Filling gaps in MODIS NDVI data using hybrid multiple imputation–Machine learning and DINCAE techniques: Case study of the State of Hawaii

Abstract

Normalized difference vegetation index (NDVI) data are vital for monitoring vegetation dynamics and health. However, NDVI time-series data obtained via remote sensing often contain missing values due to factors such as cloud cover, snow, and hardware failures. To address this problem and fill gaps in NDVI data from the Moderate Resolution Imaging Spectroradiometer (MODIS), this study combines the multiple imputations by chained equations (MICE) model with three machine learning techniques: Knearest neighbor, multilayer perceptron (MLP), and boosted regression tree. Additionally, the data interpolating convolutional auto-encoder (DINCAE), a recently proposed imputation method, is employed for imputation and comparison. The performance of all these models is evaluated using MODIS NDVI data from Oahu, Hawaii for training and validation. Synthetic scenarios with gap sizes of 20 %, 40 %, 60 %, and 80 % are created to assess the models’ feasibility for each gap size. Furthermore, all models are tested using data from Hawaii Island and Maui. Results indicate that the MICE-MLP model achieves the highest accuracy in imputing missing NDVI values on Oahu, with root mean square error (RMSE) values of 0.1028, 0.1112, and 0.1224 for missing ratios of 20 %, 40 %, and 60 %, respectively. Similarly, MICE-MLP outperforms other models using Hawaii Island and Maui data at gap sizes below 80 %. While the DINCAE model demonstrates superior accuracy at an 80 % gap size, its computational speed is slower than MICE-MLP. Overall, the findings underscore the robustness and accuracy of the MICE-MLP model in imputing missing NDVI data, making it a reliable alternative to existing methods.