Generalized Fine-Resolution FPAR Estimation Using Google Earth Engine: Random Forest or Multiple Linear Regression

Accurate estimation of fine-resolution fraction of absorbed photosynthetically active radiation (FPAR) is urgently needed for modeling land surface processes at finer scales. While traditional methods can hardly balance universality, efficiency, and accuracy, methods using coarse-resolution products as a reference are promising for operational fine-resolution FPAR estimation. However, current methods confront major problems of underrepresentation of FPAR-reflectance relations within coarse-resolution FPAR products, particularly for densely vegetated areas. To overcome this limitation, this article has developed an enhanced scaling method that proposes an outlier removal procedure and a method weighting the selected samples and models FPAR through weighted multiple linear regression (MLR) between the coarse-resolution FPAR product and the aggregated fine-resolution surface reflectance. Meanwhile, a random forest regression (RFR) method has also been implemented for comparison. Both methods were particularly applied to Landsat 8 OLI and moderate resolution imaging spectroradiometer (MODIS) FPAR data on the Google earth engine. Their performance was tested on a regional scale for an entire year. The results of the enhanced scaling method were closer to the in situ measurements (RMSE = 0.058 and
R
²
= 0.768) and were more consistent with the MODIS FPAR (RMSE = 0.091 and
R
²
= 0.894) than those of the RFR, particularly over densely vegetated pixels. This indicates that a well-designed simple MLR-based method can outperform the more sophisticated RFR method. The enhanced scaling method is also less sensitive to the number of training samples than the RFR method. Moreover, both methods are insensitive to land cover maps, and their computation efficiency depends on the number of images to be estimated.