Crop NDVI time series construction by fusing Sentinel-1, Sentinel-2, and environmental data with an ensemble-based framework

Satellite-based Normalized Difference Vegetation Index (NDVI) time series data are conducive for near-real-time (NRT) monitoring of crop progress, but they commonly suffer from data gaps due to cloudy weather conditions. Fortunately, Synthetic Aperture Radar (SAR) data, which are cloud-insensitive, has high potentials to fill these data gaps. However, the relationship between NDVI and SAR data is not generic and affected by various factors (e.g., crop canopy structure and phenology, soil exposure and humidity). Therefore, it is worth exploring to evaluate the relationships and develop a more robust SAR-NDVI fusion method for NDVI time series construction. In this study, we proposed and evaluated an ensemble-based data fusion framework that accounts for these factors, based on which dense NDVI time series for crop monitoring were constructed. The framework consisted of three steps: (1) NDVI was predicted from Sentinel-1 SAR data and auxiliary environmental factors using a random forest (RF) model; (2) an improved feature importance measurement method was proposed to reveal the different contributions of multisource data to the modeling, and suggestions for selecting optimal input parameters were presented; (3) the uncertainty of the predicted data was quantified by an ensemble-based method and incorporated into a Weighted Least Squares (WLS) method to construct dense NDVI time series. Results showed that the RF models were generally improved by the auxiliary data and achieved a satisfactory accuracy of NDVI estimation (R2 > 0.93, RMSE < 0.075) for both corn and soybean crops in Southwestern Ontario, Canada. The proposed method performed well in filling data gaps in both the vegetative and reproductive stages for corn and soybean, providing a practical and promising solution for continuous crop monitoring.