Performance of vegetation indices from Landsat time series in deforestation monitoring

https://doi.org/10.1016/j.jag.2016.06.020Get rights and content

Highlights

  • Data fusion of vegetation index can improve mapping performance, particularly in complex ecosystems.

  • The Normalized Difference Fraction Index is most suitable for mapping deforestation across ecosystems while the Normalized Difference Vegetation Index performs comparatively low.

  • For Landsat time series based deforestation mapping wetness related vegetation indices outperform greenness related indices.

  • Mapping performance sensitivity to observation frequency varies by used vegetation index.

Abstract

The performance of Landsat time series (LTS) of eight vegetation indices (VIs) was assessed for monitoring deforestation across the tropics. Three sites were selected based on differing remote sensing observation frequencies, deforestation drivers and environmental factors. The LTS of each VI was analysed using the Breaks For Additive Season and Trend (BFAST) Monitor method to identify deforestation. A robust reference database was used to evaluate the performance regarding spatial accuracy, sensitivity to observation frequency and combined use of multiple VIs. The canopy cover sensitive Normalized Difference Fraction Index (NDFI) was the most accurate. Among those tested, wetness related VIs (Normalized Difference Moisture Index (NDMI) and the Tasselled Cap wetness (TCw)) were spatially more accurate than greenness related VIs (Normalized Difference Vegetation Index (NDVI) and Tasselled Cap greenness (TCg)). When VIs were fused on feature level, spatial accuracy was improved and overestimation of change reduced. NDVI and NDFI produced the most robust results when observation frequency varies.

Introduction

Between 2000 and 2012, global forest loss increased by approximately 2000 km2 per year (Hansen et al., 2013). Deforestation contributed to around 8% of anthropogenic carbon emissions in the 2000 s (Tubiello et al., 2015), and despite forests remaining a sink, emissions from degradation were 0.80 Gt CO2 yr−1 between 1990 and 2015 (Federici et al., 2015). Mechanisms such as Reducing Emissions from Deforestation and Forest Degradation (REDD+) aim to reduce forest loss and increase carbon sequestration in forests (UNFCCC, 2016). Monitoring, Reporting and Verification (MRV) of REDD+ carbon stock changes is mandatory and requires consistent and long term monitoring of forests supported by field observations (Arino et al., 2012), (De Sy et al., 2012). The Landsat mission can be a key component of such MRV methodologies, as it provides long term medium resolution (10–60 m) remote sensing data (Gutman et al., 2008, Skole and Tucker, 1993, Townshend and Justice, 1988). The availability of free Landsat data (Woodcock et al., 2008) created a paradigm change in how Landsat data is used, away from bi-temporal analysis towards time series analysis (Hansen and Loveland, 2012). In addition, an increase in computational capacities (Evangelidis et al., 2014) resulted in forest change maps of unprecedented scale and resolution (Hansen et al., 2013). Time series analysis methods applied to Landsat data were inspired by previous developments of coarse spatial resolution systems such as the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Advanced Very High Resolution Radiometer (AVHRR) (Cihlar et al., 1997, Roerink et al., 2000, Verbesselt et al., 2010, Forkel et al., 2013, de Jong et al., 2012, Tucker et al., 2005). Compared to those missions, Landsat offers the longest running cross-calibrated globally consistent record of the Earth’s surface at medium resolution. Since the opening of the archive, many studies demonstrated Landsat’s capabilities for mapping forest cover and related changes (Hansen and Loveland, 2012), (Wulder et al., 2012, Roy et al., 2014, Loveland and Dwyer, 2012, Woodcock and Ozdogan, 2012, Trenberth et al., 2013, Pflugmacher et al., 2012) with an increasing density of Landsat time series (LTS) (Achard et al., 2014, Huang et al., 2010, Zhu et al., 2012, Broich et al., Apr. 2011, Cohen et al., 2010). Methods developed for temperate forests can often be characterized by a higher frequency of observations than in tropical areas which are characterized by persistent cloud cover (Romijn et al., May 2012). When large areas are under investigation, methods have to cope with very different observation frequencies, often requiring complex solutions (Broich et al., Apr. 2011). Empirical studies show strong correlations between vegetation indices (VIs), and vegetation parameters such as biomass and canopy closure (Avitabile et al., 2012, Myneni and Hall, 1995). VIs are useful for assessing the amount and condition of vegetation, while suppressing noise, soil background and atmospheric effects (Jackson and Huete, 1991). VIs have become a standard for the interpretation of vegetation dynamics such as deforestation, and have been applied on LTS (Forkel et al., 2013, de Jong et al., 2012, Verbesselt et al., 2012). In this paper, we compare the suitability of different VIs for mapping deforestation when using the Breaks For Additive Season and Trend (BFAST) Monitor algorithm. BFAST recently emerged as a reliable tool to detect ecosystem disturbances such as droughts, fires and vegetation changes (Verbesselt et al., 2010, Verbesselt et al., 2012, Hutchinson et al., 2015, Watts and Laffan, 2014) in agricultural (Atzberger, 2013) and forested landscapes (Schmidt et al., 2015, Lambert et al., 2013, Lambert et al., 2015). Moreover, BFAST Monitor proved its robustness when applied to more infrequent time series such as Landsat, Landsat – SAR fused series or Landsat – MODIS fused series (DeVries et al., 2015a, Reiche et al., 2015, Dutrieux et al., 2015, Hamunyela et al., 2016, DeVries et al., 2015b). The observation frequency of a pixel’s time series determines the ability of BFAST Monitor to describe the time series, and therefore affects its ability to detect deforestation. More observations can describe seasonality and deforestation with a higher temporal resolution and thus tend to be more accurate (Schultz et al., 2015, Schultz et al., 2013). So far, the algorithm has only been applied to the Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDMI) and Normalized Difference Vegetation Index (NDVI) time series. The performance of BFAST Monitor when using other VIs has not yet been assessed. The goal of this study was to identify which VIs can detect deforestation in the tropics better when applying BFAST Monitor on LTS. Given the constraint of highly varying observation frequencies in the tropics we test the capacity of VIs to produce robust and consistent results while varying observation frequency. Since different VIs might have different success rates depending on the type of deforestation mapped, our study covers three areas in the tropics characterized by different forest change dynamics. In addition, we tested whether data fusion of multiple maps at feature level can provide increased mapping accuracy. Ensemble classification has proved useful in combining various mapping outcomes of various inputs to produce one highly accurate map, and among those classification algorithms, random forest is the most prominent (Pal, 2005, Ceamanos et al., 2010, Gislason et al., 2006). We addressed the following objectives:

  • Identify the most spatially accurate VI for deforestation mapping when applying BFAST Monitor to LTS

  • Understand the VI’s spatial accuracy regarding its sensitivity towards observation frequency per site (Brazil, Ethiopia, Vietnam)

  • Explore the potential of feature level data fusion of VIs to complement each other and increase accuracy

To address these objectives, three sites in Brazil, Ethiopia and Vietnam were investigated for recent forest changes (2010–2013) using a reference database.

Section snippets

Material and methods

Fig. 1 outlines the study. First, LTS were created in three sites (Brazil, Ethiopia, Vietnam) (Section 2.1), which were selected based on their differing ecosystem characteristics, frequency of observations, and deforestation types. Landsat data processing included screening of each image for clouds and their shadows as well as non-forest masking (Section 2.2). Eight VIs were computed for the time series of each site, the EVI, Global Environment Monitoring Index (GEMI), Normalized Difference

Detection of deforestation in 3 sites by 8 VIs

Fig. 2 showed representative examples of time series for the same pixel per site for each VI. The figure demonstrated the complexity of deforestation monitoring. For example in Brazil, BFAST Monitor captured a drop of the vegetation signal within the historic period (2005) of a similar magnitude as the 2011 deforestation signal. However, it demonstrated the ability of BFAST Monitor to account for disturbances occurring in the historic period, while still being sensitive to changes occurring

Discussion

We assessed the performance of VIs for tropical deforestation mapping using BFAST Monitor on LTS. Within this section research objectives as previously stated, were answered and discussed successively. More general discussion points were provided at the end of the section.

Conclusion

This paper compared VIs’ performance of LTS for deforestation monitoring using BFAST Monitor. VIs were tested in different sites across the tropics and accuracy statistics based on reference data were produced, for each VIs, its fusion suitability and its sensitivity to observation frequency. A canopy cover sensitive VI (NDFI) performed better than wetness related VIs (NDMI, TCw) and greenness related VIs (EVI, GEMI, NDVI, SAVI, TCg) performed the worst. For tropical deforestation monitoring,

Acknowledgments

This work was funded by the German federal ministry of science and education (BMBF) LUCCI project and the research was carried out within the facilities of Wageningen University. Thanks go to black bridge company and the planet action organization for providing high resolution satellite data. We thank the USGS for making the Landsat data available and we thank the entire remote sensing community for their hard work of creating a better tomorrow using state of the art technologies.

References (87)

  • L.P. Dutrieux et al.

    Monitoring forest cover loss using multiple data streams, a case study of a tropical dry forest in Bolivia

    ISPRS J. Photogramm. Remote Sens.

    (2015)
  • K. Evangelidis et al.

    Geospatial services in the Cloud

    Comput. Geosci.

    (2014)
  • S. Federici et al.

    New estimates of CO2 forest emissions and removals: 1990–2015

    For. Ecol. Manag.

    (2015)
  • G.M. Foody

    Status of land cover classification accuracy assessment

    Remote Sens. Environ.

    (2002)
  • P.O. Gislason et al.

    Random Forests for land cover classification

    Pattern Recognit. Lett.

    (2006)
  • N.R. Goodwin et al.

    Estimation of insect infestation dynamics using a temporal sequence of Landsat data

    Remote Sens. Environ.

    (2008)
  • V.H. Gutiérrez-Vélez et al.

    Annual multi-resolution detection of land cover conversion to oil palm in the Peruvian Amazon

    Remote Sens. Environ.

    (2013)
  • E. Hamunyela et al.

    Using spatial context to improve early detection of deforestation from Landsat time series

    Remote Sens. Environ.

    (2016)
  • M.C. Hansen et al.

    A review of large area monitoring of land cover change using Landsat data

    Remote Sens. Environ.

    (2012)
  • D.J. Hayes et al.

    Estimating proportional change in forest cover as a continuous variable from multi-year MODIS data

    Remote Sens. Environ.

    (2008)
  • C. Huang et al.

    An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks

    Remote Sens. Environ.

    (2010)
  • A. Huete et al.

    Development of vegetation and soil indices for MODIS-EOS

    Remote Sens. Environ.

    (1994)
  • A.R. Huete et al.

    A comparison of vegetation indices over a global set of TM images for EOS-MODIS

    Remote Sens. Environ.

    (1997)
  • A. Huete et al.

    Overview of the radiometric and biophysical performance of the MODIS vegetation indices

    Remote Sens. Environ.

    (2002)
  • A.R. Huete

    A soil-adjusted vegetation index (SAVI)

    Remote Sens. Environ.

    (1988)
  • J.M.S. Hutchinson et al.

    Monitoring vegetation change and dynamics on U.S. Army training lands using satellite image time series analysis

    J. Environ. Manag.

    (2015)
  • R.D. Jackson et al.

    Interpreting vegetation indices

    Prev. Vet. Med.

    (1991)
  • S. Jin et al.

    Comparison of time series tasseled cap wetness and the normalized difference moisture index in detecting forest disturbances

    Remote Sens. Environ.

    (2005)
  • T.R. Loveland et al.

    Landsat: building a strong future

    Remote Sens. Environ.

    (2012)
  • J.G. Masek et al.

    North American forest disturbance mapped from a decadal Landsat record

    Remote Sens. Environ.

    (2008)
  • D.C. Morton et al.

    Mapping canopy damage from understory fires in Amazon forests using annual time series of Landsat and MODIS data

    Remote Sens. Environ.

    (2011)
  • P. Olofsson et al.

    Good practices for estimating area and assessing accuracy of land change

    Remote Sens. Environ.

    (2014)
  • D. Pflugmacher et al.

    Using Landsat-derived disturbance history (1972–2010) to predict current forest structure

    Remote Sens. Environ.

    (2012)
  • J. Reiche et al.

    Fusing Landsat and SAR time series to detect deforestation in the tropics

    Remote Sens. Environ.

    (2015)
  • E. Romijn et al.

    Assessing capacities of non-Annex I countries for national forest monitoring in the context of REDD+

    Environ. Sci. Policy

    (May 2012)
  • D.P. Roy et al.

    Landsat-8: science and product vision for terrestrial global change research

    Remote Sens. Environ.

    (2014)
  • M. Schmidt et al.

    Multi-resolution time series imagery for forest disturbance and regrowth monitoring in Queensland, Australia

    Remote Sens. Environ.

    (2015)
  • C. Song et al.

    Classification and change detection using landsat TM data: when and how to correct atmospheric effects?

    Remote Sens. Environ.

    (2001)
  • C.M. Souza et al.

    Combining spectral and spatial information to map canopy damage from selective logging and forest fires

    Remote Sens. Environ.

    (2005)
  • C.J. Tucker

    Red and photographic infrared linear combinations for monitoring vegetation

    Remote Sens. Environ.

    (1979)
  • J. Verbesselt et al.

    Phenological change detection while accounting for abrupt and gradual trends in satellite image time series

    Remote Sens. Environ.

    (2010)
  • J. Verbesselt et al.

    Near real-time disturbance detection using satellite image time series

    Remote Sens. Environ.

    (2012)
  • L.M. Watts et al.

    Effectiveness of the BFAST algorithm for detecting vegetation response patterns in a semi-arid region

    Remote Sens. Environ.

    (2014)
  • Cited by (151)

    • Not-so-random forests: Comparing voting and decision tree ensembles for characterizing partial harvest events

      2023, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text