Continuous monitoring of forest disturbance using all available Landsat imagery

Transcription

1 Continuous monitoring of forest disturbance using all available Landsat imagery Zhe Zhu, Curtis E. Woodcock, and Pontus Olofsson Dept. of Geography and Environment, Boston University 675 Commonwealth Avenue, Boston, MA 02215, USA Corresponding authors

2 Study area includes commercial timber land, agriculture and urban areas Study area (subset of November 23 rd 2002 Landsat ETM+ image Band 432 composites)

3 Steps of Continuous Monitoring of Forest Disturbance Algorithm (CMFDA) 1. Remove clouds and cloud shadows 2. Estimate surface reflectance model from past observations (LEDAPS for atmospheric correction) 3. Predict surface reflectance assuming no land cover change (forest clearing) 4. Define a forest mask 5. Compare predicted and observed images to find forest disturbance

4 Study area (all available Landsat ETM+ image from 2001 to 2002 with cloud cover less than 90%)

5 Two-step cloud/cloud shadow screening Step one global algorithm : Single-date based cloud and cloud shadow masking (Fmask freely available) Step two pixel-based algorithm : Multitemporal cloud and cloud shadow masking based using a robust fitting of TOA reflectances of the clear pixels clouds and shadows are ephemeral outliers

6 Results after initial screening and then multitemporal results Illustration of the two-step cloud, cloud shadow, and snow masking results. Left image shows a small piece of a Landsat image (shown with Bands 4, 3, and 2 in red, green, and blue). Middle image shows the results of the Fmask algorithm. Clouds are yellow and shadows are blue. Right image shows the results after use of the multi-temporal approach. Notice that the cloud and cloud shadow missed in Fmask were found in the multi-temporal approach.

7 Estimate time series models for Landsat surface reflectances Where, x = Julian date N = number of years T = number of days per year (=365) = coefficient for overall reflectance = coefficients capture the changes over i th year = coefficients for bimodal variations of surface reflectance each year.

8 Predicting surface reflectances Take away the coefficients for capturing interannual changes

9

10 Visual comparison of the observed and predicted surface reflectance for Landsat images

11

12 Two algorithms tested Two algorithms have been developed: Single-date differencing: Multi-date differencing:

13 Reference map A total of 21 rectangular areas, each with width and length larger than 3 km, were carefully interpreted to determine precisely the location and timing of forest disturbances. The reference rectangles were divided into two groups: one group used for training CMFDA and one group used for evaluating CMFDA accuracy. The reference rectangles were sorted and ranked by size and the odd number ranked rectangles (in blue) were used for help training CMFDA, and the even number ranked rectangles (in red) were used for evaluating CMFDA

14 Reference Map when and where for forest clearing

15 1. Spatial accuracies Results of single-date differencing: Testing on unseen data Reference data Single-date differencing Forest disturbance Others Total User s (%) Forest disturbance Others Total Producer s (%) Overall (%) Temporal accuracy = 90% This table shows the confusion matrix for the accuracy assessment of the single-date differencing algorithm. The overall accuracy results are not terribly revealing, as after excluding the edges of the change polygons, the change pixels left are only about 3% of the total interpreted pixels.

16 1. Spatial accuracies Results of multi-date differencing Testing unseen data Reference data Single-date differencing Forest disturbance Others Total User s (%) Forest disturbance Others Total Producer s (%) Overall (%) Temporal accuracy = 94% This table shows the confusion matrix for the accuracy assessment of the single-date differencing algorithm. The overall accuracy results are not terribly revealing, as after excluding the edges of the change polygons, the change pixels left are only about 3% of the total interpreted pixels.

17 Map of location and timing of forest clearing (over reference areas)

18 Zoom in of reference and accuracy maps. The colors show different types of errors.

19 Examples of continuous monitoring of forest disturbance using all available Landsat images in 2003

20 Discussion and conclusions (Part I) CMFDA is accurate in detecting forest disturbance both spatially and temporally, with producer s and user s accuracies above 95% and the temporal accuracy of approximately 94%. The continuous character of the monitoring makes the algorithm capable of indentifying disturbance soon after Landsat observations are available. It can be transferred to monitoring other land cover changes by applying different land cover masks. The SLC-off problem in Landsat 7 is not nearly as significant for CMFDA as compared with more conventional approaches. The expected time to find probable change and change in CMFDA is still too long to monitor changes as they are occurring. To achieve goal of global near-real time monitoring of land cover change, using more Landsat-like sensors or fusion with higher temporal frequency sensors like MODIS are choices in future studies.

21 Continuous land cover classification and change detection We processed a total of 532 Landsat TM/ETM+ images for path 12 row 31 at Eastern Massachusetts started from 1982 to Land cover changes are detected continuously using all spectral bands. The type of land cover for each curve are classified using the parameters of the time series model.

22 Community Goals? Reconstruct the history of the surface of Earth Provide maps of surface characteristics at any time (where maps change between dates correspond to locations of land cover change!!) Monitor change as it is occurring management relevant In interesting ways, there are part of the same process!

23 Study Area Path/Row =12/31

24

25 Landsat surface reflectance Band 4 surface reflectance (X10 4 ) Land cover change map Land cover classification map C B A A B C

26

27 Conclusions: Comments and Lessons for Sentinel-2 Better use of the time domain improves consistency, accuracy, timeliness and thematic detail in land cover data Dependent on ability to easily analyze free images (Landsat L1T format, easy conversion to surface reflectance via LEDAPS, automated cloud and shadow screening) Images with lots of clouds are useful as they also have lots of clear observations (getting roughly 25% of observations from images with >25% cloud cover) To extend the Landsat time series with Sentinel, we need to able to overlay images at the pixel scale! Historical observations are extremely valuable, particularly when processed consistently through time (please help the USGS effort to build a consistent dataset of all existing Landsat images)