Delineation of management zones dealing with low sampling and outliers

Purpose

Management zones (MZs) are the subdivision of a field into a few contiguous homogeneous zones to guide variable-rate application. Delineating MZs can be based on geostatistical or clustering approaches, however, the joint use of these approaches is not usual. Here, we show a joint use of both techniques. The objective of this manuscript is twofold: (1) compare different procedures for creating management zones and (2) determine the relation of the MZs delineated with i) coffee yield maps and ii) the summarizing power of each method for each input variable inside the MZs delineated.

Methods

The techniques compared to summary spatial data were: (1) summarizing the variables into a soil fertility index (SFI), (2) the MULTISPATI-PCA technique, and (3) the multivariate Min/Max autocorrelation factors (MAF) approach. Then, clustering methods were applied to perform field partition into binary MZs (grouping lower and higher values of input variables).

Results and discussion

The MAF approach achieved the best field partition regarding clustering metrics (McNemar’s test, Silhouette Score Coefficient, and variance reduction). In this paper we did not use yields as a cluster variable but as a measure of success. MAF also was the best one for separating low- from high-yielding areas over the MZs. The results show that the proposed approach could be effectively used for management zone delineation.

Conclusions

This methodology facilitates evaluating innovative approaches in challenging spatial modeling scenarios, such as low-sampled fields with outliers. A wide range of summarization methods and clustering techniques are available, making this agnostic approach quite interesting for delivering MZ maps. This flexible approach can guide precision nutrient management in low-sampled areas, allowing the joint use of data science tools and agronomical knowledge to delineate variable rate application strategies.

Graphical abstract