A Self-Supervised Learning Framework to Reconstruct High-Quality NDVI Time Series

It is crucial to reconstruct missing Normalized Difference Vegetation Index (NDVI) data caused by factors such as clouds. Deep learning has been applied to this task. However, existing approaches often struggle to balance the large-scale selection of training samples with the integrated utilization of spatio-temporal information. In this work, we present a novel NDVI time series reconstruction method based on self-supervised learning (referred to as "SEFVI") to further enhance the reconstruction accuracy and regional generalization. Extensive accumulated NDVI data globally is leveraged to construct self-supervised samples. A spatio-temporal reconstruction network is introduced, incorporating Transformer to model global dependencies across temporal tokens and long short-term memory with convolution (LSTM-Conv) to extract local spatial-temporal relationships. Experimental results demonstrate that the SEFVI surpasses the state-of-the-art methods for NDVI time series gap-filling in both accuracy and efficiency.