Spatiotemporal imputation of MODIS Normalized Difference Vegetation Index (NDVI) data using machine learning techniques
Abstract
Normalized Difference Vegetation Index (NDVI) time series are essential for monitoring vegetation dynamics and health as well as land-atmosphere interaction at regional and global scales. However, the time series of remotely sensed NDVI observations often have missing values due to cloudiness, snow, hardware failures, etc. Therefore, gap-filling (or the so-called missing data imputation) should be conducted to reconstruct satellite-derived NDVI time series. In this study, four Machine Learning (ML) techniques namely, K-nearest neighbor (KNN), Multilayer perceptron (MLP), Boosted regression tree (BRT), and the state-of-the-art Data Interpolating Convolutional Auto-Encoder (DINCAE) methods are used to fill gaps in MODIS NDVI data over Oahu, Hawaii. The relative amount of missing data was varied from 0.1 to 0.3 to investigate the optimal hyper-parameters of the models and compare their performance for imputation of missing values. The results indicated that all ML models, except KNN, outperformed the DINCAE method. The BRT approach had the best performance among the four utilized models, but at a high computational cost, while KNN showed the worst performance. To further test the feasibility of calibrated ML models on Oahu, they are used to fill the gaps in NDVI data in two other islands, namely Hawaii Island and Maui, i.e., the ML approaches were not trained for these two islands and they were used to test the models. The results indicate that BRT was more accurate than the other models on both islands. Overall, the findings showed the superiority of the BRT over the MLP, DINCAE and KNN for spatiotemporal imputations of NDVI on all three islands, suggesting that the BRT model might be the best choice for other areas.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H22P1049B