High Resolution Satellite Data for Forest Management - Algorithm for Tree Counting - Kiyoshi HONDA ACRoRS, Asian Institute of Technology NASDA ALOS (NASDA JAPAN) 2.5m Resolution Launch in 2002 Panchromatic Remote sensing Instrument for Stereo Mapping (PRISM) It has three independent catoptric systems for nadir, forward and backward looking to achieve along-track stereoscope. Each telescope consists of three mirrors and several CCD detectors for push-broom scanning. Nadir-looking telescope provides 70 km width coverage and forward and backward 35 km. Forward and backward telescopes are inclined about +-24 degrees from nadir to realize base to height ratio = 1. Its 2.5-meter resolution data will be used for extracting highly accurate digital elevation model (DEM).
IKONOS 1-meter Resolution Hishron Museum and Sculpture Garden Washington DC, USA 1999 Source: Space Imaging Corp. 1999 SPIN-2 2 meter Resolution Fulton County Stadium. Downtown Atlanta Georgia, USA. 1995 Source: SPIN-2 Corp. 1995
Future Name Present IKONOS Present & Future High Resolution Satellites SPIN-2 Russia - Multispectral 4 4 Panchromatic 1 1 ALOS NASDA / Japan 2002 PRISM Panchromatic 1 2.5 Quick Bird OrbView-3 Country Space Imaging / USA 1999 Earth Watch / USA Launch 2000 Orbimage / USA 2000 OrbView-4 Orbimage / USA 2001 Sensors Optical Telescope Assembly KVR-1000 - Photograph Pushbroom Linear Arrary Types channels (meters) Panchromatic 1 2 Multispectral 4 4 Panchromatic 1 1 Multispectral 4 4 Panchromatic 1 1 Multispectral 4 4 Panchromatic 1 1 EROS-A West Indian Space, Ltd./ USA 2000 Panchromatic 1 1.5 EROS-B1 West Indian Space, Ltd./ USA 2001 Panchromatic 1 0.82 IRS-P5 ISRO / India 2000 LISS IV Panchromatic 1 2.5 IRS-P6 ISRO / India 2001 LISS IV Panchromatic 1 6 SPOT-5 SPOT / France 2001 HRV Panchromatic 1 5 Source: tree counting algorithm development Introduction Emerging Super High Resolution Satellites (1m ~ 6m) ALOS, IKNOS, SPIN2. Possibility for identifying each tree crown from space Getting more detail and accurate information of forest Tree Crown Identification, Number of Trees, Crown Size, Volume, Identify big trees. Usual Classification algorithms ( pixel wise approach ) are not applicable for tree crown identification -> Develop a new algorithm for identifying tree crown
Flow of the development of the algorithm I. Model image of high resolution satellite data using Japanese Cedar model and 3D computer graphics. ( Ideal Image ) II. Algorithm Development based on the model image. III. Simulation Image of high resolution satellite data from an aerial photograph by degrading the resolution IV. Accuracy Check Japanese Cedar Tree Model Cross section of a Tree Crown 0.5 y = αx (K. TAKESHITA, 1985) α = f ( Height, Density ) Height of tree 30(m) density of standing trees 494(n/ha) tree crown curve y = 0.79x 0.5 α = 0.79
The 3D image of the Japanese Cedar model Diameter: 4.6m Front view Perspective view CG Image of the Japanese Cedar Forest Model Perspective view Top view
Model image of the high resolution satellite image (CG) 1m Resolution 3D image of the model image ( Height is a value of pixel)
Identification of Tree Crown by Threshholding ( The threshold is average of pixel value) Pixel value is more than average Pixel value is less than average New Algorithm for Extracting Tree Crown High Tree crown Tree crown Tree Crown Not Tree Crown Pixel value Threshold Low P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Pixel P:Target pixel P1:Neighbor 1 P2:Neighbor 2
Ravine Bottom Recognition using the difference in pixel value of neighbor p(1) p(2) p(3) p(4) p(0) p(5) p(6) p(7) p(8) 8-neighbor of p(0) Difference between P(0) and P(n) q(1) = p(0) - p(1) q(2) = p(0) - p(2) q(3) = p(0) - p(3) q(4) = p(0) - p(4) q(5) = p(0) - p(5) q(6) = p(0) - p(6) q(7) = p(0) - p(7) q(8) = p(0) - p(8) The condition of bottom of a ravine in the image 1) {q(1)<0 & q(8)<=0} or {q(1)<=0 & q(8)<0} 2) {q(2)<0 & q(7)<=0} or {q(2)<=0 & q(7)<0} 3) {q(3)<0 & q(6)<=0} or {q(3)<=0 & q(6)<0} 4) {q(4)<0 & q(5)<=0} or {q(4)<=0 & q(5)<0} true Other Image value = 0 Image value = 1 Identification of Tree Crown by Ravine Bottom Recognition Not the bottom of a ravine the bottom of a ravine
Tree Crown Identification by New Algorithm ( Thresholding & Ravine Bottom Recognition ) Tree crow Not tree crown Counting Tree Crown by Labeling
Satellite Simulation Image from an Aerial Photograph Target place Experimental forest of Mie university 975 trees per hectare Simulation image of the high resolution satellite data (1) Aerial photograph (21 cm/pixel) Simulation image (80 cm/pixel)
Simulation image of the high resolution satellite data (2) Simulation image (1 m/pixel) Simulation image (3 m/pixel) 80 cm / pixel
1m / pixel 3m / pixel
Tree Crown Identification Result 0.8m/pixel 1m/pixel 3m/pixel Visual identification result of each tree crown N = 274
The accuracy in terms of number of tree crown in the image number of standing trees (visual) number of standing trees (suppositional) Number of standing trees 300 250 200 150 100 50 0 87.6% 69.7% 16.8% 0.8 1 3 Ground resolution (m/pixel) The accuracy in terms of identification of each tree crown 0.8m/pixel 1m/pixel 1)Identified 2)Not Identified 3)Combined 4)Divided
The accuracy in terms of identification of each tree crown at different diameter class Identified Combined Not identified Divided Rate of number of tree crown(%) 100% 80% 60% 40% 20% 0% 74 147 53 274 0 5 4 9 0 8 26 18 0 18 13 5 43 129 0.8m/piexl 49 221 0-2 2-4 4-6 all Diameter of tree crown(m) Rate of number of tree crown(%) 100% 80% 60% 40% 20% 0% 74 147 53 274 1 1 2 4 31 5 25 2 61 9 31 20 106 44 178 28 0-2 2-4 4-6 all Diameter of tree crown(m) 1m/pixel Application to Teak Bearing Forest in Myanmar 1 m / Pixel 2 m / Pixel
Visual Identification of Crown Crown Area were identified visually for accuracy check-up. These Crowns were digitised on the screen manually The result image Colour assigned to appropriate size of crown
Graph of accuracy in 1 meter Overall accuracy of 1 meter image is 72%. The composition of identified crown is 70 out of 78 (89%) Accuracy % 120.0 100.0 80.0 60.0 40.0 20.0 0.0 0-10 10-20 20-40 40-65 1m resolution Image Accuracy. 65-100 Crown Size Classes 100-200 200-300 300-400 400-500 500-600 1m. Graph of accuracy in 2 meter Overall accuracy of 2 meter image is 68%. The composition of identified crown is 66 out of 78 (84%) Accuracy % 120 100 80 60 40 20 0 0-10 10-20 20-40 40-65 2m resolution Image Accuracy. 65-100 Crown Size Classes 100-200 200-300 300-400 400-500 500-600 2m.
Conclusion Japanese Cedar Forest The accuracy in terms of number of tree crown in the image 0.8m/pixel 87.6% 1m/pixel 69.7% 3m/pixel 16.8% Accuracy of identification of each tree crown at different diameter class Teak Bearing Forest 80cm/pixel 2m~4m 87.8% 4m~6m 92.5% 1m/pixel 2m~4m 72.1% 4m~6m 83.0% 1m/pixel 72% 2m/pixel 68% New algorithm makes it possible to apply High Resolution Satellite Image to Identify Each Tree Crown for the management of forest resources. Conclusion - cont d Future work Algorithm Development For local important species or forest. More detailed forest and tree structure ( Diameter, Height, Biomass. ) Combination with other instruments Developing Database of Forest Structure: Global Environment