Generating high-resolution (10-m) vegetation greenness fraction with Sentinel-2, SMAP imagery, and machine learning

Vegetation Greenness Fraction (VGF) describes the proportion of green vegetation over an area. High resolution VGF data is widely used in many aspects. One of its applications is to monitor the vegetation growth, and crop yields, which provide valuable insight for agriculture activity (Tao et al., 2019).

Conventionally, VGF are commonly calculated with the "relative vegetation abundance algorithm" by leveraging vegetation index (VI) derived from remotely sensed data. Simply put, this algorithm calculates VGF as the ratio between the instaneous and the maximal VIs relative to the bare-soil VI. However, this algorithm requires accurate maximal and bare-soil VIs to give us a skillful VGF estimation, which is quite challenging since they can vary in different geographic regions, and vegetation types (Gao et al., 2020). The level 4 product of the Soil Moisture Active Passive (SMAP) satellite mission, SPL4SMGP, provides the VGF data. However, its 9-km spatial resolution is coarse and may not be fine enough for practical use.

In this Colab script, I demonstrated a machine learning-based approach to generate high-resolution (10-m) VGF from the Sentinel-2 (S2) imagery by leveraging the Google Earth Engine. Specifically, I used randomly sampled data from spatiotemporally coincident historical S2 Normalized Difference Vegetation Index (NDVI) and SMAP VGF from 2019 to 2021 as training data to train a random forest regression model. With this model, any updated high-resolution (10-m) S2 NDVI can be used as input to generate the corresponding high-resolution (10-m) VGF.

Original 9-km resolution SMAP VGF

Sentinel-2 high-resolution (10-m) VGF estimated with machine learning over the same area at the same time as the above figure