On the combined use of phenological metrics derived from different PlanetScope vegetation indices for classifying savannas in Brazil

Abstract

Mapping of savannas in Brazil is challenging since there is no consensus on the best remote sensing strategy to deal with the spatial variability of some physiognomies and the spectral similarity of others. In this study, we evaluated the performance of 12 land surface phenology (LSP) metrics calculated from 70 cloud-free PlanetScope (PS) satellite images and three vegetation indices (VIs) for Random Forest (RF) classification of eight savanna physiognomies. The 12 LSP metrics were: the start (SOS), end (EOS), length (LOS), and mean (MGS) of greening season; the mean spring (MSP) and mean autumn (MAU); the VI peak (PEAK) and trough (TRG); the positions of the peak (POP) and trough (POT); and the rates of spring green-up (RSP) and autumn senescence (RAU). These metrics were calculated from the Green-Red Normalized Difference (GRND), Enhanced vegetation Index (EVI), and Normalized Difference Vegetation Index (NDVI). At the protected Ecological Station of Águas Emendadas (ESAE) in central Brazil, we compared the LSP classification in the 2017–2018 seasonal cycle against the VI classification in the 2017 dry season using an existent reference vegetation map for accuracy assessment. Furthermore, we analyzed the performance of the individual and combined sets of VIs and their derived LSP metrics for RF classification of the savanna physiognomies. The results showed that LSP added gains of 19.3% (EVI), 13.1% (NDVI), and 5.4% (GRND) to dry-season VI classification. The overall accuracies of the individual and combined sets of VIs and their retrieved LSP metrics generated gains of 22.8% and 28.1% in relation to the dry-season EVI. In the classification combining LSP metrics, the most important ranked predictors originated from the NDVI and EVI (e.g., TRG, PEAK, MSP, MGS, and RSP). Our findings highlight the importance of the combined use of high spatial and temporal resolution data of the Planet's satellite constellation for the classification of Brazilian savannas leveraging the information retrieved from vegetation phenology. However, when dense time series of a given sensor are not available for retrieving the phenological metrics, an alternative is to use combinedly different VIs calculated in the dry season, when the frequency of cloud cover is reduced over Brazilian savanna areas.