potential of deep learning in satellite-based fire detection and ... · nceo king’s college...

Tianran Zhang, Martin Wooster, Daniel Fisher

NCEO King’s College London

Potential of Deep Learning in Satellite-based Fire Detection and Emission Estimation Study

2

Tianran Zhang | [email protected] | NCEO King's College London

Aqua-MODIS15 Sept 2014

Sept 2014

Jan 2015

April 2014

Copernicus Atmosphere Monitoring

Service (CAMS)

MODIS FRP

Based

Global Fire Assimilation System (GFAS)

3


VIIRS Channels

Potential Application 1:

Fire Detection Alg. Using Multi-spectral Remote Sensing Data

Input: 28x28 window data for all 16 VIIRS-M band channels

Output: Cloud/water/land/fire Mask

4



Fire Detection Alg. Using Multi-spectral Remote Sensing Data

Results Analysis: Overfitting 5


Data Scientist Solution

• More training data.

• Change model structure.

• Apply regularization.

Illustration of overfitting.

(source: Wikipedia). Remote Sensing Scientist Solution

• Look into data details (geometry, seasonal/regional special pattern).

• Improve data quality by using other source/higher resolution data.

Results Analysis 6


t-Distributed Stochastic NeighborEmbeddingNonlinear dimensionality reduction technique for embedding high-dimensional data for visualization in a low-dimensional space.

1

2 3

4 5


Real Time Multi Masks Generation 7


8


MODIS VIIRS

VIIRS 26/07/2019

Current fire products suffer 10-40% false fire detection in certain urban area of Eastern China.


Improving Accuracy of Fire Detection Alg.

9


GlobeLand30 land cover product (Chen et al, 2015) re-gridded to 0.01degree in eastern China (111-123° E, 27-40° N).

• Landcover map product has long delay.

• Data not available for real-time fire detection.

• Dynamic world (urbanisation).• Spatial & temporal resolution.(Seasonal fire patterns vs. yearly map).



10


Source: University of Cincinnati

VIIRS SDR data(L1B product)Training Stage Labels:

VIIRS Active Fire Product + Landcover MapVegetation MapCultivation Map……

Deploying StageVIIRS SDR data(L1B product)

Results



Deep or Deeper 11


x2 x3

Params/Accuracy

Simple DLResnet20

(n=3)

Params 55,565 858,389

Fire 97% 97%

Land 95% 95%

Cloud 88% 96%

Water 92% 94%

False Fire 67% 84%

12



Smoke Plume IdentificationSEVIRI Active Fire Detections

MODIS AF Detections

MODIS AF Detections Remapped

MODIS 10 km AOD Product

CAMS Global Fire Assimilation System 13


potential of deep learning in satellite-based fire detection and ... · nceo king’s college...

Documents