IGFUL: An Integration Gap-Filling, Unmixing, and Learning Approach for Estimating Continuous High-Resolution Landsat Normalized Difference Vegetation Index Time Series

The Normalized Difference Vegetation Index (NDVI) is the most popular index used in remote sensing for monitoring vegetation dynamics. While numerous moderate-resolution NDVI products are available, there is an increasing demand for fine-resolution, temporally continuous NDVI time series, particularly for monitoring vegetation dynamics in areas with high surface heterogeneity. A major challenge in producing such datasets is the irregular Landsat data acquisition, often hampered by extensive cloud contamination. This study introduces an Integration Gap-Filling, Unmixing and Learning (IGFUL) approach for creating continuous 30-m, 8-day NDVI time series by integrating Landsat surface reflectance with 250-m GLASS NDVI data. The workflow comprises four stages: (1) constructing the first Random Forest (RF) model using cloud-free Landsat surface reflectance as input and GLASS NDVI as output to obtain physically consistent Landsat clear-sky NDVI; (2) for cloudy pixels on a target day, identifying similar pixels based on seasonal similarity curves and utilizing their spatio-temporal correlation information for pre-filling missing Landsat NDVI values; (3) identifying reference images using Structural Similarity Index (SSIM), calculating temporal and spatial predictions through unmixing and Thin Plate Spline (TPS) methods respectively, and developing the second RF model using temporal prediction, spatial prediction, GLASS NDVI on the target day, and Landsat NDVI and GLASS NDVI on the base day as inputs, with Landsat NDVI on the target day as output; (4) weighting multiple initial predictions from the second RF model to generate the final Landsat NDVI reconstruction results. IGFUL was evaluated under varying cloud conditions and outperformed three benchmark methods in both accuracy and efficiency. It achieved high accuracy at the CIA site (RMSE = 0.013, CC = 0.983) and remained robust at the cloud-heavy Qionglai site (RMSE = 0.033, CC = 0.882). IGFUL's ability to reduce reconstruction errors while preserving phenological and spatiotemporal consistency underscores its potential for generating high-resolution and temporally continuous NDVI time series in challenging remote sensing environments.