SAR time series in forest research Biomass

Size: px
Start display at page:

Download "SAR time series in forest research Biomass"

Transcription

1 SAR time series in forest research Biomass Thuy Le Toan Centre D Etudes Spatiales de la Biosphere (CESBIO) Toulouse, France Thuy.Letoan@cesbio.cnes.fr

2 The research question on the global Carbon budget Emissions CO2 in atmosphere 8.3±0.4 GtC/yr 90% 4.3±0.1 GtC/yr 46% Absorbed by ocean 1.0±0.5 GtC/yr net flux + 26% 2.5±0.5 GtC/yr Global Carbon Project, %? 28% To close the budget 2.6±0.9 GtC/yr (not measured at all!) Absorbed by land

3 Forests play a major role in the carbon cycle Carbon sinks Photosynthesis uptakes CO 2 from the atmosphere and store it under the form of biomass Carbon sources The quantity of burned biomass is directly linked to the amount of CO2 released Despite our knowledge of the role of forests, we still have large uncertainties on the spatial distribution, the quantity and dynamics of forest carbon stocks, sources and sinks

4 Why using Synthetic SARs Aperture to observe Radars forests? to observe the world forests? Transmit and receive polarised waves (here Horizontal and Vertical) H V Penetrate into the forest cover

5 Questions to be addressed The contribution of SAR time series (and multitemporal, multi baselines ) data for : 1. Monitoring of disturbances 2. Above-ground-Biomass retrieval Focus on global observations of tropical forests, the most critical biome regarding the carbon cycle

6 What cause temporal variations in SAR measurements of forests? Vegetation properties scatterers size,shape scatterers orientation scatterers number density scatterers water content Ground properties soil moisture & surface roughness

7 What cause temporal variations in SAR measurements of forests? Vegetation properties scatterers size,shape scatterers orientation Natural changes scatterers number density scatterers water content o o o o Change with vegetation growth Diurnal and seasonal change Wind, rain, temperature effects Ground properties soil moisture & surface roughness o Rain, inundation effects o Freeze/thaw o Weathering effects o..

8 What cause temporal variations in SAR measurements of forests? Vegetation properties scatterers size,shape scatterers orientation Disturbances scatterers number density scatterers water content o o o o o Change in scatterers density (degradation) Change in scatterers orientation (wind throw) Changes in water content (fire) All scatterers removed (clear cutting) Ground properties soil moisture & surface roughness o o Soil surface with post deforestation debris

9 Which scatterers seen by the SARs? Pinus Nigra X-band λ= 3 cm C-band λ= 5 cm L-band λ= 27 cm P-band λ= 70 cm TerraSAR-X COSMO-SkyMed Sentinel-1 RADARSAT PALSAR(-2) NISAR (2021) BIOMASS (2021)

10 The scattering mechansisms ) Scattering from the crown 2) Crown-ground interaction 3) Scattering from the trunk 4) Multiple trunk-ground scatterin 5) Attenuated ground scattering 6) Scatterig from the ground

11 At C-band, the main scattering mechanisms in mature forests are from the tree crown 1 6 At C-babd 2 1) Scattering from the crown 2) Crown-ground interaction 6) Scattering from the ground

12 At P-band, all mechanisms are involved At P-band 2 1) Scattering from the crown 4) Multiple trunk-ground scatterin 2) Crown-ground interaction 5) Attenuated ground scattering 6) Scatterig from the ground

13 How it works Physical with forest background growing? Indicator: forest above-ground biomass 1. Experimental evidence: campaigns conducted in temperate, boreal and tropical forests since early 90s. 2. Physical modelling At L and P-band Model simulation P-band L-band Villard et al. Contrast between mature forest and bare ground is very high: > 15 db at P-band and >10 db at L band Sensitivity to AGB up to 100 t/ha at L-band, and > 300 t/ha at P-band

14 How it works with forest growing? At C band Observation and modelling results at the Landes forest Picard et al., 2002 Contrast between mature forest and bare ground is small, 1-3 db, depending on surface properties Sensitivity to AGB negligible or very low

15 Question: time series of Sentinel-1 can be used to detect forest disturbances? Clear cutting for rubber plantation in Cambodia and Vietnam Clear cutting In Peru Clear cutting in Brasil Indonesia

16

17 Here No cloud comes cover Sentinel-1 radar images: S1 over Vietnam 6 October 2014 Every 12 /6 days except few gaps

18 A CNES-CESBIO processing chain to handle the large amount of S1 data Generation of stacked time series from a set of downloaded pre-processed images T. Koleck & S. Mermoz PEPS: CNES platform that provides free access to data from the Sentinel satellites OTB Orfeo ToolBox: CNES open source for state-of-the-art remote sensing

19 A CNES-CESBIO processing chain to handle the large amount of S1 data Generation of stacked time series from a set of downloaded pre-processed images T. Koleck & S. Mermoz o Fast parallel processing of 10m resolution images o Automatically downloads, calibrates, orthorectifies and filters speckle noise o Multi-image filtering particularly adapted to dense time series o Images subset directly superposed to Sentinel-2 100x100 km 2 tiles

20 An important element is the multi-image filtering Original Images I1 I I M Filter Filtered images J 1 J J M Purpose of filter: (1) Preserve radiometry unbiased I k ( x, y) = J ( x, y) 1 k M (2) Minimise the variance of k J k Lopes & Bruniquel, 1997 Quegan & Yu, 2001

21 An important element is the multi-image filtering Non filtered 4 looks Spatial filter Spatial + multitemp. filter γ o HV -8 db 2 1 ENL=20 Mean = db SD = 1.58 db 4 ENL=4 Mean = db SD = 2.47 db 2 Filtered 20 looks 3 ENL=20 Mean = db SD = 1.21 db 3-20 db 4 1 ENL=4 Mean = db SD = 2.31 db 21

22 Multitemporal forest monitoring with Sentinel-1 Sentinel-1 acquisition segment Rubber plantation area near Dầu Tiếng (Bình Dương Province) 78 Sentinel-1 images (VH and VV) have been acquired in this area between October 2014 and 12 March 2017: Almost every 12 days until September 2016 Almost every 6 days after October 2016, when Sentinel-2B entered its operational phase

23 Multitemporal forest monitoring with Sentinel-1 Color composite image from VH backscatter R (3 Feb 2015), G (26 Aug 2015), B (17 Jan 2016)

24 Multitemporal forest monitoring with Sentinel Saison humide: juin à octobre Bouvet el al., 2016

25 γ 0 VH 78 dates are available between October 2014 and April Only every second image is shown here.

26 Forest/Non-Forest from γ 0 VH Forest in green Backscatter change in yellow Need time series for disturbance detection

27 Detection of the first date where change occurs Logging date o Using backscatter change detection (Mermoz and Le Toan, 2016) o Confirmation by backscatter variance after the disturbance (Bouvet et al.,)

28

29 Logging date map between October 2014 and April 2016

30 Possible detection at national scale (110km x 110km) Logging date map between Nov & Jan. 2017

31 Forest disturbances from Sentinel-1 In Cambodia Sentinel S. Mermoz et al., 2017

32 Forest disturbances from Sentinel-1 In Cambodia Sentinel

33 Disturbance detection performance depends strongly on the disturbed area - Vegetation type - Dimension of the disturbed area - Environment conditions: rain, seasonality - Post-disturbance state of the area - Topography Need to develop disturbance indicators other than backscatter change

34 Disturbance detection using ascending & descending orbits H2020 EOMONDIS project

35 ASCENDING orbit Viewing direction Satellite pass SAR shadowing

36 DESCENDING orbit SAR shadowing Viewing direction Satellite pass

37 Disturbance detection using shadowing effect Forest Logged area Forest Shadowing areas can be detected in the form of a quick decrease of backscatter that remains during a long period after the logging event. Bouvet et al., 2017, in prep.

38 Disturbance detection using ascending & descending orbits Shadows detected in Sentinel-1 ascending and descending orbits, in Peru

39 Use of SAR data to detect disturbances and to measure forest biomass C-band Sentinel-1 BIOMASS L-band ALOS, NISAR, SAOCOM.. NRT Disturbances AGB 500t/ha AGB 150t/ha

40 L-band ALOS PALSAR to estimate biomass up to 100 t/ha ESA DUE project 40

41 L-band ALOS PALSAR to estimate biomass up to 100 t/ha Bouvet et al., in review

42 For tropical forests biomass, no dedicated sensors The Congo basin AGB map at 1 km Saatchi et al. AGB 500 t/ha Fair agreement between existing biomass maps Based mainly on IceSAT/GLAS, MODIS and in situ data AGB map at 500 m Baccini et al. AGB map at 1 km Avitabile et al. 0 42

43 What SAR system to measure biomass of all forest biomes? Mapping forest biomass requires a radar sensor with long wavelength: 1. to penetrate the canopy in all forest biomes 2. to interact with woody vegetation elements 3. to enable repeat pass interferometry with a single satellite for advanced measurement techniques This implies a radar at P-band, of wavelength ~70 cm, the longest possible from space

44 The Biomass mission was proposed to measure forest biomass o ESA 7th Earth Explorer mission o To observe forests of different biomes but focus is on tropical forests, the most critical biome in terms of carbone budget o To be launched in 2021 Biomass Objective: To reduce uncertainties in the spatial distribution, the quantity and dynamics of forest carbon stocks, sources and sinks

45 BIOMASS will map global forest biomass and forest height Forest biomass Forest height Disturbances 1. The crucial information need is in the tropics: deforestation (~95% of the Land Use Change flux), regrowth (~50% of the global biomass sink) 2. Biomass accuracy of 20% at 4 hectares 3. Detection of deforestation at 0.25 ha 4. Repeated measurements over multiple years to identify deforestation and growth

46 PolSAR: past experimental results showed consistency in the SAR intensity-biomass relationship Biomass (t/ha) > 260 Le Toan et al., RSE 2011 Estimated biomass (t/ha) Retrieved biomass (ton/ha) Beaudoin et al., RMSE= 9.46 t/ha In situ biomass (t/ha) In-situ stand biomass (ton/ha) 0

47 POLSAR: For high biomass, topography (and other effects) obscure the SAR sensitivity to AGB Tropical forest, French Guiana Backscatter at single polarisation (HV) in db Paracou Nouragues r = 0.16 Paracou Nouragues In situ biomass (t.ha -1 ) Correction for topographic effects and scattering mechanisms using polarimetry and a DEM. Polarimetric biomass indicator (db) Paracou Nouragues r = Villard et al., 2014 In situ biomass (t.ha -1 )

48 Use of advanced technique to enhance biomass retrieval performance BIOMASS combines advanced techniques to enhance retrieval performances PolSAR (SAR Polarimetry) PolInSAR (Polarimetric SAR Interferometry) TomoSAR (SAR Tomography) z x z x z x o o o y y y

49 Use BIOMASS of advanced combines technique advanced to enhance biomass techniques retrieval to performance enhance retrieval performances PolSAR (SAR Polarimetry) PolInSAR (Polarimetric SAR Interferometry) TomoSAR (SAR Tomography) z x z x z x o o o y y y Conditions: high coherence with repeat pass interferometry

50 Tower based experiments to test long term variation of the SAR measurements o o 1. Tower-based radar observing a forest 2. Automatic and systematic acquisitions of fully polarimetric data (HH, HV, VH and VV) 3. Tomographic capability (to have a vertical discrimination of backscattering mechanisms) 4. Associated with in situ measurements Tropiscat o o o Experiment in French Guiana Measurements every 15 mn Started in 2011, end 2013 (with interruptions) Afriscat Experimentna in Ghana 2 x 3 hours per day: 4:30-7:30 am/pm o Started on 20/07/2015, end 4/05/2017

51 Measure of long term coherence for repeat pass interferometry P-band L-band γ pq = K K k= 1 i min k= 1 i min i max i max S S k t1 pq k t1pq ( i) ( i). S 2 K k t 2 pq i max k= 1 i min ( i) S * k t 2 pq ( i) 2 Hamadi et al., 2017

52 Measure of long term coherence for Biomass repeat pass interferometry TropiScat (French Guiana) AfriScat (Ghana)

53 Diurnal cycle in intensity and coherence Intensity : cycle linked to dielectric constant wrt tree functioning Hamadi et al., 2015 Coherence : cycle linked to tree functioning and covective wind

54 Pol-InSAR provides the canopy height estimates Interferometry provides height information The measured height depends on polarisation PolInSAR retrieves canopy height using models Key requirement: HV VV High temporal correlation between two images, hence P-band is essential HH

55 PolInSAR provides the canopy height estimates Hajnsek et al., IEEE GRS Peat Swamp forest, Mawas, Kalimantan 0m Height (m)

56 TomoSAR provides 3 D images of the forests Tomography generates images of different forest layers from multiorbit SAR acquisitions SAR resolution cell SAR Tomography resolution cell Enables Isolation of: Canopy Ground Topography Tebaldini et al. Height (m) z o x y Normalised backscatter intensity

57 TomoSAR provides 3 D images of the forests Boreal forest Kryclan, Sweden z x o y Tropical forest Paracou, French Guiana

58 SAR Tomography provides high accuracy biomass maps Ground range [pixel] Biomass map obtained by inversion power layer 30m (t.ha -1 ) Azimuth [pixel] BIOMASS map at 50 m, by tomography Ho Tong et al., 2015, 2016

59 Consistent measurements in French Guiana and in Gabon Height (m) Normalised 0.2 backscatter 0.1 intensity 0 HV intensity of TomoSAR layers at 15, 20,.., 35, 40 m Red: French Guiana (2009) Black: Gabon (2015) 5 0 Layer 15m, r = 0.47, RSMD = 78% P 0 Layer 20m, r = 0.72, RSMD = 43% P 0 Layer 25m, r = 0.9, RSMD = 19% P HV [db] -20 HV [db] -20 HV [db] Best fit line in-situ Lope in-situ 4 Paracou Layer 30m, r = 0.97, RSMD = 9% P 0 Layer 35m, r = 0.96, RSMD = 10% P 0 Layer 40m, r = 0.9, RSMD = 18% P HV [db] -20 HV [db] -20 HV [db] Above-ground biomass (100 t/ha) Above-ground biomass (100 t/ha) Above-ground biomass (100 t/ha)

60 Summary 1. The dynamic nature of SAR measurements of forested areas; changes occur at different time scales 2. Understanding these changes will contribute to develop robust methods for forest observation o Detection of disturbances o Biomass measurements 3. Complementarity with other SAR and optical sysptems to be exploited