Learning the variations in annual spectral-temporal metrics to enhance the transferability of regression models for land cover fraction monitoring

Monitoring the Earth by annually mapping land cover (LC) fractions helps to better understand the ongoing processes and changes of land use and land management. At 10 to 30 m spatial resolution, the combination of time-series data aggregation, specifically spectral-temporal metrics (STM), and regression-based unmixing models has been shown to be highly effective in quantifying LC fractions over large areas. However, STM are subject to variations in data densities within and between years, which may lead to variations in prediction accuracies and limit the transferability of models through time. To better understand the influence of annual data density on multi-year monitoring, we systematically tested the accuracy and spatial-temporal transferability of regression models for LC fraction mapping. Additionally, we introduced a novel strategy, Random Observation Selection, in the STM generation to enhance model transferability and compared its results to those obtained from regular STM. We used STM from multi-spectral Sentinel-2 satellite data to estimate Impervious surface, Woody and Low vegetation at a regional scale in northern Germany and Poland for the years 2017 to 2022. The study period is characterized by substantial inter-annual variation in data density and well suited to test temporal transfer: in the extremely hot and dry years of 2018 and 2019 data density is above average and considered relatively high, while in the wetter years of 2017 and 2021 it is low. The mapping quality of models trained in single years varied considerably between years with high and low data density. Specifically, models trained in years with high data density performed well in that same year but poorly when transferred to years with lower density. Conversely, models trained in years with low data density demonstrated less of a decrease in accuracy when transferred. A multi-year model trained in all six years performs the best for each individual year. The use of Random Observation Selection improved the transferability of all models, particularly for those trained with data from a high data density year when transferred to years with lower data density. Here, the mean absolute error showed the highest relative improvement of 16%. In conclusion, STM proved useful for multi-year regression-based monitoring, but care must be taken when annual data density varies. Incorporating Random Observation Selection reduced this influence and improved the spatial-temporal transferabil