A land-cover-assisted super-resolution model for retrospective reconstruction of MODIS-like NDVI data across the continental United States by blending Landcover300m and GIMMS NDVI3g data

Abstract

Long-term archives of remote sensing data hold values for identifying temporal changes occurring on the land surface. Moderate-spatial-resolution data acquired by sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) have proven useful in large-scale studies. The absence of such data prior to the launch of MODIS in 2000 necessitates the retrospective reconstruction of MODIS-like datasets. While data fusion techniques are capable of generating spatiotemporally continuous data, challenges remain in capturing interannual variation of land surface dynamics at a spatial resolution where real observation data are lacking. This study introduces a novel deep learning-based model, termed the Land-Cover-assisted Super-Resolution SpatioTemporal Fusion model (LCSRSTF), designed to produce biweekly 500-meter MODIS-like data spanning from 1992 to 2010 across the Continental United States (CONUS). LCSRSTF integrates Landcover300m and the Global Inventory Modelling and Mapping Studies (GIMMS) NDVI3g data. The model exacts moderate-resolution class features from annual Landcover300m data at the target year, incorporates GIMMS NDVI3g time series to capture seasonal fluctuations, and employs the Long Short-Term Memory (LSTM) method to mitigate sensor differences. Evaluation against observed MODIS images confirms the robustness of our model in generating MODIS-like data across CONUS. The root mean square error (RMSE) of the model results is 0.094 from 2001 to 2010, while that of GIMMS NDVI3g data is 0.154. The linear regression coefficient for the model simulation is 0.872, compared to 0.844 for GIMMS data. The model exhibits reasonable predictive capabilities in reconstructing retrospective data when assessed using Landsat data prior to 2000. The developed method as well as the MODIS-like dataset spanning from 1992 to 2010 across CONUS hold the promise in extending the temporal span of moderate-spatial-resolution data, thereby facilitating comprehensive long-term studies of land surface dynamics.