Abstract
In this paper, we present a new concept for predicting satellite-derived land surface temperature (LST) under cloudy skies over vegetated areas in the Alps. Although many different reconstruction methods have been developed, they require rarely available inputs, or they restore missing pixels from clear-sky observations with low spatial resolution (1-5 km), which makes them unreliable in heterogenous ecosystems. Given these limitations, we propose a station-based procedure to predict cloud-covered grids from 1-km Terra MODIS LST at 250 m spatial resolution. First, we explored correlations between ground-measured LST and air temperature in conjunction with other geo-biophysical variables under cloudy-sky conditions derived from ESRA clear-sky radiation model. Considering a high site dependency driven by different landcovers, in-situ data were aggregated into three groups (forest, permanent crops, grassland) and then, models were established. Next, the regressions were applied to 250-m gridded predictors to estimate cloud-covered LST pixels for six Terra MODIS LST images in 2014. While for permanent crops and forest group linear modelling was the most efficient, neural networks achieved the best performance for grasslands. The reconstructions showed reasonable LST distribution considering landscape heterogeneity of the region. The results were validated against timeseries of ground-measured LST in 2014. The models achieved reliable performance with an average R2 of 0.84 and RMSE of 2.12C. Despite some limitations, mainly due to diversified character of cloudy-sky conditions and high heterogeneity of gridded predictors, the method can effectively reconstruct overcast MODIS data at subpixel level, which shows great potential for producing cloud-free LSTs in complex ecosystems.