Comparisons between temporal statistical metrics, time series stacks and phenological features derived from NASA Harmonized Landsat Sentinel-2 data for crop type mapping

Spectrotemporal features that capture changes in reflectance over time are useful for characterizing the land cover of highly dynamic crops. Currently, temporal statistical metrics, time series stacks and phenological features are the three spectrotemporal features commonly used in crop type mapping. The three types of features differ in their calculation methods and physical implications. However, there has been limited investigation on the performance comparisons between them for crop type mapping. The objective of this study was to evaluate and compare the effectiveness of the three features derived from Harmonized Landsat Sentinel-2 (HLS) data for crop type mapping. The HLS data were first pre-processed with cloud masking, temporal compositing and gap filling to create the gap-free time series for extracting the three spectrotemporal features. Crop reference data were obtained through a field survey conducted over a study area of 14.5 km by 8 km near College Station, Texas, USA. For the calibration of the Random Forest (RF) classification model with different sets of spectrotemporal features, 30% of the total reference data were used, and the remaining 70% were used for quantitative accuracy assessment. Results showed that although all three spectrotemporal features yielded accurate crop type maps, time series stacks performed better in crop classification with an overall accuracy (OA) of 96.62% and Kappa of 0.95, compared to temporal statistical metrics (OA of 92.19% and Kappa of 0.88) and phenological features (OA of 90.87% and Kappa of 0.86). In addition, time series stacks outperformed temporal statistical metrics and phenological features for all individual crop types mapped in terms of user’s accuracy, producer’s accuracy and F1-score. Moreover, the effects of temporal density, interval and depth on time series stacks were analyzed. The analysis suggested that the optimal crop mapping results for time series stacks were achieved using the monthly composites of the combined Landsat-8 and Sentinel-2 data from March to October. Supplementary experiments conducted in two additional areas confirmed the consistency of the results from this study, thereby demonstrating the scalability of the methods used. This research provides valuable insights into spectrotemporal feature selection and optimization for accurate crop type mapping. And finally, a new web-based application named “Crop Mapper” was developed with Google Earth Engine to facilitate the availability of crop type maps derived from monthly gap-free Landsat Sentinel-2 time series for the areas once the training samples were available.