Integrating 250 m MODIS data in spectral unmixing for 500 m fractional vegetation cover estimation

•The 250 m MODIS data are integrated with the 500 m MODIS data for FVC estimation.•The 250 m features are stacked with the 500 m features for spectral unmixing.•The proposed method increases the accuracy of FVC estimation obviously.•The method has great potential for enhancing current 500 m FVC products. Fractional vegetation cover (FVC) is an important indicator to measure the potential for carbon neutrality in the terrestrial ecosystem. Spectral unmixing is a common choice for FVC estimation. To deal with the lack of endmembers and intra-class spectral variation, machine learning-based spectral unmixing methods have been developed, but the training samples were constructed based on the data acquired at a time different from the prediction time, which leads to great uncertainty in FVC estimation. Moreover, the used sample features may not cope with areas with strong heterogeneity, where vegetation cannot be reliably identified from the complex background. In this paper, for 500 m FVC estimation, the recently developed the spatio-temporal spectral unmixing model is applied to deal with the inconsistency between training and predicting data by extracting training samples at prediction time directly. Furthermore, more features from 250 m MODIS data are proposed to be integrated with the original 500 m MODIS data. Specifically, for sample construction, new features from 250 m bands (Red and NIR) are stacked with the original 500 m features. For model training, the random forest (RF) model is applied, which can handle high-dimensional data composed of features from different sources. The trained model is employed to predict the FVC of all 500 m MODIS pixels. The proposed method of incorporating 250 m data in 500 m FVC mapping was validated through experiments in five different areas. The results show that the accuracy of FVC mapping can be increased obviously by integrating 250 m features, and the training samples extracted at the prediction time are more reliable for modeling training. This paper has great potential for enhancing current 500 m FVC products at the global scale.