Review comments: Major Revision Recommended

The development of the History of Anthropogenic Phosphorus Inputs (HaPi) dataset presented in this manuscript is a significant undertaking with considerable potential value to the global change research community. The goal of creating a long-term, high-resolution global dataset of P inputs is laudable and addresses a genuine need.

However, from the perspective of remote sensing ecology and spatial data science, several fundamental issues regarding the choice of foundational spatial data, the realism of key temporal assumptions, and the strategy for spatial validation must be addressed before the dataset can be considered reliable for regional-scale applications. These concerns form the basis for my recommendation of major revision.

1. Core Issue: Foundational Flaws in Land Use/Land Cover Data

This is the most critical weakness, as it undermines the entire spatial allocation algorithm.

Problem: The manuscript relies exclusively on the model-simulated and historically reconstructed HYDE 3.2 and LUHv2 datasets for land use, without any cross-validation or correction against satellite-observed land cover data.

Technical Rationale:

Known Biases vs. Remote Sensing Observations: HYDE and LUHv2, particularly for historical periods, are known to contain regional inaccuracies. The authors mention this lightly in Section 4.3 as a "limitation," using the overestimation of cropland in India as an example. This is not a minor limitation but a core source of systemic error. In regions like Africa and South America, historical reconstructions of cropland and pasture are highly uncertain. Distributing P inputs based on biased land area data can lead to severe distortions in the spatial patterns and trends of input fluxes.

Missed Opportunity to Leverage Remote Sensing Products: For the post-1980 period, multiple global Land Use/Land Cover Change (LUCC) products derived from satellite observations exist (e.g., ESA CCI-LC, MODIS Land Cover). The manuscript makes no attempt to use these observational data to constrain or improve the spatial allocation for the last 40 years, which reduces the dataset's credibility for the modern era.

Actionable Suggestions:

1.It is strongly recommended to perform a consistency assessment and, if possible, a correction of the HYDE/LUHv2 cropland and pasture distributions for the period after 1980 using at least one remote sensing-based LUCC product (e.g., ESA CCI-LC).

The discussion must be substantially expanded to analyze how potential land use biases might affect the main conclusions, such as the shifts in regional hotspots.

1. Critical Assumption: The Unrealistic Temporal Invariance of Fertilizer Application Rates

Problem: A fixed, crop-specific fertilizer application rate circa 2000 (F_{crop,i,2000}) is applied to the entire period from 1860 to 2020.

Technical Rationale:

Contradicts Agricultural Practices: Agricultural technology, economic conditions, and environmental policies have changed dramatically over 160 years. The fertilization levels of the year 2000 are entirely unrepresentative of early Green Revolution or 19th-century practices. This assumption likely severely underestimates the phosphorus use efficiency of early agriculture while overestimating early application intensities, thereby distorting the historical trajectory of P inputs.

Undermines the Dynamic Crop Rotation Innovation: While the paper rightly highlights "dynamic crop rotation" as an advance, using a static application rate significantly diminishes the impact of this "dynamic" element. The primary driver becomes changes in crop area, not the evolution of management intensity.

Actionable Suggestions:

A sensitivity analysis of this critical assumption is mandatory. A simple scenario (e.g., assuming a linear increase in application rates from a baseline in 1860 to the value in 2000) should be designed to quantify its impact on global and regional P input trends.

In the discussion, this point must be elevated from a "limitation" to a "major source of uncertainty," explicitly acknowledging how it may lead to over- or underestimations in early historical inputs.

1. Insufficient Validation: Lack of Spatially-Explicit Validation Against Remote Sensing or Independent Data

Problem: The validation is confined to comparing global totals with previous studies (Table 3), which only verifies that the "magnitude is reasonable." For a high-resolution spatial dataset, the accuracy of its spatial patterns is paramount, and this is entirely unaddressed.

Technical Rationale:

Underutilization of Remote Sensing Proxies: While soil P cannot be directly sensed, several remote sensing proxy variables correlated with P inputs can be used for indirect validation.

Agricultural Intensity: Night-time light data or time series of vegetation indices (e.g., NDVI from MODIS/AVHRR) can identify agricultural hotspots and should be correlated with the spatial patterns of high P input.

Environmental Impact: P inputs drive eutrophication. The estimated P inputs for major watersheds (e.g., Mississippi, Yangtze) should be compared against time series of satellite-derived chlorophyll-a concentration or water turbidity in downstream lakes and coastal waters.

Lack of Sub-national Case Studies: The claim of revealing "subnational heterogeneity" is unsupported. There are no case studies at provincial, state, or watershed scales comparing HaPi against finer-resolution survey data or regional models.

Actionable Suggestions:

A spatial validation analysis must be added. As a minimum, the spatial distribution of HaPi should be overlain and correlated with the remote sensing proxy variables mentioned above for 2-3 representative regions (e.g., US Corn Belt, Western Europe, Eastern China).

It is recommended to seek existing regional or national agricultural P budget studies and aggregate HaPi data to these regions for comparison, assessing its ability to capture inter-regional differences.

1. Overly Qualitative Uncertainty Analysis

Problem: The uncertainty analysis in Section 4.4 identifies sources but concludes with only a generic "minimum of 20%" estimate. This is insufficient for a dataset intended to force numerical models.

Technical Rationale: In remote sensing and geoscience data processing, the goal is to provide spatially explicit, quantitative uncertainty information. For instance, a Monte Carlo approach or error propagation model could be used to generate an uncertainty map accompanying the P input data, synthesizing uncertainties from land use, application rates, etc.

Actionable Suggestions:

The authors are strongly encouraged to attempt a more quantitative uncertainty analysis. If a fully spatialized product is infeasible, a differentiated discussion of uncertainty levels for different periods and regions (e.g., developed vs. developing countries) should be provided.

Ideally, the published dataset should include per-grid-cell uncertainty estimates, which are critical for downstream users employing this data in models.

Conclusion

In summary, while the concept and global totals of the HaPi dataset are impressive, the issues with its spatial data foundation, temporal dynamics representation, and spatial validation significantly constrain its reliability for applications in regional ecology, environmental management, and models coupled with remote sensing. These are not minor textual issues but fundamental methodological concerns affecting the core quality of the dataset.

Therefore, I recommend Major Revision. The authors must undertake substantial revisions to:

Systematically evaluate the land use data against remote sensing products.

1.Quantify the impact of the static fertilizer rate assumption via sensitivity analysis.

2.Provide robust, spatially-explicit validation using remote sensing proxies and regional case studies.

3.Deliver a more quantitative and spatially-informed uncertainty assessment.

4.Only after addressing these core issues can the HaPi dataset become a truly reliable and powerful tool for the remote sensing ecology community and related fields.

This manuscript is built-upon the anthropogenic nitrogen input dataset (HaNi product;Tian et al., 2022), and reconstructs a comprehensive dataset with high spatial resolution (5 arc-minutes) of anthropogenic phosphorus input (mineral fertilizer, manure, and atmospheric deposition) history from 1860 to 2019 over the global terrestrial biosphere. Given the importance of phosphorus as a limiting element and the urgency of the current global phosphorus planetary boundary, long-term and spatially explicit anthropogenic P input datasets are valuable. The manuscript is well constructed and written. However, there are several major issues on methodology should be addressed.

L145: Annual national P fertilizer inputs from FAOSTAT, constant country/regional specific fertilizer partioning rate to cropland and pasture, and crop-specific fertilization rates by 2000 from Mueller et al., 2012 were used to derive gridded cropland and pasture fertilizer application. The authors used crop-specific fertilization rates from 2000 provided by Mueller et al. (2012) as a base dataset and back-calculates fertilization amounts from 1961 to 2020 by combining FAOSTAT national total P fertilizer amount. This method relies heavily on the accuracy of the 2000 data. It is not clear (1) how Mueller et al. (2012) dataset is different from Nguyen et al., 2024; (2) why not use Nguyen et al., 2024; (3) how they are different from crop-specific fertilizer use for several periods provided by FUBC1-9 (Ludemann et al., 2024). It should be noted that P fertilizer application rate for a crop can change a lot across the time. Assuming a constant rate is a strong assumption that needs further justification and discussion.

Furthermore, as described in Zou et al. (2022) cited by the authors, national-scale P fertilizer application intensity follows the environmental Kuznets curve, so that assuming P fertilization per unit area remains constant over time may beunreasonable. In addition, the constant partitioning rate between crop fertilizer use and pasture use (country or region specific value) is a coarse estimate, and could be wrong given the changes in the importance of crops/pasture in a country. For example, FUBC1-9 (Ludemann et al., 2024) provide fertilizer used for pasture/grasses for a few countries.

L169-L175: Yu et al. (2020) only provided crop distribution around 2010. Using machine learning to predict crop distribution in other years, considering that topography and soil variables will not change significantly in the short term, the changes in crop planting structure are almost only affected by changes in climate variables. The authors should further justify the method, and if there is no alternative/better choinces, please discuss the limitation and uncertainties specifically and probably ways to improve the estimates.

L218-L220: The authors used a fixed manure N/P ratio and manure N amount to estimate manure P between 1860 and 2019. However, with changes in animal feed composition, the manure n N/P ratio has varied over the past few decades for different livestock. The authors need to provide further discussion on the limitation and probably ways to improve the estimates.

L220-L223: The latest GLW dataset provide gridded data for several period based on consistent methods. This study relies only on GLW3, which neglected such variation in the past decades. I suggest the authors to revise the dataset with the latest GLW4.

L236-L238: This study attributes atmospheric P deposition to both natural and anthropogenic sources, yet the title is anthropogenic Phosphorus inputs. Is it possible to derive anthropogenic P deposition (e.g., excluding pre-industrial ones?)

L273-L274: How P deposition from 1860 to 1899 was extrapolated?

L333: I recommended to change the abbreviation of Southwestern South America, as SSA is easily confused with Sub-Sahara Africa.

Table 3: Zou et al. (2022) and Ludemann et al. (2024) also provide relevant results for P fertilizer and manure application, which need to be compared with the estimates from current study. In particular, Ludemann et al. provided the FAOSTAT nutrient budget reference database, and much of the data in this paper comes from FAOSTAT.

This manuscript presents a rich dataset supported by substantial and careful work. The resulting anthropogenic phosphorus input dataset will help advance our understanding of the global phosphorus cycle. However, I still have several concerns regarding the data-processing methods and some important details of the dataset development. I therefore recommend major revision and hope that the authors can address these concerns in a revised manuscript.

General comments

1. Validation is still weak (Lines 443-456). The comparison mainly focuses on global totals, and most reference datasets use the same FAOSTAT sources. This does not validate the 5-arcmin spatial patterns. I would like to see comparisons with independent national or subnational statistics for several representative regions.

1. The estimation of manure-derived P may involve considerable uncertainty. Lines 210-230: Could the authors clarify whether the manure P estimates are sensitive to the fixed animal-specific P:N ratios and the use of GLW3 livestock spatial patterns from a single reference year? Although annual FAO livestock statistics were used to scale national livestock numbers, the within-country spatial distribution appears to remain static, which may introduce uncertainty in regions where pig and poultry production has shifted substantially over time.

1. Lines 188-199: Uncertainty in the pre-1961 reconstruction. As I understand it, the 1961 spatial distribution was used as a reference pattern and scaled backward using the global temporal change rates from Cordell et al. (2009). If so, the pre-1961 maps may mainly reflect changes in global total P fertilizer use, rather than changes in regional or national spatial patterns.

I do not fully understand why this approach is appropriate for extending the dataset back to 1860. In my view, applying the 1961 spatial pattern to nearly 100 years of earlier history may seriously weaken the reliability of the pre-1961 data, because crop composition, cropland distribution, fertilizer accessibility, and agricultural practices likely changed substantially during this period. I strongly suggest that the authors more carefully address the uncertainty introduced by this backward extrapolation and provide clearer information in the dataset, such as reliability classes, or confidence flags for the pre-1961 period.

Minor comments

Lines 34-55 and 233-246: HaPi is described as an “anthropogenic” P dataset, but atmospheric deposition includes dust, sea salt, volcanoes, and other natural sources. The terminology should be reconsidered.

Lines 145-154: The manuscript states that 18 crops were included, but only 17 are listed. Please check the crop list.

Lines 182-185 and Table 1: The land-use data are reported to end in 2019, while HaPi continues to 2020. How was the 2020 land-use map produced?

Lines 233-263: The atmospheric deposition period is inconsistent: 2001-2019, 2000-2019, and extension back to 1900 are all mentioned. The treatment of 1860-1899 and 2020 should be explained.

Lines 247-263: Downscaling 2.5° model output to 5 arcmin does not necessarily provide real information at 5-arcmin resolution. The authors should avoid overstating the effective resolution.

Lines 525-537: Could the authors briefly justify the use of a uniform 20% uncertainty for FAO-based data? Data quality may vary across countries, time periods, and variables, so a short explanation would be helpful.

Line 378: “Consumer of manure P” is not appropriate because manure application and deposition are inputs rather than consumption. “Manure P input” would be clearer.