towards improved deforestation monitoring by fusing ...€¦ · towards improved deforestation...

Towards Improved Deforestation Monitoring by Fusing Optical and Synthetic Aperture Radar Time Series

Chris HoldenJanuary 11, 2017

Background

2

● Forests:○ Biodiversity○ Material resource○ Regulatory ecosystem services○ Carbon pool and potential sink

http://www.nasa.gov/topics/nasalife/features/globe-workshop.htmlHansen, et al. 2013. High-Resolution Global Maps of 21st-Century Forest Cover Change

● Satellite remote sensing○ Relatively long (~40 years)

record with global coverage○ Landsat: “locally meaningful”

spatial resolution

Landsat Limitations

3

Wulder, et al. 2016. The global Landsat archive: Status, consolidation, and direction.

Landsat Limitations

4

Wulder, et al. 2015. Virtual constellations for global terrestrial monitoring

Landsat and AWIFS Landsat, AWIFS, and CBERS

Medium Resolution Record

5Reiche, et al. 2016. Combining satellite data for better tropical forest monitoring

“Fusion”Wikipedia:

“Sensor fusion is combining of sensory data or data derived from disparate sources such that the resulting information has less uncertainty than would be possible when these sources were used individually.”

Research Objectives:Develop and assess methods of fusing Landsat and PALSAR time series that,

1. Increase accuracy of forest cover change detection2. Decrease time to detect forest cover change3. (Perhaps) Improve land cover characterization thematic detail 6

Fusion Methods

7

Fusion: “Residual Monitoring”● Create predictive model explaining time series observations and monitor for

change in characteristic of forecast residuals● Lineage:

○ Verbesselt et al., 2014: “BFAST Monitor”■ Recursive MOSUM

○ Brooks et al., 2014: “On-the-Fly Massively Multitemporal Change Detection”■ Exponentially Weighted Moving Average (EWMA)

○ Zhu and Woodcock, 2014: Continuous Change Detection and Classification■ Scaled residuals heuristics

● Fit different models based on characteristics of each data source○ Landsat: y ~ 1 + time + harmonic○ Radar: y ~ 1 + time

8

Fusion: “Residual Monitoring”

9

NDMI L-HH

Fusion: “Residual Monitoring”

10

NDMI

MODEL

FORECASTRESIDUAL

L-HH

MODEL

FORECASTRESIDUAL

Fusion: “Residual Monitoring”

11

EWMA CHANGE DATA

NDMI

MODEL

FORECASTRESIDUAL

L-HH

MODEL

FORECASTRESIDUAL

Fusion: “Probability Monitoring”● Create classification model predicting “forest likelihood” and monitor for

change in characteristic of probabilities● Lineage:

○ Huang, et al. 2010: Vegetation Continuous Tracker■ Monitor forest z-scores

○ Solberg and Huseby, 2008: MODIS and ENVISAT ASAR for snow state monitoring■ Hidden Markov Model

○ Reiche, et al. 2015: “A Bayesian Approach…”

● Gather training data and estimate classification models for each data source

12

Does it matter?

13

Simulated other observing conditions

1. Randomly choose some percentage of Landsat data to throw out

2. Run algorithm on reduced Landsat time series with and without ALOS observations

3. Compare number of instances in each simulation that we found change correctly

Chocó, Colombia?

Congo?

SE Asia?

In Progress… Yurimaguas, Loreto Region, Peru

14

15

16

Questions - Poster Session!

17

Backup Slides

18

Fusion: “Probability Monitoring”

19

EWMA CHANGE DATA

FORESTLIKELIHOOD

ALOSFEATURES

FORESTLIKELIHOOD

LANDSATFEATURES

CLASSIFIER CLASSIFIER

Fusion: “Probability Monitoring”

20

EWMA CHANGE DATA

FORESTLIKELIHOOD

ALOSFEATURES

FORESTLIKELIHOOD

LANDSATFEATURES

CLASSIFIER CLASSIFIER

Fusion: “Probability Monitoring”

21

EWMA CHANGE DATA

FORESTLIKELIHOOD

ALOSFEATURES

FORESTLIKELIHOOD

LANDSATFEATURES

CLASSIFIER CLASSIFIER