A user-friendly manual to apply the water availability predictive model from D'Agostini et al. 2023 and D'Agostini et al. 2024 to your phytolith dataset

List of contents: What is this manual for What you need to know before starting You can apply the full model if… Prepare your datasets Apply the model Interpret the result Make predictions on plant water availability using a subset of morphotypes References Do you need help? Contact us! Required software and materials: A spreadsheet editor capable of supporting .xlsx and .csv formats (e.g., LibreOffice https://www.libreoffice.org/). R (https://cran.r-project.org/mirrors.html) or RStudio (https://dailies.rstudio.com/). A dataset of modern phytoliths (training dataset) derived from plants grown under controlled watering conditions, where samples of both irrigated and non-irrigated plants are available. Access links are provided below or download the file “Modern.csv” directly from this upload. A dataset of archaeological phytoliths (testing dataset). 1. What is this manual for Between 2021 and 2024, we developed a predictive model to reconstruct water availability for drought-tolerant crops in archaeological contexts using phytoliths. The model is intended to build hypotheses on the presence or absence of irrigation systems. The model's output provides a measure of the likelihood that the archeological phytolith assemblage under observation had formed from irrigated plants. In this step-by-step guide, we show you how to apply the model to your dataset and the theoretical framework behind it. The predictive model has been previously published in two articles: In D'Agostini et al. (2023, https://doi.org/10.1016/j.revpalbo.2022.104783), the production of phytoliths in three crop species was analyzed and modeled in relation to water availability, based on data from one experimental cultivation (conducted in 2019), and applied to an archaeological case study from the Indus Valley Civilization. D'Agostini et al. (2024, https://doi.org/10.1007/s00334-024-01012-9) combines the experimental results of 2019 with those obtained from an additional cultivation (replica) conducted in 2020. The 2019 and 2020 cultivations were conducted in the same location and under the same experimental conditions. Therefore,either dataset included in the 2023 and 2024 publications is valid as training data, but the 2024 dataset is more complete and based on a larger number of samples (361 entries). 2. What you need to know before starting The model was developed using three C4 cereal species: finger millet (Eleusine coracana (L.) Gaertn.), pearl millet (Cenchrus americanus (L.) Morrone, syn. Pennisetum glaucum (L.) R.Br.), and sorghum (Sorghum bicolor (L.) Moench). Ten landraces of each species from Ethiopia, Kenya, Pakistan, Sudan, and Tanzania were tested. For further information on the origin of the grains as well as on growth and productivity data, please refer to the supplementary database “FileS1-Experimental cultivation.xlsx” in D’Agostini et al. (2024) at the following link https://doi.org/10.5281/zenodo.10246367. Each of the landraces analyzed was subjected to two distinct water treatments to simulate rainfed conditions in an arid context and irrigated conditions that also acted as a control. For data regarding the amount of water received by each replicate in total and weekly, harvest and flowering time, and additional experimental data refer to “FileS1-Experimental cultivation.xlsx” at the following link https://doi.org/10.5281/zenodo.10246367. To understand the basis on which we set the treatments, refer to the methodology explained in both D'Agostini et al. (2023) and D’Agostini et al. (2024). Phytolith analysis was conducted on the leaves and chaff of 4 finger millet, 5 pearl millet, and 5 sorghum landraces. The morphotypes identified and counted include: ACUTE BULBOSUS, BILOBATE, BLOCKY, BULLIFORM FLABELLATE, CROSS, ELONGATE SINUATE clavate, ELONGATE SINUATE crenate, ELONGATE DENTATE, ELONGATE ENTIRE, ELONGATE SINUATE, POLYLOBATE, RONDEL, SADDLE, STOMA, TRACHEARY. For information on identification, classification, and photos of these morphotypes, please refer directly to the supplementary file “FileS2-Phytolith data.xlsx” (sheet Metadata - Phyto. classification) in D’Agostini et al. (2024), at the following link https://doi.org/10.5281/zenodo.10246367. You can consult the raw phytolith counts, the calculation of concentrations (expressed as grams of dry tissue), and the percentages of the total phytoliths counted in the same file in sheet Data. 3. You can apply the full model if… The water availability model can be applied to archeological assemblages of phytoliths using our phytolith dataset for model training (as done in D'Agostini et al., 2023 for sites of the Indus Valley Civilization) in two ways. Either utilizing the complete training dataset, which includes all morphotypes, or a subset of morphotypes. Check the conditions listed below to determine if your archaeological dataset meets the criteria for using the complete training dataset or just a subset of morphotypes. You can use the complete training dataset if… Your archaeological phytolith assemblage includes ACUTE BULBOSUS, BILOBATE, BLOCKY, BULLIFORM FLABELLATE, CROSS, ELONGATE SINUATE clavate, ELONGATE SINUATE crenate, ELONGATE DENTATE, ELONGATE ENTIRE, ELONGATE SINUATE, POLYLOBATE, RONDEL, SADDLE, STOMA, TRACHEARY. C₄ Panicoideae and Chloridoideae are considered predominant for the agro-environmental context/setting under study (e.g., in the Sahel regions of Africa where millets constitute the predominant part of agricultural production). Finger millet, pearl millet, and sorghum were likely present in the past because they have been identified through phytolith analysis and/or alternative means (e.g., charred grains, isotopic signal of C₄). …otherwise you can apply the model using a subset of morphotypes if: It is reasonable to assume that if most of the morphotypes listed here are part of your archaeological assemblage, then the presence of C₄ cultivated crops is likely. Therefore, by selecting only the phytoliths included in this list from your archaeological assemblage, the model can be applied. In addition, take into account that phytoliths from C₃ cereals are also effective in predicting water (Ermish and Boomgarden 2022; Jenkins et al., 2016, 2020; Madella et al., 2009). If C₃ cereals are identified in the archaeological context under study, morphotypes that are redundant in both C₃ and C₄ cultivated crops but that cannot be taxonomically distinguished can still be used (e.g., RONDEL). You work in areas dominated by C₄ grasses but have a few of the listed morphotypes. In this case, the absence of certain morphotypes could be due to: (a) taphonomic factors that affect some morphotypes more than others (see for example Cabanes and Shahack-Gross 2015). In this instance, we suggest removing the missing morphotypes from the training dataset and apply the model using a subset of morphotypes; (b) the fact that not all three cultivated species used in the training model are/were present but the presence of at least one of these species can be confirmed through a combination of diagnostic morphotypes and/or other proxies. If you are going to use the full model, proceed to sections 4-6. If you need to use a subset of morphotypes or only one morphotype (or several morphotypes but analyzed individually) for your prediction, then proceed directly to section 7. 4. Prepare your datasets Two datasets will be needed: the dataset of modern morphotypes to be used as training data and the dataset of morphotypes derived from an archaeological context to be used as test data. The two datasets must meet the following criteria: They should be two separate CSV files. Both files must have the same column headers and number of columns. This means that the list of phytoliths identified in the archaeological record that are not present in the modern training dataset should be excluded (e.g., those clearly associated with C₃ like SPHEROID and CRENATE). Both datasets must express the concentrations of morphotypes as a percentage of the total phytoliths counted (no absolute counts, no concentration). If some of the columns have been excluded from the analysis, it is essential to retain the percentages as they are (even if the total no longer adds up to 100%) so that the relative abundance of the phytoliths excluded from the analysis is still accounted for. Prepare the training dataset The training dataset can be downloaded directly from this guide (below files “Modern.csv”) or from the supporting data of D'Agostini et al. (2024) “FileS2-Phytolith data.xlsx" at https://doi.org/10.5281/zenodo.10246367. In the Data sheet, select and copy only the columns "Sample_Acronym", "Species", "Treatment", and "Origin" to obtain a unique identifier for each modern sample. Subsequently, select the columns corresponding to the percentages of each morphotype: TriPER, SaddlePER, RondelPER, BulliParallelPER, BulliFlabellatePER, BulliTOTPER, ClavatePER, SinuatePER, EntirePER, CrenatePER, StomaPER, CrossPER, BilobatePER, PolylobatePER. Create a new CSV with the column copied from the training dataset and name it: Modern.csv. Be sure to use the filing name suggested, so that the R code will run without need to be modified. If you prefer to use a different file naming, make sure you change them in the R code as well, when appropriate. Prepare the test dataset (archaeological) Select the same morphotypes in your archaeological dataset and proceed with the creation of a CSV file, analogous to “Modern.csv”, with identical columns and name it: Archaeological.csv. Since the information regarding Species and Treatment is not available for the archaeological samples (and specifically Treatment is the variable we want to make predictions on), fill the Species column with a label of your choice that seems representative of the sample and useful for interpretation purposes. In D’Agostini et al. (2023), for example, we filled it with the type