Using Hidden Markov Models for Land Surface Phenology: An Evaluation Across a Range of Land Cover Types in Southeast Spain

Abstract

Land Surface Phenology (LSP) metrics are increasingly being used as indicators of climate change impacts in ecosystems. For this purpose, it is necessary to use methods that can be applied to large areas with different types of vegetation, including vulnerable semiarid ecosystems that exhibit high spatial variability and low signal-to-noise ratio in seasonality. In this work, we evaluated the use of hidden Markov models (HMM) to extract phenological parameters from Moderate Resolution Imaging Spectroradiometer (MODIS) derived Normalized Difference Vegetation Index (NDVI). We analyzed NDVI time-series data for the period 2000–2018 across a range of land cover types in Southeast Spain, including rice croplands, shrublands, mixed pine forests, and semiarid steppes. Start of Season (SOS) and End of Season (EOS) metrics derived from HMM were compared with those obtained using well-established smoothing methods. When a clear and consistent seasonal variation was present, as was the case in the rice croplands, and when adjusting average curves, the smoothing methods performed as well as expected, with HMM providing consistent results. When spatial variability was high and seasonality was less clearly defined, as in the semiarid shrublands and steppe, the performance of the smoothing methods degraded. In these cases, the results from HMM were also less consistent, yet they were able to provide pixel-wise estimations of the metrics even when comparison methods did not.