Abstract. The primary objective of this study was to compare the density-based (OPTICS-Ordering Points To Identify the Clustering Structure) and graph-based (Spectral Clustering) algorithms for identifying Normalized Difference Vegetation Index (NDVI) -based crop growth clusters in multi-temporal satellite imagery from two crop fields in southeast Wyoming, USA. The clusters represent spatial groupings of pixels with similar NDVI values, corresponding to relative crop growth and vigor conditions across the fields. This study evaluated the similarities and differences in the clusters generated by these two unsupervised Machine Learning (ML) algorithms. Spectral clustering cannot find the number of clusters, so eigengap was used to estimate the number of optimal clusters. The same number was enforced in OPTICS which used the parameters: min samples, max eps and Xi. OPTICS generated more fragmented and fine-scale clusters, especially in higher NDVI ranges, whereas Spectral Clustering produced smoother, more contiguous zones, particularly in moderate to low NDVI areas. The cluster output images generated by Spectral and OPTICS had only 55% overlap. It showed that OPTICS is better suited when the objective is to detect fine-grained variability in crop vigor, especially in dense vegetation, while Spectral Clustering is more effective for identifying broad growth zones and overall field patterns. The choice of algorithm should therefore depend on whether detailed local differences or generalized field-wide structures are of greater importance.
Patel et al. (Mon,) studied this question.