Phenology-Based Residual Trend Analysis of MODIS-NDVI Time Series for Assessing Human-Induced Land Degradation

Chris H.q. Ding (丁宏强), Hao Chen, Xiangnan Liu, Chao Ding, Fang HuangSensors 2018 · DOI

TL;DR

This paper introduces a novel "phenology-based residual trend analysis" (P-RESTREND) method to improve the accuracy of monitoring human-induced land degradation. By incorporating the timing of vegetation growth (phenology) and considering precipitation before the growing season, P-RESTREND more effectively distinguishes degradation caused by human activities from natural climate variability compared to existing methods.

What Problem Does This Paper Address?

Land degradation is a critical global issue, threatening ecosystem services and human well-being. Accurately monitoring and understanding its drivers, particularly distinguishing between climate-driven changes and human impacts, is challenging. Traditional methods like the Residual Trend Analysis (RESTREND) often oversimplify the growing season and ignore the lagged effects of precipitation, leading to potential misinterpretations of land degradation. This paper aims to address this by developing a more nuanced approach that accounts for the dynamic nature of vegetation growth and its relationship with precipitation.

Methods

The core innovation is the Phenology-Based Residual Trend Analysis (P-RESTREND). This method builds upon the standard RESTREND by:

  • Phenological Signal Detection: Instead of assuming a fixed growing season, P-RESTREND uses the Normalized Difference Water Index (NDWI) time series to dynamically identify the start (SOS) and end (EOS) of the growing season for each pixel. This is crucial because vegetation growth periods vary significantly across space and time.
  • Incorporating Pre-Growing Season Precipitation: The model acknowledges that rainfall before the main growing season can influence vegetation productivity. P-RESTREND includes precipitation from the month preceding the identified growing season in its analysis.
  • Bivariate Linear Regression: A bivariate linear regression is used to model the relationship between accumulated NDVI (representing vegetation productivity) and precipitation (both during the growing season and pre-growing season). This allows for a more accurate estimation of expected NDVI based on climatic factors.
  • Residual Trend Analysis: The difference between the observed NDVI and the predicted NDVI (the residual) is then analyzed for trends over time. A negative trend in these residuals indicates land degradation that cannot be explained by precipitation alone, thus pointing to human-induced causes.

Datasets

  • MODIS MOD09A1 Surface Reflectance Data: 8-day composite reflectance data from Terra MODIS (2000-2015) was used to derive NDVI and NDWI. This dataset is atmospherically corrected and provides a 500m spatial resolution.
  • Daily Precipitation Data: Downloaded from the Meteorological Data Center of the China Meteorological Administration (2000-2015). This data was resampled to a 500m spatial resolution to match the MODIS data.
  • MODIS Land Cover Type Product (MCD12Q1): Used to identify and retain only grassland areas that remained unchanged between 2001-2013 for the study area.
  • High-Resolution Remote Sensing Images (Google Earth): Used for visual verification of "missed pixels" – areas identified as degraded by P-RESTREND but not by standard RESTREND.

Key Insights and Ideas

  • Phenology is Key: The timing of vegetation growth is not a static parameter. Dynamically identifying the growing season for each pixel is essential for accurately assessing land degradation, as fixed growing seasons can lead to misinterpretations.
  • Pre-Growing Season Matters: Rainfall prior to the main growing season has a significant, albeit lagged, influence on vegetation productivity. Ignoring this can lead to overattributing changes to human activity when they are, in part, climatically driven.
  • P-RESTREND is a Better Discriminator: The P-RESTREND method demonstrates a superior ability to differentiate between land degradation caused by climate variability and that caused by human activities. This is a critical step towards targeted land management.
  • "Missed Pixels" Highlight P-RESTREND's Value: The identification of "missed pixels" – areas showing degradation signs in P-RESTREND but not in standard RESTREND – and their subsequent visual confirmation through high-resolution imagery strongly supports the improved sensitivity of the new method.

Key Results

  • Improved VPR Significance: P-RESTREND achieved a higher mean R² (0.45) in the vegetation-precipitation relationship (VPR) compared to standard RESTREND (0.38) and a pixel-by-pixel RESTREND approach (0.35). This indicates a better fit of the model to the data, suggesting more accurate removal of precipitation influence.
  • Higher Proportion of Significant Trends: Approximately 75% of pixels met the significance criteria for VPR analysis using P-RESTREND, compared to 61% for standard RESTREND. This suggests P-RESTREND is more robust in identifying meaningful trends.
  • Detection of Human-Induced Degradation: P-RESTREND identified 2.44% of pixels with significant negative trends (degradation), while standard RESTREND identified only 1.89%. Crucially, P-RESTREND also identified a higher proportion of significant positive trends (recovery) at 27.37% compared to RESTREND's 25.68%.
  • Effective Identification of "Missed Pixels": Visual interpretation of high-resolution imagery confirmed that many of the "missed pixels" identified by P-RESTREND indeed showed clear signs of human-induced land degradation, such as desertification and loss of grassy areas.

Limitations and Future Work

  • Phenology Extraction Sensitivity: The accuracy of P-RESTREND relies on the accurate extraction of the growing season. Any errors in phenological signal detection could adversely affect the entire technique. Future work could focus on developing more robust and practical methods for phenology detection.
  • Linear VPR Model: The study used a linear model for the vegetation-precipitation relationship. Nonlinear responses between precipitation and NDVI, which can occur in arid ecosystems, were not fully considered and could lead to less accurate results in such scenarios.
  • Abnormal VPR in Arid Ecosystems: In arid ecosystems experiencing major structural changes, abnormal VPR might occur, and P-RESTREND may not perform optimally. Investigating these abnormal changes is a direction for future research.
  • Verification Scale: The visual verification method used for "missed pixels" was applied at a small, pixel-level scale and cannot be considered a comprehensive regional validation tool.

Summary updated

BibTeX
@article{2018phenology,
  author = {Chris H.Q. Ding (丁宏强) and Hao Chen and Xiangnan Liu and Chao Ding and Fang Huang},
  title = {{Phenology-Based Residual Trend Analysis of MODIS-NDVI Time Series for Assessing Human-Induced Land Degradation}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18113676},
}
Figure