Mapping Vegetation Dynamics in Wyoming: A Multi-Temporal Analysis using Landsat NDVI and Clustering

Abstract. This research focused on mapping vegetation growth patterns in non-irrigated fields using five early season (2019–2023) Normalized Difference Vegetation Index (NDVI) derived from Landsat imagery. As part of this study, we compared the outputs generated by two unsupervised machine learning algorithms with a conventional image clustering technique. NDVI data were screened for outliers using the interquartile range method. Gaussian Mixture Models (GMM), Affinity Propagation (AP) and a traditional rule-based classification were used for clustering the pixels in the five NDVI images. GMM assigns data points probabilistically, assuming data are generated from a mixture of Gaussian distributions, while AP identifies clusters by finding representative exemplars without needing a predefined number of clusters. To evaluate the performance of these clustering algorithms, we assigned the clusters into six classes based on vegetation growth patterns observed over the five-year period. Class 1 represents five years of good vegetation growth, class 2 represents four years, etc. We used Intersection over Union (IoU) score to measure how well the classes represented in the final products compared to each other. When compared to rule-based classification product, AP generated product had an aggregate IoU score of 0.63, while GMM generated product had 0.59. GMM and AP detected finer NDVI variations that the rulebased method missed. AP’s exemplar-based approach provided a better understanding of vegetation pattern compared to GMM and rule-based method. This study highlights the importance of using advanced clustering techniques over traditional approaches for vegetation analysis, contributing to improved environmental monitoring and management decisions.