Burnt area detection using two post fire images 1. Istituto Nazionale di Geofisica e Vulcanologia, Italy [email protected], [email protected]
2. The University of Manchester, UK 3. Istituto di Fisica Applicata Nello Carrara
Background
• .
Stefania Amici 1 2, Leonardo Santurri3, Alessandro Piscini 1
•The dynamic of open fire phenomena makes remote sensing from satellite a good tool for both
fire detection and burn area assessment.
• A multi-temporal approach relies on pre and post fire multispectral images to derive vegetation
maps (e.g. Normalized Differential Vegetation Index-NDVI
•While Mediterranean countries have an high likehood of cloud free satellite data in VIS-SWIR the
north European countries are more challenged and SAR data may be used.
•We show preliminary results obtained in two case studies :
•Case Study 1 : Santa Marinella wildfire- 20 July 2017- in Latium Italy
•Case study 2: Stalybridge wildfire - 10th April 2016- in The Peak District National
What we learned
Methods
RESULTS
Conventional Approach: •Pixels affected by fire trivially show a vegetation decrease • such pixels show high positive dNDVI values -white areas in (b) •The dNDVI histogram shows a significant separation between the burned area pixels and the unburned area pixels. •The two classes can be therefore separated by a simple thresholding technique, generating a low number of classification errors Proposed alternative a posteriori approach to be used when an image acquired before the fire is not available shows promising results •The vegetation re-growth in a burned pixel is supposed to be higher than a pixel not affected by fire •Pixels characterized by high re-growth show lower values of dNDVIp •The dNDVIp histograms (i) and (o) shows that burned area pixels are still relatively separated by the unburned ones. They can be therefore separated by a adaptive thresholding technique depending on a range of parameter including phenology •Time interval among the two acquisitions is a key parameter It should be sufficiently large to allow a significant re-growth in the burned area; however, a too large period could originate error because of significant changes in the unburned vegetation •The proposed technique based on dNDVIp values could be usefully adopted also in conjunction with other procedures such as SAR data •Work is in progress regards NBR index study and accuracy evaluation by using in field additional data
WHAT WE LEARNED
R.Carlà, L. Santurri, L.Bonora, C. Conese. “Multitemporal burnt area detection methods based on a couple of images acquired after the fire event” In Remote Sensing for Agriculture, Ecosystems, and Hydrology XI, SPIE Remote Sensing 2009. Berlin, Germany Burnt area detection in PDN using two post fire images, S . Amici, L. Santurri; l. Piscini UK Wildfires Conference 2017 - Wildfire resilience in a UK context (Wildfires2017), Bournemouth, 7 - 8 November 2017, doi.org/10.5281/zenodo.1164092
Acknowledgements •The authors thanks Prof. Julia McMoprrow for valuable feedback and support. •A special thanks to Dr. Gail Millin-Chalabi for organizing the field trip, taking pictures and describing the vegetation species. •Catherine Porter from PDNP is thanked for information about occurred fires •Dr. Massimo Selva from IFAC-CNR s thanked for the help in data processing. •SNAP by ESA and ENVI-HARRIS have been used to process the data •Sentinel playground visualization tool by synergize http://apps.sentinel-hub.com/sentinel-playground/ •for a quick look of available data
Case Study 1 Santa Marinella wildfire 20 July 2017 Latium Italy
Case study 2 Stalybridge wildfire 10th April 2016 The Peak District National
Sentinel 2A 20 April 2016
Sentinel 2A 19 July2016
Visualization in R (SWIR2), G (SWIR1) and B (Green) wavelength allows to visually identifying burned areas
•A multi-temporal approach rely on pre and post fire images to derive vegetation maps
•Two post fire images can be used to detect burn area •Vegetation phenology plays key role
•dNDVIp values histograms show relatively good separabilty •Adaptive thresholding to be used •Phenology impact on Sentinel 2A data and need further investigation
Sentinel 2 A Band 4 - 0.665 µm
Band 8 - 0.842 µm
(a) Usual multitemporal approach based on a couple of images acquired one before and one after the fire
(b) dNDVI map from Sentinel 2A data. White area is formed by pixel with high dNDVI values related to a vegetation decrease ; burned area is highlighted by a red line
(c) Histogram of dNDVI for burned area pixels (red) and not burned area pixels (green)
(d) Usual approach based on a single image acquired after the fire (image before the fire is not available)
(e) NDVI map from Sentinel 2A data. Dark area is formed by pixel with low NDVI values related to a low vegetation ; burned area is highlighted by a red shape
(f) Histogram of NDVI burned area pixels (red) and not burned area pixels (green)
(b) (c)
NDVIindex is used to measure healthy, green vegetation in remote sensing
data •The differential version of NDVI, named dNDVI is commonly used in assessment of vegetation change; if defined as dNDVI = NDVIpre – NDVIpost
high dNDVI values are related to a vegetation decrease, Pixels affected by fire trivially show a vegetation decrease; such pixel shows high positive dNDVI values
«a posteriori» dNDVIp When data before (pre) the fire are not available, an «a- posteriori» dNDVI is defined, based on data acquired after the event dNDVIp = NDVIpost – NDVIpost -regrow
Pixels affected by fire usually show a high re-growth of vegetation; in such case they show negative high modulus values of the dNDVIp (black areas in dNDVIp map)
(l)
(o)
(k)
(g)
(n)
Case Study 1- Conventional approach
(a)
Location Ignition date Sentinel 2A -
cloud free
Stalybridge
Fire
10 April
2016
No-pre data
post -20 April
2016
Post-19 July2016
Santa Marinella
Fire
Pre-11 May 2017
20 July 2017 20 July 2017
Post- 9 August
2017
Post - 27
November 2017
X NDVI
(d)
(h) (i)
(e) (f)
Case Study 1- Proposed alternative a posteriori approach
X NDVI
(j)
Case Study 2- Conventional approach
(m)
dNDVIp
dNDVIp
dNDVI
In situ validation point- image
acquired on July 2016 Case Study 2- Proposed alternative a posteriori approach
Sentinel2A cloud free data
Wildfire located on Google Earth
X
Sentinel 2A Pre -not available
Sentinel 2A Pre -11 May 2017
Sentinel 2A 20 July 2017 Active
Sentinel 2A Post 9 Aug. 2017
Sentinel 2A Post 27 Nov. 2017
(g) Proposed multitemporal approach based on images acquired after the fire event
(h) dNDVIp (a-posteriori dNDVI) map from Sentinel 2A data: black area is formed by pixels with lower values, that are related to a vegetation regrowth after the fire. In red is the burned perimeter extracted by visual selection of pixels on original data.
(i) Histogram of dNDVIp values for burned area pixel (red) and not burned area pixel (green)
(j) Conventional approach based on a single image acquired after the fire (image before the fire is not available)
(k) NDVI map from Sentinel 2A data. Dark area is formed by pixel with low NDVI values related to a low vegetation; burned perimeter is highlighted by a red . The blue star indicates the in situ validation point showing traces of burned peat under re-growing moorland vegetation
(l) Histogram of NDVI burned area pixels (red) and not burned area pixels (green)
(m) Proposed multitemporal approach based on images acquired after the fire event
(n) dNDVIp (a-posteriori dNDVI) map from Sentinel 2A data: black area is formed by pixels with lower values, that are related to a vegetation regrowth after the fire. In red the burned area extracted by visual selection of pixels on original data.
(o) Histogram of dNDVIp values for burned area pixel (red) and not burned area pixel (green)
References
REMARKS