*5.2. Combining Samples from Multiple Regions*

A valid question arises as to whether combining samples from multiple regions can make models more robust for prediction across environments. Higher performance was obtained when the US and ET samples were combined. However, these higher *R*<sup>2</sup> values are likely artifacts of a larger spread among the dataset (larger range of levels of nutrients across locations).

It is worth noting that in three instances, the *RMSEP* from the combined USET dataset was greater than the minimum value of the nutrient in ET (see Table 2). These included Ca and Mg at the plant level for the PLS-Full model (Table 6) and Ca at the plant level for the PLS-Wave model (Table 7). Upon further investigation, it was determined that two of these samples were outliers (PLS-Full Mg and Ca), where the value was more than two standard deviations from the mean. Thus, these issues do not appear to be widespread or have a large impact on predictive capabilities of the models. However, it is important to keep this point in mind when combining data from differing regions with varying environmental factors and agricultural practices, where there may be large disparities in nutrient content. These disparities may result in the identification of between-site variation, requiring a greater variation of nutrient values to fully establish a generalized model.

Nevertheless, the results from the combined dataset are promising because they suggest that increasing within-sample variation can improve PLS model predictability across study areas. It is also worth mentioning that this practice could result in a reduction of accuracy of measurement, especially for field measurements, as the data may take on more noise. In this case, the use of single region analytics may be more beneficial. Nonetheless, improvement of predictability has been a similar finding in past studies that have referred to a need for global (i.e., USET) versus local (i.e., US and ET) modeling within similar measurement collections such as those in soil geochemistry studies [58,59]. Since it is unlikely though that this range of variation would be captured within a single site in many cases, it becomes necessary to combine samples from a variety of locations, and likely from a variety of study sites and research groups. To this point, we recommend building a more open and

transparent culture of data sharing within the remote sensing community that can permit data sharing to advance our modeling capabilities and promote development of more generalizable predictive models. We recognize that a shift is already underway in many disciplines to promote data and code sharing, with some journals mandating these components accompany manuscript submission to allow reproducibility checks. Our findings here suggest that sharing data may have broader impacts beyond simply reproducing tests, but rather could result in more robust predictive models.
