The advent of high-resolution instruments for time-series sampling poses added complexity for the formal definition of thematic classes in the remote sensing domain-required by supervised methods-while unsupervised methods ignore expert knowledge and intuition. Constrained clustering is becoming an increasingly popular approach in data mining because it offers a solution to these problems; however, its application in remote sensing is relatively unknown. This article addresses this divide by adapting publicly available constrained clustering implementations to use the dynamic time warping (DTW) dissimilarity measure, which is sometimes used for time-series analysis. A comparative study is presented, in which their performance is evaluated (using both DTW and Euclidean distances). It is found that adding constraints to the clustering problem results in an increase in accuracy when compared to unconstrained clustering. The output of such algorithms are homogeneous in spatially defined regions. Declarative approaches and k-means-based algorithms are simple to apply, requiring little or no choice of parameter values. Spectral methods, however, require careful tuning, which is unrealistic in a semisupervised setting, although they offer the highest accuracy. These conclusions were drawn from two applications: crop clustering using 11 multispectral Landsat images nonuniformly sampled over a period of eight months in 2007; and tree-cut detection using 10 NDVI Sentinel-2 images nonuniformly sampled between 2016 and 2018.
Supplementary notes can be added here, including code, math, and images.