14.Limitations_knowledge_gaps_and_way_forward.rmd

# | Limitations, knowledge gaps and way forward 

It is of most importance to identify which regions, environments and production systems present greater potential to increase SOC stocks and mitigate GHG emissions and establish priorities for research and implementation of public policies. In this document, we provided an approach and the procedures to produce digital SOC sequestration maps using soil legacy data, process oriented SOC models and modern techniques of digital soil mapping, that would allow covering as many conditions and productive systems worldwide as possible, in a relatively simple, transparent and standardized way, without complex configuration and computational capacities. Nonetheless, the estimation of SOC sequestration potential in a harmonized way among countries, regions, and productive systems is not an easy task and there are different contentious issues and limitations that must be outlined.  

Firstly, agricultural lands (croplands and grazing lands) are selected as target areas to estimate SOC sequestration potential in this first instance, since they are managed at least on a yearly basis, and management practices could be used to increase soil organic carbon content. These lands have been identified as the options with greater potential to accumulate SOC and mitigate GHG emissions through improved management practices (Smith et al., 2008; Lal et al., 2018). Furthermore, most of the information regarding the SOC dynamics has been developed in these productive systems, and most SOC carbon models have been successfully tested under these conditions. Countries can nevertheless assess SOC sequestration potential of different land uses and deliver additional maps including other land use (other than agricultural lands). Future versions of the GSOCseq map may include other land uses, depending on national demands.  

Secondly, most SOC models are parameterized under land use, land management, soil or climatic regions. Ideally, SOC models should account for all major SOC-controlling factors, such as soil mineralogy, climate conditions, litter quality, biota activity, land use and management. These factors have extremely complex interactions, and separate analysis of controls could limit predictions of their effects on SOC (Falloon and Smith, 2009). Even the full multidimensional development of a single element of a model can rarely, if ever, be predicted precisely, and the actual consequence is that it is impossible to create "universal" models (Sinclair and Seligman, 1996). The review by Campbell and Paustian (2015) emphasizes the fact that among the different known process-oriented models used to estimate SOC changes, no one clearly outperforms the others. However, in order to obtain consistent and harmonized results, and allow comparisons between countries and regions, due to potential differences in computational, technical capacities and data availability, the use of RothC as a standard 'process-oriented' SOC model, following the proposed methodology, is requested as a first step. Nevertheless, users are encouraged to provide supplementary alternative maps developed using alternative preferred SOC models and/or methods or approaches to estimate C inputs and compare results with the proposed methodology. The use of a multi-model ensemble approach (e.g. Riggers et al, 2019; Lehtonen et al., 2020) with selected models is intended for future versions of the GSOCseq map. Moreover, the SoilR package (Sierra et al., 2012) used in the current approach already includes other SOC models like CENTURY and ICBM, that   can be used to estimate results using a multi-model ensemble approach.  

It must be also outlined that at some level of analysis all known process-oriented SOC models including RothC (see Chapter 2; see FAO, 2019), include empirical functions, so they are expected to perform best when operating in situations similar to those for which they were originally parameterized, which tend to be croplands and grasslands from the temperate zone (Jenkinson et al., 1990; Petri et al., 2010). There is relatively less available data of the performance of SOC models under tropical and arid conditions. Current SOC models, including RothC,  may be limited in their applicability to these systems, due to differences in soil fauna and their effects on SOC dynamics, the much faster turnover of slow and passive SOM, different temperature and moisture relationships with microbial activity, and differences in mineralogy (Shang and Tiessen, 1998; Tiessen et al., 1998) and solution chemistry (Parton et al., 1989) in tropical soils, or water dynamics under arid environments (Farina et al., 2013). The inability to account for cation availability or aluminium (Al) toxicity may also limit SOC model predictions (Parton et al., 1989; Shang and Tiessen, 1998). Most well-known models may be limited by failing to account for pH effects on soil carbon turnover (Jenkinson, 1988; 1996; Falloon and Smith, 2009). Soil organic carbon models generally predict faster carbon turnover than the ones observed in very acid soils (Motavalli et al., 1995), and few models can predict SOC changes in allophanic soils or soils developed on recent volcanic ash (Jenkinson et al., 1991; Motavalli et al., 1995; Falloon et al., 1998; Falloon and Smith, 2000; Falloon and Smith., 2009).  

Additionally, RothC does not accurately simulate SOC dynamics in waterlogged soils such as paddy rice. Countries are encouraged to use local adaptations or modified versions of the RothC model which have shown to improve estimations under the above mentioned conditions (e.g. Parshotam and Hewitt, 1995; Saggar et al., 1996 for volcanic soils; Shirato et al., 2004; Shirato and Yokozawa, 2005 for paddy rice; and ROTHC10 developed by Farina et al., 2013 in arid conditions). Local adaptations should be implemented following the general procedures and input data described in Chapter 5 and 6 to obtain consistent results, and/or to use their preferred model (e.g. Gilhespy et al., 2014) under these conditions, and deliver additional maps to contrast results. Further developments of the GSOCseq will include specific and standardized methods for SOC estimations in paddy rice and other specific conditions.  

The proposed approach estimates SOC changes in the first 0-30 cm. Although SOC at deeper soil layers is responsive to land management changes (e.g. Follett et al., 2013; Poeplau and Don, 2013; Schmer et al., 2014), the 0-30 cm is selected because: it is most responsive to land management changes; allows the use of GSOCmap as a baseline for SOC stocks; allows for better harmonization with national greenhouse gas inventories, and allows validation of selected models with available ground data (mostly generated at 0-30cm depth). New models and adaptations of known models have been developed to account for SOC dynamics in deep layers with different approaches (see Campbell and Paustian, 2015). For example, the DAYCENT model was modified to simulate deeper soil C dynamics by slowing SOC pool turnover and increasing allocation to passive soil C, without separating soil layers (Wieder et al., 2014). Jenkinson and Coleman (2008) modified RothC to RothPC-1 to predict the turnover of organic C in subsoils up to 1 m of depth using multiple layers and introduced two additional parameters, one that transports organic C down the soil profile by an advective process, and one that reduces decomposition processes of SOC with depth. However, there is still a strong necessity for additional data to confirm or refute hypotheses suggested by the different modelling approaches of SOC in deep layers (Campbell and Paustian, 2015). As new information is generated, future versions of the GSOC and GSOCseq maps will be able to incorporate SOC stocks and SOC changes at deeper layers.
There is also still need for a better understanding of spatial heterogeneity in SOC in the landscape and for a better prediction of potential changes in SOC dynamics on the landscape scale (Stockmann et al., 2013). Differences in drainage that may be linked to landscape position are often not accounted for in SOC models (Falloon and Smith, 2009). In this sense, three gaps in knowledge have been identified (Stockmann et al., 2013): (1) the development of optimal, but still simple, 3-dimensional representations of landscapes (vertically and horizontally), (2) the implementation of functional interactions and SOC transfers (i.e. the redistribution of SOC to different parts in the landscape due to erosion, transport and deposition) and (3) the availability of adequate datasets for model validations (especially the representation of fluxes between different landscape elements).  

It should be also outlined that the temperature is expected to increase in the next 20-50 years, especially after 2050 (IPCC, 2018) and this may impact SOC dynamics.  The proposed approach considers a 2000-2020 climate average for SOC projections after 2020.  Using a 20-year average removes the year-to-year variation. However, there is no consensus over which climate projections are the most appropriate for 2020-2040, and prior agreement between countries is required. The proposed methodology allows climate change scenarios for longer-term projections to be incorporated in future versions.  

It should be also noted that a very wide range of management practices are currently being implemented and can potentially be introduced into the world's agricultural systems, depending on climatic, soil, socio-cultural and economic conditions. In turn, different SSM C-oriented practices are often combined, making it difficult to dissociate their effects on SOC dynamics. Thus, as a first step, and to harmonize the results on a global map, and because soil carbon turnover models are the most sensitive to carbon inputs (FAO, 2019),  this manual proposed to group SSM practices into three scenarios as a standard method, based on their expected relative effects on C inputs compared to business as usual management: Low, Medium and High increase in C inputs (referred as SSM1, SSM2, and SSM3 scenarios). A 5%-10%-20% increase in C input is suggested as default values to test potential. This increase in C-inputs will not always be possible where C-inputs are already high. On the other hand, this increase in C inputs can be low in regions or productive systems with current low C inputs.  However, to obtain consistent and harmonized results, and allow comparisons between countries and regions the use of standard SSM scenarios and 5-10-20 % is kindly requested in this first stage. National experts' opinion and local data are essential in order to accurately estimate or validate the target areas and carbon input levels for the different SSM scenarios in forthcoming versions. Countries are encouraged to provide supplementary alternative maps developed using alternative % increases in C inputs or specific absolute increases in C inputs of specific SSM practices in the different agro-ecological regions and productive systems of the country, based on local knowledge or obtained from a literature search of local studies. We believe the comparison of results will greatly enrich the final product. The information generated by the different countries will allow us to select and model specific practices in forthcoming versions.  

Finally, the precision of models relies heavily on the quality and quantity of data used in executing and validating them (FAO, 2019). Often, the datasets for running models are not collected for that specific purpose but are taken from previous or ongoing studies. In many cases the format and amount of data may be inappropriate for the models. There may be several potential pitfalls for the integration of data to calibrate, drive and evaluate a SOC model. Careful harmonization of datasets and input estimation methodologies is essential to obtain consistent results across regions and countries. Ideally, calibration and driving data should match the scale of the model simulation. However, data limitations will prompt the use of data of coarser resolution and/or mixing data of varying quality from different sources. The 1x1 km scale for the GSOCseq is required in the final product to allow comparisons among countries. However, input datasets from different resolutions will be probably used, and this may introduce uncertainties (e.g. climate data that usually occurs at coarser resolutions). In the proposed approach, global sources are proposed (same resolution and quality) but countries are encouraged to develop and deliver SOCseq maps using the best available national climatic, soil, and land use data.  

Data availability for model evaluation will also affect the assessment of model accuracy, as well as its ability to support hypothesis testing. Although there is a wealth of measured data from carefully monitored long-term agronomic experiments to evaluate SOC models, especially in the northern hemisphere and temperate climate conditions, there are comparatively few similar datasets from natural ecosystems (Falloon and Smith, 2009). The suite of datasets may then become sources of uncertainty in SOC model predictions (Keenan et al 2011; Palosuo et al 2012). Datasets are also often difficult to identify or compare between SOC models, particularly in large-scale ecosystem or global analyses. Furthermore, soil carbon measurements from available experiments are rarely available in replicate and hence attributing uncertainty to these measurements, and ultimately confidence in SOC model predictions is limited (Falloon and Smith, 2003). Data availability to validate model performance will be a limitation for many countries.   

A meta-analysis should be conducted on the basis of the latest available local and regional studies to estimate how agricultural practices affect average annual C inputs, SOC sequestration rates and SOC stocks. However, meta-analyses and comparisons have often suffered from datasets based on diverging definitions (e.g. concerning definitions of sample depth, baselines for comparisons, or the components of soil respiration), (Bahn, 2009). We hope that this exercise, together with other GSP activities,  will also be an opportunity for the different countries to establish long-term observatories that will allow us to monitor the effect of different management practices on SOC stocks under different environments, and this will in turn allow us to improve model estimations.
We acknowledge that consistency among input datasets and results would be improved if there was only one actor involved in the entire process. However, it is of most importance that information is locally generated, involving local experts and institutions, building technical capacities in the process, as this will encourage countries to implement national and subnational policies, and to get involved in regional and global policies. Moreover, we believe that a 'bottom-up', country-driven approach, using country specific data and expert knowledge, is a fundamental step for iterative improvements. As it is the case of other GSP documents, this manual constitutes a living document, which will be continuously improved and refined after its use and implementation.