During the Peiligang period, the millet cropping patterns in the Guanzhong region, Central Plains, and Haidai region were dominated by the broomcorn millet, as indicated by the percentages of the counts (100 %, 32 %, and 55 %) and weights (100 %, 32 %, and 55 %) (Fig. 2). The relatively low percentages of millets in the Central Plains (Fig. 2b and e) and Haidai region (Fig. 2c and f) were mainly attributed to the quantities of rice excavated in a few sites, such as the Baligang (Deng et al., 2015), Jiahu (Zhao and Zhang, 2009), and Xihe sites (Jin et al., 2014). With a decrease in rice from the Peiligang to the early Yangshao period in the Central Plains and Haidai region, a steady rise occurred in the percentage of the counts (58 % and 85 %) and weights (62 % and 88 %) of broomcorn millet, which maintained an advantage over foxtail millet.

From the early Yangshao period to the Bronze Age, the percentage of the counts and weights of broomcorn millet dramatically decreased from approximately 60 % to 10 %, while those of foxtail millet gradually increased from approximately 30 % to 80 % (Fig. 2). Notably, during the late Yangshao period, foxtail millet surpassed broomcorn millet in most of the percentages of the counts and weights in the Guanzhong region (69 % vs. 29 %, 51 % vs. 41 %, respectively); Central Plains (72 % vs. 20 %, 57 % vs. 25 %); and Haidai region (53 % vs. 36 %, 33 % vs. 42 %).

The dominant role of foxtail millet improved during the Longshan period, as indicated by the percentages of the counts and weights in the Guanzhong region (77 % and 54 %, respectively), Central Plains (75 % and 50 %), and Haidai region (55 % and 24 %). Accordingly, the percentage of the counts (10 %–17 %) and weights (9 %–25 %) of the broomcorn millet reduced in these regions. Subsequently, the gap between the foxtail and broomcorn millets widened during the Bronze Age.

Prior to the late Yangshao period, the archaeobotanical studies in the APE were relatively rare and discontinuous. During the pre-Peiligang period, the Donghulin site in the Yanbei region excavated the earliest charred millets in China, dominated by the foxtail millet (Zhao et al., 2020) (Fig. 3b and e). Contrarily, the cropping pattern was dominated by the broomcorn millet during the Peiligang period in the Liaoxi region, accounting for 96 % and 98 % in the percentage of the counts and weights, respectively, particularly in the Xinglonggou site (Zhao, 2004) (Fig. 3c and f). Subsequently, during the early Yangshao period, although the percentage of the counts and weights of the foxtail millet greatly increased to 40 and 26 %, respectively, in the Liaoxi region (Fig. 3c and f), broomcorn millet still played a leading role in the cropping pattern, particularly in the Weijiawopu site (Sun and Zhao, 2013).

From the late Yangshao period to the Bronze Age, the percentages of the counts and weights of broomcorn millet decreased in the Ganqing (45 % to 30 % and 55 % to 27 %, respectively); Yanbei (46 % to 19 % and 65 % to 30 %); and Liaoxi regions (96 % to 35 % and 98 % to 48 %) (Fig. 3). The percentages of the counts and weights of foxtail millet gradually increased in the Yanbei (53 % to 79 % and 34 % to 58 %, respectively) and Liaoxi regions (3 % to 64 % and 1 % to 50 %), while the decrease in foxtail millet in the Ganqing region resulted from a rapid increase in wheat and barley (Fig. 3a and d). Summarily, although the cropping patterns in the APE were dominated by foxtail millet during the Longshan period and Bronze Age, broomcorn millet was still significant.

The millet cropping patterns in northern China exhibited diverse patterns in the MLY and APE. In the MLY, the millet cropping patterns strikingly transitioned from broomcorn to foxtail millet around 6000 cal. a BP (Fig. 4a), such as in the Yangguanzhai site, the only settlement enclosed by a completed moat during the Miaodigou period (6000–5500 cal. a BP) (Zhong et al., 2020). The percentage of the counts of foxtail millet rapidly increased from approximately 20 % to 80 % around 6000 cal. a BP and predominated afterward, and that of the counts of broomcorn millet changed in reverse (Fig. 4a).

By contrast, although it appeared that the transition also occurred in the APE around 6000 cal. a BP despite the inadequate data, the foxtail (∼ 60 %) and broomcorn millets (∼ 40 %) jointly made important contributions to the cropping patterns after 6000 cal. a BP (Fig. 4b). For example, the staple crops in the Shimao site, a super-large central settlement dated 4300–3800 cal. a BP, mainly comprised foxtail (73 %) and broomcorn millets (26 %) (Yang et al., 2022).

Based on the age, location, and percentage of crop counts of each site, the spatial distribution and temporal evolution of millet agriculture in northern China were divided into two stages.

Stage I (9000–6000 cal. a BP). Millet agriculture was confined to several possible centers of domestication along the marginal mountains of the Loess Plateau and Inner Mongolian plateau (Fig. 5a–b and f–g), known as the Hilly flanks or China's Fertile Arc (Liu et al., 2009; Ren et al., 2016). The primary agricultural region in northern China had been framed at Stage I despite several blank areas still existing within this region. The predominant crop was broomcorn millet in most sites of the Liaoxi region, Haidai region, Central Plains, and Guanzhong region, generally accounting for more than 60 % (Fig. 5f and g).

Stage II (6000–2221 cal. a BP). The discontinuous agricultural region linked together at Stage II and expanded to marginal areas without solid evidence of agriculture prior to 6000 cal. BP. The spread of millet intensified from the late Yangshao to Longshan periods in two directions (Fig. 5c–d and h–i): (1) westward routine to the Ganqing region along the Wei River and Hexi corridor around 5400 cal. BP (Leipe et al., 2019) and (2) northward routine to the Yanbei region along the middle Yellow River around 5200 cal. BP (Bao et al., 2018). The cropping patterns of most sites in northern China were now dominated by foxtail millet (mostly more than 60 %), although a few sites in the APE were still dominated by broomcorn millet.

Although foxtail and broomcorn millets could be identified based on clear diagnostic features of their carbonized seeds (Liu and Kong, 2004) and phytolith morphologies (Lu et al., 2009b; Zhang et al., 2011), the percentage of the carbonized millets appeared to exhibit a reverse proportion to those of the phytoliths in the archaeological context. The divergence between the macroremains and phytoliths may result from several factors, such as the different sources of crop grains and phytoliths or biases in the representativeness of quantity. Diagnostic phytoliths of millets were derived from the inflorescence bracts (Fig. 6g and i) which may have been discarded during threshing and dehusking. Besides, based on the analysis of modem crop husk phytoliths, a previous study suggested that the counts of the husk phytoliths of foxtail and broomcorn millets were proportional to the weights of their seed grains, reflecting the relative production rather than quantities of these two millet types (J. P. Zhang et al., 2010). However, even after they were converted into weights, the percentage of carbonized millets still contradicted those of the phytoliths recovered from the same sites (Fig. 6a–d), such as the Anban (Liu, 2014; J. P. Zhang et al., 2010), Lajia (Wang et al., 2015; Zhang, 2013; Zhao, 2003), Dongzhao (Luo et al., 2018; Yang et al., 2017), and Wangjiazui sites (J. P. Zhang et al., 2010; Zhao and Xu, 2004). Thus, there may be other factors contributing to this divergence, such as depositional and preservation biases.

Carbonized macroremains represented only a small and biased sample that had access to the fire, and the survival rates for the carbonization process varied among crops (Colledge and Conolly, 2014; Wright, 2003). Experiments on the effects of carbonization indicated that the carbonization temperature window of broomcorn millet (250–325 ∘C) was smaller than that of foxtail millet (270–390 ∘C), implying that broomcorn millet was less likely to be carbonized than foxtail millet in archaeological contexts and thus may be underestimated in the cropping pattern (Märkle and Rösch, 2008; Wang and Lu, 2020). However, the preservation bias may be the reverse for phytoliths. Based on empirical evidence, the fragments of η-type phytoliths from broomcorn millet were generally larger than those of Ω-type phytoliths from foxtail millet (Lu et al., 2009b), implying that broomcorn millet was more likely to be preserved archaeologically and maintain diagnostic attributes for identification. In sum, the amount of broomcorn millet may be underestimated in carbonized macroremains and overestimated in the phytolith assemblage. Thus, further taphonomic research on survival rates should be quantitatively conducted for the correction between foxtail and broomcorn millet.

Furthermore, the way of seed yield may also affect the biases between foxtail and broomcorn millets. Morphological data from the northern Chinese Loess Plateau showed that the size of broomcorn millet increased significantly during 5500–4000 cal. BP, while that of foxtail millet did not exhibit an obvious increasing trend (Bao et al., 2018). Given the contemporaneous increasing human population, the crop yield was supposed to increase to feed a large population. Nevertheless, the increasing yields of broomcorn millet may depend on the increase in the seed size, while that of foxtail millet may result from the increase in the number of seeds per plant. Thus, the different ways of seed yield may also be one cause for the increase in the proportion of foxtail millet within the total seed number of seeds. In brief, both biases in the carbonization process and the way of seed yield may have exaggerated the proportion of foxtail millet.

Though the transition from broomcorn to foxtail millet around 6000 cal. a BP occurred in both the MLY and APE (An et al., 2010; Li et al., 2021; Wang et al., 2016), the cropping patterns diversified afterwards to adapt to regional environments, which were predominated by foxtail millet in the MLY (Fuller and Zhang, 2007; Lee et al., 2007; Song et al., 2019; Yang et al., 2020; Zhang et al., 2014) and focused on foxtail and broomcorn millet in the APE (Bao et al., 2018; Sheng et al., 2018). The spatial divergence of the cropping patterns may be attributed to the crop traits and regional climate.

Compared to the foxtail millet, the broomcorn millet exhibited better resistance to saline–alkali soil and arid climates (mean annual precipitation: 400–700 mm vs. 300–500 mm) and required a shorter growing season and lower-growing temperature constraints (mean annual temperature, 8–10 ∘C vs. 6–8 ∘C) (Shanxi Academy of Agricultural Sciences, 1987; Wang, 1996), and yet its yield was typically twice as low (1170 kg hm-2 vs. 750 kg hm-2) (Dong and Zheng, 2006). In sum, foxtail millet was more productive while the broomcorn millet was more environmentally tolerant.

Furthermore, though the East Asian monsoon rain belt had migrated northwestward ∼ 150–300 km during the mid-Holocene (Dong et al., 2021; Yang et al., 2015), the persistent drying trend since the mid-Holocene in northern China (Goldsmith et al., 2017; Yang et al., 2021) may have formed diverse ecosystems and millet selection. Thus, under a gradual aridification trend since approximately 6000 cal. a BP, drought-tolerant broomcorn millet had been replaced by high-yield foxtail millet in the subhumid MLY but was still significant in the semiarid APE.

Based on the analysis of the archaeological and paleoclimatic records, it appeared that a diffusion, transition and divergence of millet agriculture occurred in northern China around 6000 cal. a BP under the combined influence of climate change and cultural factors. The reasons are as follows.

First, although there was a dispute over whether a cooling or warming trend occurred in the late Holocene (Bova et al., 2021; Marsicek et al., 2018; Osman et al., 2021), global temperature anomalies reconstructed from a multi-proxy database of paleotemperature records exhibited ∼ 0.6 ∘C of warming from 10 000 to 6000 cal. a BP (Kaufman et al., 2020; Marcott et al., 2013), known as the Holocene thermal maximum (Renssen et al., 2009). In addition, pollen-based East Asian monsoon precipitation exhibited a time-transgressive pattern of the maximum precipitation shift from southern to northern China (X. Zhou et al., 2016, 2022), and peak precipitation occurred between 8000 and 5000 cal. a BP in northern China (Fig. 7a), as indicated by pollen records in the Dali, Daihai, and Gonghai lakes (Chen et al., 2015; Wen et al., 2017; Xiao et al., 2004). Thus, the warm and wet climate during 8000–6000 cal. a BP with a maximum combination of precipitation and temperature (Chen et al., 2015; Dong et al., 2022), defined as the Holocene Optimum, may have greatly promoted the intensification of the millet agricultural shift from broomcorn millet to the productive foxtail millet in the MLY and the diffusion of millet agriculture from the subhumid MLY to the semiarid APE.

Second, the summed radiocarbon probability distribution suggested that the population in North and Northwest China experienced a rapid increase and reached a high level during 6500–5000 cal. a BP (Fig. 7b) (Wang et al., 2014). Additionally, an explosive growth of archaeological sites occurred from the early (n=217) to late Yangshao periods (n=3817) (Hosner et al., 2016), extensively distributed along the middle Yellow River (Fig. 8). Consequently, the interaction among different cultural subregions unprecedentedly intensified since 6000 cal. a BP and formed the Chinese Interaction Sphere (Chang, 1986) or Miaodigou Age (6200–5500 cal. a BP) (Han, 2012). Thus, the cultural expansion and population explosion prompted the selection of the productive foxtail millet and the spread of millet agriculture around 6000 cal. a BP in northern China, coinciding with the origin and diffusion of the Sino-Tibetan evidenced by linguistics and genetics (Wang et al., 2021; Zhang et al., 2019).

Finally, as the main growing area of millet agriculture (He et al., 2017), the Chinese Loess Plateau was generally unsuitable for intensive arable production because of the small amounts of total nitrogen and phosphorus in the Loess (Catt, 2001). Therefore, field management practices, particularly manuring (animal waste), played significant roles in the intensification of millet agriculture in northern China. Modern field experiments suggested that the manuring of cereal crops could significantly elevate their δ15N values above those grown in natural or unmanured soils (Bogaard et al., 2013), with an offset range of 5 ‰–7 ‰ for millets (Christensen et al., 2022). Elevated nitrogen isotope values of the millet grains recovered from the Baishui valley indicated the use of animal manure as a growth-promoting factor since the late Yangshao period (X. Wang et al., 2018). Besides, a recent study in the Dadiwan site also suggested that an intensive crop–livestock system was in practice around 5500 cal. BP (Yang et al., 2022). Considering that the yield of foxtail millet was typically twice that of broomcorn millet, the dramatic increase in soil fertilization may promote the conscious choice of high-yield foxtail millets to maximize agricultural productivity.