Efficient and Effective NDVI Time-Series Reconstruction by Combining Deep Learning and Tensor Completion

Abstract

Reconstruction of normalized difference vegetation index (NDVI) time series plays an imperative part in the inference of vegetation dynamics. However, it is challenging for the existing methods to achieve a good balance between accuracy and efficiency. In this article, we novelly combine deep learning with a high-precision spatiotemporal adaptive tensor completion (ST-Tensor) method and propose an end-to-end NDVI time-series reconstruction network (NIT-Net). The ST-Tensor method is first used to generate high-quality seamless NDVI data as the label data to construct sample pairs, along with the original degraded observations. A handcrafted time-series processing network is further employed for effective and rapid reconstruction of the NDVI time series. Considering the temporal continuity and spatial correlation of NDVI time-series data, we combine long short-term memory with a convolution (LSTM-Conv) structure and utilize residual learning and dense connection strategies to mine the spatiotemporal features in depth. Multidimensional gradient constraints are introduced in the loss function to retain critical information. The experiments conducted on moderate resolution imaging spectroradiometer NDVI data show that the NIT-Net framework is superior to most of the comparison methods. The mean correlation coefficient between the reconstruction results of NIT-Net and ST-Tensor can reach 0.9955, while NIT-Net achieves a more than 14 times speed-up on a CPU, compared with ST-Tensor, and a 115 times speed-up on a GPU, which fully demonstrates its efficient performance and great practical application value.

Keywords

• Deep learning
• long short-term memory (LSTM)-conv
• normalized difference vegetation index (NDVI) time-series reconstruction
• tensor completion