**4. Discussion**

Overall, the GNDVI derived from remote sensing allowed for the discrimination of zones within the village, with a reasonably good explanation of the variability in measured N and OC. The GNDVI values were low in years in which the minimum growing season temperatures were higher than in other years (e.g., 2008 c.f. 2011). This may have resulted in m the crop developing faster, meaning that by the time the satellite data were acquired (grey box in Figure 2) the crops might have been at an advanced developmental stage i.e.past flowering, meaning higher senescence rates causing lower GNDVI values. For years when minimum growing season temperatures were lower, meaning a longer wheat growing season, by the time our images were acquired, the wheat would still be at the flowering stage and therefore with less senesced material. In addition, there were differences in GNDVI index values between TM and OLI because of the differences in the wavelengths for each of the band that the sensor collected. The image data were not corrected for this. There was a very narrow range of the measured N and OC within the village along with the unmeasured variability in field managemen<sup>t</sup> and inputs, both of which would have affected the relationships. From a statistical point of view, the soil N and OC showed a narrow range of values. However, from an agronomic point of view, the values of OC ranging between 0.5 and 1.2% have big differences in terms of impact on soil chemical processes and impacts on yield. In fact, soil organic matter is a reserve for nutrients and an agen<sup>t</sup> that improves soil structure. It is a storage pool of plant nutrients. In addition, the humus (which is the stable fraction of the soil organic matter) adsorbs and holds nutrients in a plant-available form. Soil organic matter also releases nutrients in a plant-available form upon decomposition [40].

Satellite images of crops provide an indirect tool to obtain spatial information of crop growing conditions for a given year and therefore are a good tool to quantify spatial field variability [26,41]. The use of a longer remotely sensed time-series enabled the quantification of temporal stability within the spatial context of the village. Satellite images of the bare soil also provide an indirect method to obtain spatial information about the variability in soil conditions, in particular soil moisture, which affects crop growth. However, one limitation of the soil brightness is that it is a weak proxy of soil moisture because the soil–plant relationship is deeper than the first few cm of soil. In addition, whilst there was a significant correlation between soil brightness and measured soil OC, this relationship was not strong (0.19). The soil brightness was reconstructed by a private company and it is not a good proxy of the soil samples measured in the field during the study. This is because the resulting soil brightness image is captured by the company on a given date at a time when the soil is bare (it could have been many months before the soil sampling), and given the high temporal variability in soil nitrogen concentration the time when soil samples were collected do not reflect the spatial variability of the brightness map.

The subdivision of the village into 3–4 zones improved the explanation of the variability in measured soil parameters. However, there is a trade-off between the number of zones and site-specific management. The spatial coherence of the zones also needs to be considered, since it is likely to be more economically and practically efficient to zone the village fields into "blocks" rather than by individual field. It has been found in the literature that three zones are a common number adopted in commercial precision agriculture solutions regardless of the size of the field [42]. More zones could have been defined, but this would translate into more managemen<sup>t</sup> recommendations and organisation of farmers into more co-operative "clusters", which would add to the complexity and time commitment for co-operative leads.

The results of this study can be considered as a preliminary method based on the integration of di fferent remotely sensed data to delineate MZs at the village scale. More studies are needed to further refine them for guiding site-specific managemen<sup>t</sup> in small scale farming systems. In addition, incorporating measurements of field level yields would aid in validating this approach in the future.

The main limitation of this study is in the spatial resolution of the satellite data. If a higher resolution and consistently measured dataset had been available over a similar timescale, a more accurate measurement of spatial and temporal variability at the field scale would have been detected. However, recent advancements in sensors on free satellite products (e.g., Sentinel-2) will make the acquisition of a long temporal series of images with higher resolution easier in the future. In addition, the MZ approach could be improved with data on historic managemen<sup>t</sup> (e.g., timing of key operations) and crop yields. The lack of mechanization and extension services at the village level might hamper the application of modern technologies. Even the subdivision of a village into zones might be of no help if it is not coupled with additional information on how to translate this information into agronomic management.

The approach used in this study was developed considering the limited data availability in small scale farming systems used in the NCP. Therefore, it may not be the best approach if more data are available. In order to improve crop managemen<sup>t</sup> (e.g., sowing dates, fertilizer amount and timing), a decision support system (DSS) should be developed in order to integrate MZs and agronomic prescriptions. The design of a DSS should provide a science-based approach to quantify the optimal practices that can evaluate the trade-o ff between economic and environmental benefits. The DSS should be system-based in order to take into account the dynamic interactions between soil–plant–atmosphere–agronomy continuum [5,16]. Future research is therefore needed to overcome the limits highlighted in [19] for a better link between the remotely-sensed approach to define zones and crop models. It is likely that the development of simplified crop growth models will be a step forward for better integration. In this regard, [43] developed simplified models that are less data-intensive. In particular, [44] developed a simple scalable and satellite-based yield model to predict yield for canola (Brassica Napus L.) and wheat (Triticum spp.) at di fferent regional scales. The results of this study could be integrated within the modelling approach highlighted in [44] to provide the system-based DSS approach. Future research would also be needed to concentrate the e fforts to consider the impacts of other agronomic practices such as crop rotation (wheat–maize), the use of manure and tillage on the zoning.
