Tropical Forest Monitoring Using Landsat Time Series and Community-based Data

Ben DeVries

NASA Goddard Space Flight Center, May 12, 2016

Co-authors, Collaborators, Project Partners, Thesis Students

    Jan Verbesselt, Lammert Kooistra, Martin Herold, Arun Kumar Pratihast, Mathieu Decuyper, Achim Zeileis, Shijo Joseph, Valerio Avitabile, Loïc Dutrieux, Johannes Reiche, Niki De Sy, Brice Mora, Nandika Tsendbazar, Frehiwot Tessema, Dereje Beyene, Elias Gebremeskel, Kalkidan Mulatu, Svane Bender Kaphengst, Daniela Tunger, Bianca Schlegel, Mesfin Tekle, Muluken Mekuria, Kafa Zone Bureau of Agriculture

About me

Born and raised in St. Thomas, ON, Canada

BSc: University of Toronto - 2005

MSc: Wageningen University - 2010

PhD (Remote Sensing): Wageningen University - 2015

Towards Near Daily Monitoring of Inundated Areas Over North America Through Multi-Source Fusion of Optical and Radar Data

Chengquan Huang (UMD), Megan Lang (FWS)

John Jones (USGS), Irena Creed (UWO)

Monitoring tropical forest dynamics using Landsat time series and community-based data


        Martin Herold


        Jan Verbesselt

        Lammert Kooistra

    Available online


Remarks on open data, tools and methods

Introduction to bfast and bfastmonitor

Monitoring small-scale disturbances: Kafa, Ethiopia

Monitoring post-disturbance regrowth: Madre de Dios, Peru

Integrating community-based monitoring data and LTS to characterize forest changes


Open data / methods

    Landsat legacy represents longest continual space-borne observation record

    Opening of the archive in 2008 - paradigm shift in change monitoring

    Streamlining L1 data generation (LEDAPS, FMASK, ...)

    Bi-temporal (e.g. MAD, post-classification), Annual (e.g. LandTrendR, VCT), "All-available" (e.g. CCDC, bfastmonitor)

..."openness" not just limited to Landsat...

    How can these data and tools be leveraged to address current problems?

    e.g. REDD+ MRV - transparency and reproducibility are key to maintaining credibility



    Breaks For Additive Season and Trend

    Decomposition of time series into harmonic, trend and noise (residual) components

    Segment a time series based on "most likely" breakpoints

    e.g. - Detecting phenological change due to fire regimes in forests

    Verbesselt et al. (2010), Remote Sensing of Environment.



    Breakpoints are determined by some measure of deviance from "expected" values

    e.g. Cumulative Sums (CUSUM), or more localized Moving Sums (MOSUM) of residuals

    Normalized by expected deviance (noise) in history data

    Econometrics: measure of "structural change"

    Chu et al. (1995), Biometrika. DOI:10.1093/biomet/82.3.603

Small-scale disturbances

Application of bfastmonitor to dense Landsat time series

Kafa Biosphere Reserve
Southwestern Ethiopia

    UNESCO Biosphere Reserve - inaugurated in 2011

    Moist Afro-montane cloud forests

    One of few remaining natural coffea arabica habitats

    Forest conservation and restoration project by NABU, Germany

    Participatory Forest Management (PFM) and Community-Based Monitoring (CBM)

Small-scale changes can be captured by breakpoints

    Despite sparse time series, small-scale changes are mapped using breakpoints

    Data gaps -- temporal uncertainty

    Do all breakpoints signal disturbance? What about magnitude?

Change magnitude (independent of breakpoint) measured as median of residuals in each 1-year interval

Visualizing change dynamics using breakpoints and magnitude

    Calibration and validation based on visual interpretation of LTS data

    The best model for disturbances considers breakpoints and magnitude (median of residuals)

    Gaps still exist, especially for degradation

    Limitations include index, cal/val data (more on this later)


R package for pre-processing LTS and applying bfastmonitor

Post-disturbance Regrowth

What happens after a disturbance?

  • Permanent land use change (forest → cropland)
  • Secondary forest regrowth (forestry, shifting agriculture)

How well can we track post-disturbance regrowth with LTS?

Madre de Dios, Peru

    Southeastern Peru

    Dense canopy forests

    Recent infrastructure development leading to rapid changes to forests

    Very dynamic: forest loss often followed by land abandonment, secondary regrowth

    "Gold rush" within the last decade a major driver of forest change

MOSUM test for regrowth

    Recall: MOSUM is a measure of deviation from a "stable" history model

    At and after a disturbance, the MOSUM become very large

    Forest regrowth following a disturbance should bring the MOSUM back to near-zero conditions

    Different MOSUM bandwidths (h) affect its sensitivity to this process

MOSUM test with added constraints

    w - minimum time between disturbance (tB) and regrowth (tR)

    s - minimum time in which MOSUM < critical boundary


R package for monitoring regrowth with LTS

Integrated Monitoring:


Integrated Monitoring:


    Community-Based Monitoring (CBM) is seen as a key component to REDD+ Measuring, Reporting and Verification (MRV) systems

    Increased engagement, ownership, benefits-sharing

    Some research has shown potential use of CBM data for monitoring purposes (e.g. AGB measurements), but research on utility of data is lacking

ODK -- Open Data Kit

Simple smart-phone based tool for field data collection

Exploits GPS, camera, and interactive features of smart phones to make data collection easy and convenient

Kafa Biosphere Reserve (revisited)

    30 forest rangers recruited from 10 local districts by Kafa Zone Bureau of Agriculture

    Reforestation, community plantations, community outreach, extension services, monitoring activities

    After monitoring trainings (incl. ODK tools), >1500 geo-located forms returned

    Prototype for local forest monitoring system - potential for nesting in national FMS

    Various attributes describe landscape, changes (if any), drivers of change, etc.

    Geo-location and photo evidence are essential for QA

    Pratihast et al. (2014), Forests. DOI: 10.3390/f5102464

    Single robust linear model fit to entire time series

    bfast used determine presence/absence of breakpoint in time series

    Temporal covariates derived for each spectral band and index:

          - trend (+delta), amplitude (+delta), overall trend, overall intercept

    Training (personnel) and data collection took place in phases

    Do the predictions improve over time as monitoring continues?

    Degradation often confused for no-change by model in the training phase

    Improvement of degradation predictions in operational phase

    Increase in number of training data as well as improved quality of data

    Band/Index-specific importance scores

    Normalized ranks (compared class accuracies) over many model fits

    SWIR[-derived] bands are essential

Local experts acting as "in situ sensors"

Ground-based, a priori identification of subtle change can enhance forest monitoring using LTS

          - e.g. understorey degradation precedes subtle changes to the canopy

Continual capacity building can pay dividends in local forest monitoring systems

Drawbacks: how representative and sustainable are CBM data?

Data and Code are available

Take home messages

Landsat time series is a very powerful tool

Change monitoring sometimes limited by arbitrary class definitions

          - e.g. 'deforestation' versus 'degradation', 'regrowth' versus 'no regrowth'

          - consider more 'objective ecological' measures of forest condition and dynamics (Sexton et al., 2016)

          - integration of LTS with in situ data/sensors can be of great help here

Community-based monitoring is a potentially cost-saving measure in monitoring systems

Some useful tools


Further Reading

Even More Reading

Thank you!