events in section 4.6. We propose to include some additional background on Australian droughts, including the Millenium and Tinderbox droughts, in the introduction.

2.    While the manuscript describes the datasets used, it would be helpful to include a consolidated table summarizing the key datasets, variables, spatial resolution, and time periods. This could be placed in the Methods section or as a supplement.

RESPONSE: We appreciate that the descriptions of the datasets are quite brief. Therefore, we propose to provide more information about the datasets in a new table. 

3.    Was there any explicit quality control or filtering applied to the input data (e.g., rainfall, temperature) before simulation? If so, briefly describe this process. Otherwise, consider referencing previous studies that confirms the reliability of the datasets used.

RESPONSE: Regarding the observed climate data, the SILO gridded dataset of rainfall and temperature is a high-quality, managed dataset used widely in studies throughout Australia. We propose to include some additional information about the quality control process. 

4.    Consider including a flowchart that visually summarizes the overall methodology (not only for the AADI system). A brief caption accompanying the figure would support reader comprehension.

RESPONSE: We appreciate that verification of AADI requires a complex workflow. We propose to include an additional process chart to illustrate the overall methodology, including cross-validation, model simulations and forecast verification.

5.    It would be better to clarify the spatial analysis methodology used when aggregating or averaging forecast data across Australia. Explain whether area-weighted averaging or other geostatistical techniques were applied to spatially aggregated time series.

RESPONSE: Averaging the forecasts across Australia for Figure 8 uses simple averaging. We propose to include this detail.   

6.    The study focuses on model-based skill evaluation using pseudo-observations. Please address the rationale for this choice in more detail. How do the authors justify relying on model-to model comparison without validating forecasts against real ground-truth data?

RESPONSE: The reviewer is correct that our forecasts are evaluated against pseudo-observations, a concept which we introduce in lines 66-69.  We completely agree about the importance of further comparing AADI predictions compare to real-world data. To this end, we have prepared another paper currently under review in NHESS, which compares AADI simulations with ground truth data. We have already referred to this study around lines 180 and 416. As noted in lines 416-418, the end users of AADI are presented with both forecast skill and indicator skill, as both dimensions are crucial for decision making. 

Furthermore, developing an Australia-wide map requires modelling to generate continuous data at 5km resolution. We note in lines 415-416 that for variables like pasture, real observed data isn’t available, and it is necessary to validate against proxy variables like stocking rates or fodder cost. 

7.    Explain the rationale for using ACCESS-S2 exclusively. If alternatives (e.g., ECMWF, SEAS5, NMME) were considered, briefly note why ACCESS-S2 was selected.

RESPONSE: One of the reasons for selecting ACCESS-S2 was the availability of real-time forecasts at the commencement of the project. Forecasts such as SEAS5 are now more easily available through the Copernicus Data Store, and it would be possible to include them in AADI. We propose to add a sentence explaining the choice of ACCESS-S2.

8.    In the Discussion section, please consider addressing the uncertainties associated with the input datasets and how these may affect the reliability of the forecast outputs.

RESPONSE: We believe that this is an important point and ties into the discussion about forecast skill in the model space compared to ground truth data. We have provided some discussion around input errors in the farm input data around line 400. However, we propose to expand the uncertainty discussion to touch on the other uncertainties, for example in weather data and soil types.

Response to Reviewer 3:

Our author responses are in italics.

The paper addresses an important and timely topic of drought forecasting for agricultural enterprises, going beyond rainfall indices. Its main objective is to assess the forecasting performance of multiple drought indictors within the AADI system. The results show that farm profit indicator has a high forecasting skill in comparison to rainfall indices. Assessment of historical events shows the usefulness of this indicator in predicting drought impacts. Overall, the manuscript is well written, but some clarifications and improvements will be useful for the readers.

General Comments

1.    There appears to be some inconsistency in abstract and introduction regarding what is being assessed. From the abstract, I thought the paper will use commodity prices (along with crop growth) to forecast occurrence of droughts. But then in lines 52-53, it appears the goal is to forecast farm enterprises financial performance i.e. impacts of droughts using farmpredict model. And then again in line 70, there is mention of using farm profit as drought indicator. It would be helpful if the introduction consistently explains whether the paper aims to forecast drought via farm profit or to evaluate farm profit for drought impact assessment. Additionally, the introduction will also benefit with an explicit statement of its main aims, such as: “The aim of this study is…”

RESPONSE: We agree that there are some inconsistencies in the wording which can be improved. AADI predicts agriculturally relevant drought indicators, of which farm profit is one indicator in addition to crop yield and pasture growth. Farm profit can be impacted by meteorological drought and/or prices. Either way, we would identify this as agricultural or economic drought. We propose to modify the abstract and introduction to be clearer in our definitions of drought, what AADI is predicting, and include an explicit statement of aims.

2.    It would help to specify the months, fiscal year boundaries, or assumptions that go into your farm profit calculations, so Section 4.2 can be understood without needing to reference other works.

RESPONSE: We agree that the results should be interpretable without heavily relying on other works. The farm profit model is described in more detail in a companion paper currently also under review in NHESS. Nevertheless, we propose to include some additional detail about the farmpredict assumptions in our methods, to aid in the interpretation of the results.

3.    Over the hindcast period, the farm profit indicator is shown to have a good predictive skill as presented in section 4.6. However, the performance is assessed for historically declared droughts. It doesn’t become immediately clear to me how such prediction will be used in real-time forecasting to attribute changes in farm profits to droughts. I assume from the information provided in section 5 about inclusion of factors like soil moisture (Lines 383-385) within the AADI system, however, such information is not presented in the methods section and should be included briefly despite the reference to Hughes et al. (2024a).

RESPONSE: This is an interesting point that appreciates that farm profit and meteorological or hydrological drought are related but not always coincidental. The AADI models, in particular the crop models and pasture models, which themselves feed into farmpredict, contain soil moisture elements. We propose to clarify these details in the methods. As for real-time prediction, the purpose is not to strictly attribute drought to farm profit, but to provide early warning of where agricultural enterprises may be impacted by the climate, antecedent conditions, economics or some combination of these.

4.    From the manuscript, it doesn’t become completely clear to me how the paper addresses the stated shortcoming of drought propagation (line 14-15) and drought evolution (Line 36-37) in the drought forecasts.

RESPONSE: The statements about drought propagation and evolution are more criticisms of standard practice to use lagged rainfall indicators as a measure of agricultural drought. AADI is a fully integrated system that takes weather forecasts and integrates them through a suite of agricultural and economic models, to predict future conditions that may include the onset or conclusion of drought, through an agricultural lens. For example, it is quite possible that high soil moisture stores sustain a high yielding crop, despite low rainfall, thereby considering elements of drought propagation.

5.    Drought can raise or lower commodity prices. Is this aspect captured in your profit-based indicator, or are prices assumed exogenous? Please clarify.

RESPONSE:  Yes. The farm profit modeling does attempt to account for drought impacts on key Australian commodity prices, including prices of grain outputs (wheat, barley, sorghum) and fodder inputs (hay and grain for livestock feed). This involves a statistical model calibrated to historical data (as documented in Appendix B: of Hughes et al. 2022 Defining drought from the perspective of Australian farmers).  The approach assumes that global commodity prices are fixed (i.e., independent of Australian weather conditions) but allows Australian fodder prices, and the grain price ‘basis’ (Australian grain prices relative to global prices) to increase in drought years.

6.    The Discussion should reference related work on drought forecasting indicators in agriculture to help place your results in wider context.

RESPONSE:  Our introduction sets the scene for our work in the context of agricultural drought indicators; however, we agree that it makes sense to review our results in light of the other studies. We propose to provide relevant comparisons in a revised Discussion.

Specific comments

1.    Lines 62-63: “…policy planners might be interested in outlook for winter and summer crops…”Do you mean total production and not only yields? Please clarify.

RESPONSE:  We will clarify. This refers to potential yield. 

2.    Line 72-73: Please give examples for threshold and categorical forecasts.

RESPONSE:  We agree to include.

3.    Line 95: It would be useful for the readers to start with a brief introduction to ACCESS-S2 climate model.

RESPONSE:  We will add a brief description..

4.    Line 138: “MOF” used without definition

RESPONSE:  Thank you, we will correct.

5.    Line 231-233: “….. indicating skill in some regions in the Austral Spring."It is not clear how you reached this result. Either mention in text that this is not shown or add reference figure.

RESPONSE:  As per line 229, the percentiles are calculated on spatial pooling, and therefore the conclusion about skill occurring in some regions follows naturally.

6.    Line 245: IDR instead of interdecile range

RESPONSE:  We will correct, thank you.