Estimating Spatial and Temporal Trends in Environmental Indices Based on Satellite Data: A Two-Step Approach

Brigitte Colin, Kerrie L. Mengersen2019 · DOI

TL;DR

This paper introduces a novel two-step approach to analyze long-term spatial and temporal trends in green vegetation cover using satellite data. By combining linear regression for trend extraction with Boosted Regression Trees (BRT) for spatial modeling, the method effectively captures complex spatio-temporal dynamics, offering valuable insights into environmental changes in semi-arid regions.

What Problem Does This Paper Address?

Monitoring changes in green vegetation cover is crucial for understanding environmental shifts, particularly in semi-arid regions vulnerable to climate variability. While satellite data offers a powerful tool for this, analyzing long-term spatial and temporal trends presents challenges. Traditional methods often struggle to capture the complex interplay between location and time-dependent changes. Existing approaches might focus on single points in time or fail to adequately model the spatial variations in these trends. This paper aims to bridge this gap by developing a method that can effectively estimate and visualize these spatio-temporal trends across large geographical areas over extended periods.

Methods

The core of this paper is a novel two-step approach:

  1. Linear Regression for Trend Extraction: For each grid cell in the satellite imagery, a simple linear regression model is applied to estimate the trend of green vegetation cover over time (years). The slope coefficient from this model represents the annual rate of change in green vegetation for that specific location. This step effectively translates the temporal dimension into a spatial variable (the slope coefficient).
  2. Boosted Regression Trees (BRT) for Spatial Modeling: The extracted slope coefficients (representing spatial trends) are then used as the response variable in a BRT model. The explanatory variables in this BRT model are the geographic coordinates (latitude and longitude) of each grid cell. BRT is a powerful, flexible machine learning technique that can handle complex non-linear relationships and interactions between variables, making it well-suited for modeling spatial patterns. This step allows the researchers to understand how the extracted temporal trends (slope coefficients) vary spatially across the landscape.

This combination of a linear model for trend extraction and a non-linear machine learning model for spatial prediction is the key innovation.

Datasets

The primary dataset used in this study is:

  • Fractional Cover (FCover) data derived from Landsat imagery: This dataset provides information on the fraction of green vegetation cover within each satellite pixel. The study utilized 30 years of this data (1987-2017) for a specific region in the Northern Territory, Australia.

The data was pre-processed by aggregating individual pixels into larger grid cells (100x100 pixels) to reduce computational load and create a manageable dataset of 5530 grid cells.

Key Insights and Ideas

  • Spatio-temporal trends are complex: The study highlights that green vegetation trends are not uniform across space and time. While a general linear trend can be observed over 30 years, finer-scale analyses reveal more nuanced patterns.
  • Geographic coordinates as powerful predictors: Latitude and longitude, acting as proxies for North-South and East-West gradients, are shown to be significant predictors of the spatial distribution of green vegetation trends. This suggests that large-scale environmental factors correlated with geography play a crucial role.
  • BRT's ability to capture spatial complexity: The BRT model effectively captures the spatial variation in the linear trends derived from the first step, demonstrating its utility in modeling complex environmental relationships.
  • Dominance of North-South gradient: In the overall analysis, the North-South gradient (latitude) was found to be slightly more influential than the East-West gradient (longitude) in explaining the spatial patterns of vegetation trends.
  • Surprising positive trends: Contrary to expectations of desertification in semi-arid regions, the study found that a significant majority (84.48%) of extracted slope coefficients showed neutral to slightly positive trends in green vegetation. This suggests localized factors or specific land management practices might be influencing vegetation cover positively in this particular study area.

Key Results

  • Good model fit: The BRT model demonstrated a strong linear relationship between observed and predicted slope coefficients, indicating good predictive performance. The Root Mean Square Error (RMSE) values for the test data were generally low across different scenarios (e.g., 0.1150 for the entire 30-year dataset).
  • Majority of positive trends: 84.48% of the extracted slope coefficients indicated a neutral to slightly positive trend in green vegetation over the 30-year period.
  • North-South gradient influence: The North-South gradient (latitude) was consistently more influential than the East-West gradient (longitude) in explaining the spatial patterns of vegetation trends across most analyses.
  • Decadal analyses showed similar trends: Splitting the data into three 10-year decades did not reveal significantly different temporal trends compared to the overall 30-year analysis, suggesting that the long-term linear trend is a dominant factor.

Limitations and Future Work

  • Linearity assumption: The initial linear regression step assumes monotonic linear trends, which may not capture abrupt changes or turning points in vegetation cover.
  • Limited covariates: The BRT model only used geographic coordinates as predictors. Future work could incorporate additional environmental covariates (e.g., climate data, soil types) to further enhance the model's explanatory power.
  • Location-specific findings: The study focused on a single semi-arid region. Further research is needed to assess if these findings are generalizable to other regions or if they are specific to the unique environmental conditions of the study site.
  • Potential for alternative methods: The authors suggest that other statistical machine learning methods (e.g., Generalized Additive Models, Random Forests, Support Vector Machines) could be explored as alternatives for either the trend extraction or spatial modeling steps.
  • Multivariate response: The current approach uses a univariate response (slope coefficient). Future work could explore a multivariate approach, using regression estimates and their associated standard errors or RMSE as inputs for the second step.

Summary updated

BibTeX
@misc{colin2019estimating,
  author = {Brigitte Colin and Kerrie L. Mengersen},
  title = {{Estimating Spatial and Temporal Trends in Environmental Indices Based on Satellite Data: A Two-Step Approach}},
  year = {2019},
  doi = {10.3390/s19020361},
}
Figure