Sieving the weeds from the grains: an R based package for classifying archaeobotanical samples of cereals and pulses according to crop processing stages

Vegetation History and Archaeobotany https://doi.org/10.1007/s00334-024-01006-7 ORIGINAL ARTICLE Sieving the weeds from the grains: an R based package for classifying archaeobotanical samples of cereals and pulses according to crop processing stages Elizabeth Stroud1 · Glynis Jones2 · Michael Charles1 · Amy Bogaard1 Received: 29 September 2023 / Accepted: 12 July 2024 © The Author(s) 2024 Abstract The R package CropPro is an open-access resource to classify archaeobotanical samples as products and by-products of different stages of the crop processing sequence for large-seeded cereal and pulse crops in south west Asia, Europe and other Mediterranean regions. It builds on ethnographic research and analysis conducted by Jones (Plants and ancient man: studies in palaeoethnobotany. Balkema, Rotterdam, pp 43–61, 1984), (J Archaeol Sci 14:311–323, 1987), (Circaea 6:91– 96, 1990) and a modified method by Charles (Environ Archaeol 1:111–122, 1998). CropPro provides functions, which allow users to construct triplots, to conduct discriminant analysis comparing archaeobotanical samples with ethnographic crop processing stages and to plot the discriminant analysis results. This paper provides two worked examples of the use of CropPro: the early medieval site of Stafford in the UK and the Bronze Age site of Tell Brak in Syria. These examples illustrate the use of the package for identifying crop-processing stages, and for assessing the relevance of taphonomic pathways other than crop processing. Keywords Crop-processing · Discriminant analysis · Weed seed attributes · R package · Cereal and pulse processing Introduction Understanding the crop processing stages represented by archaeobotanical remains is essential for identifying activity areas, seasonal activities, and storage protocols at early agricultural sites. The series of steps required to convert harvested crop material into clean grain has been recognized as one of the causes of variation in archaeobotanical samples (Dennell 1972, 1974, 1976; Hillman 1973). For this reason, determining the crop processing status of archaeobotanical samples is necessary in order to recognise the biases imposed by such activities on the composition of archaeobotanical samples, and to consider this bias during Communicated by F. Antolín. Elizabeth Stroud 1 School of Archaeology, University of Oxford, Oxford, UK 2 Department of Archaeology, University of Sheffield, Sheffield, UK interpretation. This includes changes in the proportions of different weed species, which can be particularly important when using weed species as indicators of cultivation regimes (e.g. Bogaard et al. 2005). Ethnobotanical studies on crop processing highlight how crop-processing sequences alter both the crop and weed composition of a sample (Hillman 1981; Jones 1984, 1987, 1990). Several archaeobotanists have conducted or used ethnographic research to understand the processing sequence of a range of crop species (see for example Hillman 1981, 1984a, 1985; Jones 1984; D’Andrea and Haile 2002; Peña-Chocarro and Zapata Peña 2003 for temperate cereals and pulses; Reddy 1997, 2003; Thompson 1998; Lundström-Baudais et al. 2002; Harvey and Fuller 2005 for millets and rice). Such research has been taken further, with the proportions and ratios of particular items within such ethnographic data used to infer the crop processing status of archaeobotanical material (see for example Hillman 1984b; Jones 1984, 1990). Jones (1984, 1987) used ethnographic data of the weed seed characteristics as a discriminant model, which provides a way of recognising the effect of crop processing on archaeobotanical samples. Ethnographic 13 Vegetation History and Archaeobotany work, conducted on the Greek island of Amorgos in the 1980s laid the foundation for statistical models used to identify archaeobotanical samples as the products and byproducts of different stages in the traditional crop processing sequence for large-seeded cereal and pulse crops in south west Asia, Europe, and other Mediterranean regions (Jones 1984, 1987). By collecting and characterising these (by-)products of processing, data were obtained for three different statistical models that allow a comparison between ethnographic and archaeobotanical data. Although the processing of these crops is applicable to a wide range of cereals and pulses, these models are not suitable for all crops, such as small-seeded cereals like millets, or those that are harvested without weeds like maize. The full details of this model is described in Jones (1984, 1987). This paper presents the R package CropPro, which provides, for the first time, openly accessible tools to conduct the same types of analysis as Jones (1984, 1987) and Charles (1998), as well as open access to the dataset behind the models, allowing anyone to use this method (ESM 1). CropPro enables the classification and comparison of archaeobotanical samples against the ethnographic data from Amorgos (ESM 1, Jones 1990). Three methods can be employed: triangular plotting, which compares the proportions of grains, rachis nodes and weed seeds, in order to gain insight into the processing of free-threshing cereals (see Jones 1990); a discriminant analysis that utilises the attributes of weed seeds to identify the products and by-products of cereal and pulse crop-processing (see Jones 1984, 1987); and another application of discriminant analysis, which again employs the attributes of wild/weed seeds, to assess the relevance of crop-processing versus alternative taphonomic pathways such as dung burning (see Charles 1998). Background Using the ethnographic data collected on Amorgos, Jones (1984, 1987) introduced a method for characterising products and by-products of the crop processing sequence from which archaeobotanical material is derived. Data from the processing of cereals and pulses (bread and macaroni wheat, six rowed hulled barley, oat, pea, lentil, common vetch, and grass pea) has been used to create predictive models to classify suitable archaeobotanical samples (e.g. those with a sufficient number of items). Three by-products and one product were selected for sampling because these would most likely be kept for later use, and so potentially recovered archaeologically. Discriminant analysis, a multivariate statistical technique and form of machine learning, was used to create a model based on key physical characteristics of the weed seeds accompanying the crop during processing. This model was subsequently used to classify the archaeobotanical samples. The three characteristics of the weed seeds used are: (1) the size of the seeds relative to the fine sieve mesh used to separate small weed seeds from cereal grain, (2) the tendency of the seeds to remain in seed heads, spikes or clusters after threshing and (3) aerodynamic properties (see Table 1) (Jones 1984). By utilizing these characteristics instead of specific species to distinguish crop-processing stages (...truncated)