Remote sensing of vegetation trends: A review of methodological choices and sources of uncertainty

Abstract

Long-term satellite data provide up-to-date measurements of terrestrial ecosystems. To accurately estimate vegetation trends, it is essential to carefully consider the methodological details within the satellite time series data. This review aims to i) identify the main methods for trend estimation by focusing on their requirements, and ii) highlight the potential sources of uncertainty in remote sensing of vegetation trends. Results showed that 92% of the studies used a linear model (linear regression, Man-Kendall test and Theil-Sen slope estimator). Non-linear patterns were considered in 8% of the studies by using Piecewise Linear Regression (PLR) and Breaks For Additive Seasonal and Trend (BFAST). Although linear methods have several important advantages, they require a thorough understanding of their limitations. Particularly, when data do not meet the assumption of linearity, they oversimplify complex patterns across sub-periods and cannot detect potential abrupt and unimodal changes. Results indicated that utilizing of different methods and sensors could validate the results. However, uncertainties stemming from methodological choices and cross-sensor comparisons complicate the interpretation of findings, adding to the complexity of the topic. We conclude that bias in sampling and irregularity of satellite observations over space and time are the main sources of uncertainty in vegetation trends, which has been addressed in the literature using maximum value composites. However, selecting maximum values for a specified time of a year loses other temporal information, and it can also be sensitive to outliers, leading to incorrect value composites and bias in results. We emphasize the need for a deep understanding of the complexities in time series data as a necessity for accurately estimating trends through detailed methodological choices.