Effectiveness of gap-filling in tropical tree canopy cover modeling

Satellite-based land surface phenology information (e.g. multitemporal spectral metrics and seasonal composites) can be used to model tropical tree canopy cover, but missing observations caused by clouds and cloud shadows limits their applications. A recently proposed gap-filling method named Missing Observation Prediction Based on Spectral-Temporal Metrics (MOPSTM) has shown advantages in predicting observations in Landsat time series, which has potential for solving the limitations of Landsat images in tree canopy cover modeling. This gap-filling method models the relationship between valid observations in time series and spectral-temporal metrics in the k-Nearest Neighbor regression to predict the missing observations. To evaluate the effectiveness of gap-filled images in times series with respect to tree canopy cover modeling, we compared the tree canopy cover modeling performance based on Landsat images that have used gap-filling and have not used gap-filling. The study area was a heterogeneous Afromontane landscape in Taita Hills, Kenya. Landsat Operational Land Imager 8 Collection 1 Level-2 Surface Reflectance images acquired in 2015 were pre-processed, including clouds and cloud shadows removing, using Fmask, and then they were gap-filled by MOPSTM. The reference tree canopy cover data were derived from the Airborne Laser Scanning (ALS) data. We generated seasonal composites (i.e. median composites) from original images and gap-filled images in four seasons respectively and fit them in the random forest regression. The results indicated that tree canopy cover modeling using gap-filled seasonal composites yielded smaller root mean square error than that using the original seasonal composites, which proved the effectiveness of gap-filling in tropical tree canopy cover modeling. In terms of seasonal results, the improvement of tree canopy cover modeling by gap-filling is different for four seasonal composites. To be specific, seasonal composites derived from images in higher quality (more valid observations) tended to show a slight improvement by using gap-filling while those derived from images in lower quality tended to show a large improvement by using gap-filling. This indicates the necessity of gap-filling applied to tree canopy cover modeling in cloud-prone areas.