Combined Use of SAR and Optical Time Series Data towards Near Real-Time Forest Disturbance Mapping

Transcription

1 Background Image Source bbc.co.uk Human Planet 2011 BBC Manuela Hirschmugl, Janik Deutscher, Karl-Heinz Gutjahr, Carina Sobe, Mathias Schardt Joanneum Research Earth Observation Services for Monitoring Dynamic Forest Disturbances Combined Use of SAR and Optical Time Series Data towards Near Real-Time Forest Disturbance Mapping Coordinated by: Partners: Supported by: This project has received funding from the European Union s Horizon 2020 research and innovation programme under grant agreement No

2 1) Background & Aims Tropical forest monitoring from optical data is often hindered by cloud cover, specifically when a near realtime monitoring is requested. SAR data alone is able to capture deforestation patches, but has shortcomings in reliably detecting the specific extend of smallstructured logging being considered as forest degradation. Therefore the aim is to develop algorithms to optimize the combination of dense time series by using multitemporal optical and RADAR data in detecting near real time disturbances in forest Multitemp 2017, Page 2

3 2) Overall system Multitemp 2017, Page 3

4 New optical Image - preprocessed New optical Image - preprocessed Step repeated Preprocessed optical time series data Disturbance indicator incl. probability Disturbance indicator incl. probability F/NF Class If needed: postprocessing Status Product F/NF Probability combi. NRT disturbance product Probability combi. New NRT disturbance product New Status Product F/NF Disturbance indicator incl. probability Disturbance indicator incl. probability Step repeated Preprocessed SAR time series data New SAR image - preprocessed New SAR image - preprocessed

5 Input data S2 image for validation Rapideye image for validation / /2016 Optical data (12) for F/NF Optical data (6) for disturbance map Optical disturbance map 10/2016 F/NF Map Combined disturbance map 12/ /2016 SAR data (10) for disturbance map Timeline SAR disturbance map Multitemp 2017, Page 5

6 3) Testsite Peru Very dynamic area with many disturbance drivers like agriculture (small, but also palm oil plantations), mining, logging, settlement expansion. High cloud cover probability High mountains in the western part Multitemp 2017, Page 8

7 4) Methods - Optical processing Pre-processing Geometric adjustment necessary to join S2 (S1) and Landsat8 in one time series by automated image matching Atmospheric correction necessary (currently Sen2Cor used) (Radiometric adjustment and topographic normalization depending on test site) Cloud detection based on Sen2Cor scene classification plus post-processing Multitemp 2017, Page 9

8 Forest disturbance mapping Which indicators are best suited? Feature selection exercise NDVI is classified into forest disturbance probability based on ground truth data bands Feature R² red 0,6846 NIR 0,0579 SWIR1 0,3606 SWIR2 0,5418 red, NIR 0,7033 red, SWIR1 0,6869 red, SWIR2 0,6849 NIR, SWIR1 0,6487 NIR, SWIR2 0,694 SWIR1, SWIR2 0,7132 NDVI 0,7364 Multitemp 2017, Page 10 indices NDII5 0,6301 NDII7 0,6953

9 Features in the time series Black = cloud shadow in the time series Great advantage of NDVI! Multitemp 2017, Page 11

10 4) Methods SAR processing Main steps: Efficient pre-processing (Multi-looking, gamma nought correction, multi-temporal & ev. Speckle filter) Calculation of coefficient of variation and backscatter trend Classification of probabilities to be forest disturbance by applying ground truth-derived thresholds Multitemp 2017, Page 12

11 4) Methods SAR processing Details given in the presentation held on Tuesday Paper online: Deutscher et al Multitemp 2017, Page 13

12 4) Combination of SAR and optical result Combination is based on a Bayesian approach (probability combination) Inputs: each result with probability plus a reliablity value generated from an accuracy assessment Optional: use of a reliability map to define regions of higher/lower reliability of the data types (e.g. mountains for SAR data; data quality flags on optical) Multitemp 2017, Page 14

13 5) Results: optical timeseries Landsat S2 image Forest/nonforest Red possible map disturbance Blue disturbance patterns confirmed (2 of 4) Multitemp 2017, Page 15

14 NRT disturbance mapping - optical S2 Landsat Blue Confirmed old disturbances disturbances Red possible disturbances Pink disturbance patterns confirmed (2 of 4) Multitemp 2017, Page 16

15 NRT disturbance mapping - optical S Blue, pink old disturdisturbances bances Red possible disturbances Multitemp 2017, Page 17

16 5) Results: Validation info Rapideye Gray: nonforest Blue: visually digitized forest disturbances Multitemp 2017, Page 18

17 5) Results: SAR only S1 probability map Probability: Multitemp 2017, Page 19

18 5) Results: optical only S2/L8 probability map Probability: Multitemp 2017, Page 20

19 5) Results: combined Combined probability map Orange: Combined classified map Blue: reference Probability: Multitemp 2017, Page 21

20 5) Results Plot-based Plot size nr. of plots % detected S-1 only % detected optical data only % detected S-1 & optical combined > 1 ha % 96.4 % % 0.5 ha -1 ha % 91.3 % % 0.2 ha % 94.3 % 94.3 % ha < 0.2 ha % 86.4 % 95.5 % all plots % 92.6 % 97.2 % Area-based ha ha 67.1 ha ha 43.6 % 79.9 % 83.7 % Multitemp 2017, Page 22

21 6) Conclusions & Outlook Combination of SAR and optical time series results gives an added value! Commission error has still to be evaluated with additional/other data Transfer to other areas will be tested Near realtime component will be further improved Multitemp 2017, Page 23

22 Thank you for your attention The presented work has received funding from the European Union s Horizon 2020 research and innovation programme under grant agreement No (Project EOMonDis) and from the Austrian Research Promotion Agency FFG under grant agreement No (Project LCX-SAR). Multitemp 2017, Page 24