**1. Introduction**

Small scale farms represent about 80% of the farming area of China. Given the need to produce more food on the same amount (or less) of land while also reducing environmental pollution, such areas are faced with tough challenges. Farmers manage their fields by experience and need science-based evidence to make the system more efficient. The mismanagement of nitrogen (N) fertilizer is a known problem in the smallholder farming systems of China [1].

Precision nutrient managemen<sup>t</sup> can be achieved either through a sensor-based or a map-based approach. The former uses sensors to guide site-specific N managemen<sup>t</sup> based on the quantification of crop reflectance. However, a sensor-based approach is affected by the inter-annual interactions between soil and weather. Colaço & Bramley [2] demonstrated how the sensor-based approach could be improved by also considering the impact of other environmental covariates. The latter approach consists of using multiple images (e.g., soil, remote sensing, yield monitor) with the aim of dividing the fields into managemen<sup>t</sup> zones (MZs). MZs are defined as sub-regions within the field that have similar combinations of yield-limiting factors and are managed accordingly [3,4]. Basso et al. [5] stated that understanding the factors that lead to the spatial and temporal variability of a crop within the field is the first step for optimal agronomic management. In addition, the delineation of fields into MZs helps with obtaining soil/crop samples cost-effectively and applying site-specific agronomic input [4–7].

Several approaches have been proposed to define MZs at the field level. One approach is based on gathering soil or landscape information, such as sampling the soil using electrical conductivity (EC) sensor, sampling for soil physical and chemical properties or using remotely sensed data for estimating soil and landscape properties [8–12]. Another approach uses yield maps or remotely sensed images to reconstruct spatial and temporal patterns of crop growth within the field to define a given number of MZs [13–16]. Finally, the integrated approach uses both soil-landscape and crop information to define MZs at field level [4,17].

However, the small size of single farms in the North China Plain (NCP) does not allow for cost-effective on-farm managemen<sup>t</sup> using farm-scale MZs. Moreover, yield maps are not available in most small-scale farming systems and, in such systems, farmers do not measure field-level yield at harvest [18], but measure total grain that they sell from all of their fields. Jin et al. [19] combined a crop simulation model with remotely sensed data to map yield heterogeneity on smallholder farms in East Africa. High-resolution satellite images were used to define a smallholder farming system, but the prevalence of small field size was one of the challenges in improving the performance of the approach proposed [19]. In addition, high-resolution satellite data usually has to be purchased from a private provider (e.g., RapidEye) or obtained from a contemporary open-source sensor such as Sentinel-2, which does not ye<sup>t</sup> have enough years available to capture the long-term inter-annual variability. A practical strategy is to divide fields in a village into several MZs disregarding the current field managemen<sup>t</sup> structure. Some of the common open-source satellites, such as Landsat that has a long time-series of data, may allow the mapping of fields in a village for MZ delineation purposes; therefore, each field can be classified in a given MZ, maintaining the existing boundary structure. Fields in the same MZ could then use similar managemen<sup>t</sup> practices or inputs optimised for their particular conditions and requirements regardless of their geographical proximity.

The objective of this work was to integrate soil information derived from remote sensing with crop yield proxies derived from historical satellite images to define MZs at the village scale.
