Reconstructing systematically missing NDVI time series in cropland: A GAN-based approach using optical and SAR data

Abstract

Normalized Difference Vegetation Index (NDVI) time series data is essential for monitoring cropland dynamics and assessing crop conditions. However, these data often suffer from large-scale systematic missing patterns due to atmospheric variations and satellite revisit cycles, significantly compromising monitoring accuracy, particularly for capturing rapid surface changes. Existing methods primarily concentrate on recovering cloud-covered data, often overlooking systematic data gaps. To address this limitation, we propose a Periodic Imputation Generative Adversarial Networks (PIGAN) model to reconstruct large-scale systematic missing NDVI remote sensing data. The model integrates optical and synthetic aperture radar (SAR) data as inputs and employs a Generative Adversarial Networks (GAN) to impute NDVI missing values. Specifically, Pearson correlation coefficients and Random Forest (RF) algorithms are utilized to select vegetation-sensitive indices as inputs for the generator. The generator employs a dual-stream architecture and ConvLinBlock to accommodate dual-source data inputs and effectively handle extensive missing patterns. The discriminator transforms the traditional task of distinguishing real from fake data by evaluating the proportion of the real data within fixed intervals using dilated convolutions, thereby addressing the systematic missing patterns in time series. Experimental results demonstrate that the proposed method outperforms existing models across different crop types and varying environmental conditions, achieving over 10% improvement in widely used metrics such as RMSE and MAE. Furthermore, the model exhibits superior performance in NDVI spatial–temporal recovery, highlighting its potential for practical applications. The PIGAN code is publicly available at https://github.com/hydai-00/PIGAN.