Availability of Sentinel-2-based time-series observations: which vegetation phenology-based metrics perform best for mapping farming systems in complex landscapes?

The characterization of intra- and interannual variations in optical satellite observations has proven to be effective to differentiate between land use classes. However, there is relatively little knowledge of the accuracy and usefulness of various satellite-basedphenology variables and metrics. This letter provides quick insights into the processing of vegetation seasonality and phenology data, applied to the European Space Agency (ESA) Sentinel-2 satellite observations. The following time-seriesdata methods were compared for their potential to enhance land use mapping in two study areas in Kenya and Tanzania: a) monthly maximum Normalized Difference Vegetation Index (NDVI) composites, b) NDVI-basedtime-seriesstatistics, c) NDVI-basedharmonic regressions, and d) threshold-basedNDVI-basedphenological metrics. When using the maximum NDVI composites, as classification inputs, the overall accuracies were found to be 7-18% higher than when using the other methods (predictors). Overall, the Muranga site (Kenya) showed a higher classification accuracy for the best performing predictor set (overall accuracy = 86.5%) than the Kilimanjaro site (Tanzania) (overall accuracy = 80.9%). These findings confirm the potential of phenology-basedcompositing methods for large-scalemapping of agro-ecologicalfarming systems.