Fusing MODIS and AVHRR products to generate a global 1-km continuous NDVI time series covering four decades

Xiaobin Guan, Huanfeng Shen, Yuchen Wang, Dong Chu, Xinghua Li, Linwei Yue, Xinxin Liu, Liangpei Zhang2021 · DOI

TL;DR

This paper presents a novel method to fuse data from the long-term AVHRR and higher-resolution MODIS satellite records to create a continuous, 1-km global NDVI time series spanning four decades (1982-present). This new dataset addresses the critical trade-off between temporal coverage and spatial resolution in existing NDVI products, offering a valuable resource for global vegetation studies.

What Problem Does This Paper Address?

Monitoring global vegetation health and its changes over time is crucial for understanding terrestrial ecosystems and their response to climate change. Satellite-derived Normalized Difference Vegetation Index (NDVI) is a key metric for this. However, existing NDVI products present a dilemma:

  • AVHRR (Advanced Very High-Resolution Radiometer): Provides a very long time series (since the early 1980s) but suffers from coarse spatial resolution (around 1-8 km) and lower data quality. This limits its utility for detailed regional studies.
  • MODIS (Moderate Resolution Imaging Spectroradiometer): Offers higher spatial resolution (1 km) and better data quality but has a shorter record (since 2000).

This gap means that studying vegetation changes over multiple decades at a fine spatial scale has been challenging. The authors aim to bridge this gap by creating a 1-km resolution NDVI time series that combines the long temporal coverage of AVHRR with the superior spatial characteristics of MODIS.

Methods

The authors developed a multi-step spatio-temporal fusion framework to achieve their goal:

  1. Temporal Filtering: An improved trend filtering method (combining L1-norm and Whittaker filtering) was applied to denoise both the AVHRR (GIMMS 3g) and MODIS time series, addressing noise and errors caused by clouds and atmospheric contamination.
  2. Pixel-by-Pixel Normalization: To account for sensor differences, the AVHRR NDVI data was normalized to the MODIS data level. This was done using a least-squares linear regression model applied independently for each pixel, using overlapping data from 2000-2014. This ensures the fused product has MODIS-like radiometric characteristics.
  3. Spatio-Temporal Fusion: A non-local means filter-based fusion algorithm was employed to predict fine-resolution NDVI data for the AVHRR period (pre-2000). This method leverages both AVHRR and MODIS data, along with reference images, to predict NDVI at MODIS spatial resolution. A key innovation is the calculation of the spatial weighting function, which considers spatial relationships between pixels.

Datasets

  • AVHRR GIMMS 3g NDVI: Used for the long-term temporal coverage (July 1981 - December 2014).
  • MODIS MOD13A3 NDVI: Used for its higher spatial resolution and data quality (January 2000 - December 2014).
  • Limited AVHRR LAC/HRPT data: Used for validation purposes, particularly for the pre-2000 period.
  • Global Land Cover Map (MCD12Q1): Used to validate the product's performance across different vegetation types.
  • GLAP (Global Land 1-KM AVHRR Project) data: Used for validation of the pre-2000 fused product.
  • ANCUS (1-km AVHRR NDVI composites) data: Used for temporal consistency evaluation over the conterminous U.S.

Key Insights and Ideas

  • Bridging the Resolution-Time Gap: The core idea is that by carefully fusing data from different sensors with complementary strengths, we can overcome the limitations of individual datasets and create a more comprehensive record.
  • Normalization is Key: Simply combining data isn't enough; normalizing the older AVHRR data to match MODIS's radiometric characteristics is crucial for a consistent time series.
  • Spatio-Temporal Fusion for Detail: The spatio-temporal fusion technique is essential for upscaling the AVHRR data to MODIS's finer resolution, revealing more detailed spatial patterns that were previously obscured.
  • Equatorial Regions Remain Challenging: Despite the advancements, areas with persistent cloud cover and high rainfall (like the equator) continue to pose challenges for accurate remote sensing, impacting the consistency of both original and fused data.
  • Fusion Enhances Data Quality: The fusion process not only increases spatial resolution but also significantly improves data quality by reducing noise and inconsistencies present in the original AVHRR data.

Key Results

  • A 1-km global continuous NDVI time series (STFLNDVI) from 1982 to the present has been successfully generated.
  • Simulated experiments and real-data assessments confirm the high accuracy of the STFLNDVI product in terms of spatial distribution and temporal variation.
  • The fused product shows a similar global spatial distribution to the original AVHRR data but with significantly higher spatial resolution and richer details.
  • Correlation coefficients (r) between the fused data and MODIS are generally high (e.g., >0.87 for growing season, >0.88 for off-season), with low mean absolute differences (MAD) and bias.
  • The STFLNDVI product demonstrates good consistency with true AVHRR observations (GLAP and ANCUS), with comparable spatial distributions and high temporal consistency.
  • The fusion process significantly reduced the number of pixels with high MAD and bias values compared to the original AVHRR data.

Limitations and Future Work

  • Dependence on Original Data Consistency: The accuracy of the fused product is inherently limited by the consistency between the original AVHRR and MODIS data. Regions with naturally poor relationships (e.g., equatorial areas due to cloud cover) still show some unsatisfactory results.
  • AVHRR Data Quality: The temporal dynamics for the pre-2000 period are derived solely from AVHRR data, which has known quality issues. While the fusion process mitigates some of these, the underlying limitations of AVHRR remain.
  • Cumulative Uncertainties: The multi-step processing (filtering, normalization, fusion) can introduce cumulative uncertainties.
  • Underutilized Overlapping Information: The current method could potentially improve by better utilizing the overlapping time series information between AVHRR and MODIS to reduce cumulative errors.
  • Future Work: The authors suggest developing new processing methods that better integrate temporal and spatial information to further improve accuracy. They also note that if future satellite data remains consistent with MODIS, it can be directly added to the STFLNDVI product to extend the time series even further.

Summary updated

Abstract

Satellite normalized difference vegetation index (NDVI) time-series data are an essential data source for numerous ecological and environmental applications. Although various long-term global NDVI products have been produced with different characteristics over the past decades, there is still an apparent trade-off between the spatiotemporal resolution and time coverage. The Advanced Very High-Resolution Radiometer (AVHRR) instrument can provide the only continuous time series with the longest time coverage since the early 1980s, but with the drawback of a coarse spatial resolution and poor data quality compared to the observations of later instruments. To address this issue, a spatio-temporal fusion-based long-term NDVI product (STFLNDVI) since 1982 was generated in this study, with a 1-km spatial resolution and a monthly temporal resolution. A multi-step processing fusion framework was employed to combine the superior characteristics of Moderate Resolution Imaging Spectroradiometer (MODIS) and AVHRR products, respectively. Simulated and real-data assessments both confirm the ideal accuracy of the fusion result with regard to the spatial distribution and temporal variation. Only a few relatively unsatisfactory results are found due to the poor relationship between the original AVHRR and MODIS data. The evaluations also show that the proposed fusion framework can obtain stable results similar to MODIS data in different years and seasons, even when the temporal distance between the fusion data and the reference data is large. We believe that the STFLNDVI product will be of great significance to characterize the spatial patterns and long-term variations of global vegetation. The NDVI product is available at DOI: http://doi.org/10.5281/zenodo.4734593 (Guan et al., 2021).
BibTeX
@misc{guan2021fusing,
  author = {Xiaobin Guan and Huanfeng Shen and Yuchen Wang and Dong Chu and Xinghua Li and Linwei Yue and Xinxin Liu and Liangpei Zhang},
  title = {{Fusing MODIS and AVHRR products to generate a global 1-km continuous NDVI time series covering four decades}},
  year = {2021},
  doi = {10.5194/essd-2021-156},
}
Figure