Evaluation of Multi-temporal and Multipolarization ASAR for Boreal Forests in Hinton David G. Goodenough Hao Chen Andrew Dyk Tian Han Pacific Forestry Centre Natural Resources Canada Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts © May 16, 2005 Project Objectives Radarsat-2 has a significant improvements in capability compared with Radarsat-1: better spatial resolution and fully polarimetric. Envisat ASAR alternating polarization (AP) mode can help to study the advanced capabilities of Radarsat-2 and make progress towards defining methods for potential use of multi-polarization Cband SAR data for forest applications. Objectives: • Develop a forest land-cover classifier for multitemporal and multi-polarization ASAR APP data; focus on clear-cut, reforestation, coniferous, deciduous, and mixed wood in northern forested areas in Canada; determine effectiveness of C-band ASAR APP data for boreal forest mapping and change detection. • Perform data fusion analysis with multitemporal ASAR and Landsat ETM+ data and determine improvements of fused data over single sensor data sets. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Radarsat-2, ASAR, and Convair-580 Radarsat-2, operating at 5.405 GHZ, will be able to image at spatial resolutions ranging from 3 to 100 meters with nominal swath widths ranging from 10 to 500 kilometers. In addition, Radarsat-2 will offer multi-polarization, a capability that aids in identifying a wide variety of surface features and targets. Envisat ASAR is a C-band sensor with dual-polarization and ability to acquire data with broad swath coverage, range of incidence angles, polarization, and modes of operation. This study uses Envisat ASAR data for multitemporal SAR classification. The Convair 580 C-band SAR can provide quad-polarization data (HH/HV/VH/VV), all in one scene, for polarimetric or for dual polarization combination analyses. Convair 580 data will help the study of Radarsat-2 data for forest applications and transformation of the technology to the end users of Radarsat-2 in the remote sensing service industry. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Study Site in Hinton, AB Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Landsat TM channel 5, 4, and 3 are shown as red, green, and blue channels. The white rectangles indicate the 29 township-ranges. Image Acquisitions Image acquisitions Satellite Envisat Envisat Envisat Envisat Landsat Sensor ASAR ASAR ASAR ASAR ETM+ Collection date 1/10/2004 2/14/2004 7/3/2004 8/7/2004 8/23/2002 Polarizations HV/HH VV/VH HV/HH VV/VH N/A Beam IS7 IS7 IS7 IS7 N/A Product Type ASA_APP_1P ASA_APP_1P ASA_APP_1P ASA_APP_1P Scene based Image information ASAR Alternating Polarization Precision (APP): Channels: Two co-registered image channels corresponding to one of HV/HH, VV/VH, and HH/VV Spatial resolution: Approximately 30 m Pixel spacing: 12.5 m Beam mode: IS 7 (42.5 – 45.5D) Landsat ETM+: Channels: 6 Spectral channels Spatial resolution: Approximately 30 m Cloud coverage: 0% Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts SAFORAH Data Grid Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts http://www.saforah.org Image locations ASAR APP metadata Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts ASAR Data Preprocessing Single dual-pol. ASAR APP image: Radiometric calibration: Power Speckle filtering (Lee adaptive): 5x5 _ R (t ) I (t ) W (t ) I (t ) (1 w(t )) where the weighting function W is given by W (t ) 1 C / C (t ) 2 u 2 I _ ( Cu u /u Noise variation coefficient) A. Lopes, R. Touzi and E. Nezry, "Adaptive speckle filters and Scene heterogeneity", IEEE Transaction on Geoscience and Remote Sensing, Vol. 28, No. 6, pp. 992-1000, Nov. 1990. Textures: Mean, Standard deviation, Entropy (11x11 window size) Orthorectification: 12.5 m pixel resampling ASAR APP data fusion (HV/HH/VV/VH): Image fusion from a pair of two winter images: 35 days apart Image fusion from a pair of two summer images: 35 days apart Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts LOGIT non-Gaussian classification To determine the effectiveness of the C-band Envisat ASAR data for forest typing, structure recognition, and disturbance detection, a hierarchical logistic classifier, LOGIT, was developed. LOGIT is a hierarchical logistic non-parametric classification program: SAR: Poisson distribution. ETM+: Gaussian distribution Standard Gaussian Distribution Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Poisson Distribution Curves 1 All Classes 3 2 Class Number Class Name 1 Coniferous 5 39 Clouds 1,5,6,11,18,19, 26,30,34,40 "Barren Classes" 4 Mixed wood 5 1,5,18,19, 30,34,40 6,11,26 6 Regeneration 11 Clearcut 18 Scrub coniferous 19 Scrub deciduous 26 Mining area 30 Water 34 Deciduous 6 39 7 6 Regeneration 10 11 11 Clear-cut Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts 1,5,18,19, 34,40 13 12 26 Mining 40 Cloud Shadows 1,5,18, 19,34 15 14 1,5 18,19,34 16 17 34 Deciduous 18,19 Cloud Shadow 9 30 Water 11,26 20 40 8 18 Scrub Coniferous 18 1 Conifers 19 5 M ixed Wood 21 19 Scrub Deciduous Example of a Logistic Tree Built from Landsat ETM+ Training ASAR Textures Textures were used to extract additional information from original SAR image channels by looking at the statistics and value changes in small regions. Different window sizes and grey levels were tested. The window size 11x11 and 32 grey levels were selected. The resulting texture images tend to have interesting patterns and formations that signified additional information for LOGIT classification. Textures used in this study: Mean Standard Deviation Entropy mean x n 2 n x x 2 SD nn 1 2 N 1 P i, j ( ln Pi , j ) i , j 0 Hall-Beyer, Mryka, “GLCM TEXTURE: A TUTORIAL” 2004, Department of Geography University of Calgary Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Training for LOGIT Classification Training identified from Quickbird and aerial photos; Examples of training samples from Quickbird: Clear-cut Low Regeneration Exposed Land Dense Coniferous 8 training classes were selected. Since the dominant forest in the study area was coniferous, two forest classes were defined. Coniferous contains at least 80% conifer and Mixed Wood contains all other forest types. Other classes were Clear-cut, Low regeneration (< 2 meters high), High regeneration (> 2 meters high), Scrub, Water, and Exposed land. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts LOGIT Classification on Fused ASAR APP (1) HV/HH (W) VV/VH (W) HV/HH/VV/VH (W) HV/HH/VV/VH (S) HV/HH/VV/VH (W) Class Label Accuracy Accuracy Accuracy Accuracy Accuracy Coniferous 77.2% 66.3% 79.2% 78.2% 33.2% Mixed wood Not included Not included Not included Not included 68.9% Low regeneration 49.3% 28.5% 60.5% 36.8% 53.0% High regeneration Not included Not included Not included Not included 32.8% Clear-cut 69.0% 85.7% 86.8% 60.2% 86.7% Scrub 65.4% 43.4% 82.9% 81.7% 83.0% Water 81.0% 74.5% 89.6% 93.6% 89.6% Exposed Land 54.1% 50.1% 90.1% 93.9% 90.2% Average 66.0% 58.1% 81.5% 74.1% 67.2% Overall 73.9% 66.0% 79.6% 73.9% 44.5% W – Winter data set S – Summer data set Without Mixed Wood and High Regeneration: The classification accuracies of the summer HV/HH/VV/VH image were lower than the winter HV/HH/VV/VH data. Adding in Mixed Wood and High Regeneration: Classification accuracies were greatly reduced. An example was given in the table (HV/HV/VV/VH). Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Confusion Matrix of HV/HH/VV/VH (Winter ) Adding Mixed Wood and High Regeneration to the classification of the winter HV/HH/VV/VH data set. Classification accuracy: Average : 67.2% Overall: 44.5% 26% of Coniferous was misclassified as Mixed wood, and 23.7% of Coniferous as High regeneration. Low classification accuracies also occurred for the two regeneration classes: High Regeneration and Low Regeneration Class Coniferous Mixed Wood High Regen. Low Regen. Clearcut Scrub Exposed Land Water Coniferous 33.2 12.9 21.8 3.6 0.0 0.3 0.0 0.0 Natural Resources Canada Canadian Forest Service Mixed Wood High Regen. Low Regen. 26.5 23.7 10.8 68.9 12.4 4.9 26.0 32.8 13.9 9.1 13.3 53.0 0.0 0.0 2.5 0.0 0.2 6.4 0.0 0.0 1.4 0.0 0.0 2.1 Ressources naturelles Canada Service canadien des forêts Clearcut Scrub Water 0.0 5.6 0 0.0 0.9 0 0.0 5.5 0 5.3 10.9 1.2 86.7 6.1 1.3 4.7 83.0 0.2 2.9 0.8 89.6 3.6 2.5 1.7 Exposed Land 0.1 0.0 0.0 3.4 3.4 5.2 5.3 90.2 LOGIT Classification on Fused ASAR APP (2) Multi-Pol. (W) + (S) Landsat-7 (S) Multi-Pol. (W) + L-7 Multi-Pol. (S) + L-7 Class Label Accuracy Accuracy Accuracy Accuracy Coniferous 54.1% 90.8% 88.1% 92.2% Mixed wood 69.9% 80.0% 82.1% 76.0% Low regeneration 40.3% 82.8% 94.8% 89.6% High regeneration 39.0% 84.5% 85.8% 75.8% Clear-cut 94.6% 97.0% 97.8% 97.6% Scrub 92.8% 85.1% 91.9% 88.5% Water 94.0% 95.1% 99.1% 99.7% Exposed Land 96.1% 95.6% 99.5% 99.2% Average 72.6% 88.9% 92.4% 89.8% Overall 59.4% 89.9% 89.8% 90.4% W – Winter data set S – Summer data set The Landsat-only classification had high average and overall accuracies. The Landsat +ASAR (W) data achieved 92.4% for the average, 3.5% higher than ETM+. The fused Landsat +ASAR (S) classification was at the same level as the Landsat +ASAR (W) image. The fused ASAR imagery, with both summer and winter images had the average accuracy of 72.6% and overall accuracy of 59.4%. If Landsat or other optical images are not available, this data combination should be the best alternative. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Confusion Matrix of HV/HH/VV/VH (Winter + Summer) LOGIT classification on the Multi-Pol. (W)+(S) data set: Winter: HV, HH, VV, VH, + Means Summer: HV, HH, VV, VH, + Means Classification accuracy: 72.6% Average 59.4% Overall Coniferous classification was improved from 33.2% (a single winter multi-pol. data set) to 54.1%. But, classification accuracies on High Regeneration and Low Regeneration are still low. Class Coniferous Mixed Wood High Regen. Low Regen. Clearcut Scrub Exposed Land Water Coniferous 54.1 10.1 26.9 8.7 0.0 0.1 0.0 0.0 Natural Resources Canada Canadian Forest Service Mixed Wood High Regen. 11.9 20.6 69.9 13.8 19.5 39.0 11.3 9.2 0.0 0.0 0.0 0.4 0.0 0.0 0.0 0.0 Ressources naturelles Canada Service canadien des forêts Low Regen. 10.3 4.9 9.4 40.3 1.7 3.6 0.2 2.6 Clearcut 0.0 0.0 0.1 8.9 94.6 2.4 2.1 0.0 Scrub 2.8 1.2 4.9 17.7 2.9 92.8 1.0 0.0 Water 0.0 0.0 0.0 0.0 0.3 0.2 94.0 0.9 Exposed Land 0.2 0.0 0.1 3.9 0.5 0.6 2.7 96.1 LOGIT Classification Image ASAR-ETM+ Classification Image overlaid on 20020823 Landsat-7 ETM+ Image Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Convair 580 quad polarization Data Future Study for Radarsat-2 We have acquired Convair 580 images over Hinton in September 2004. • Conv580 data: Simultaneous C-Band in four polarization combinations, HH/HV/VV/VH • Conv580 data resolution: 6 – 10 meters. This will allow us to select the combination that is best for forest applications. Compare LOGIT classification results between Convair 580 quad polarization data and fused ASAR APP data. Perform Convair 580 polarimetric data decomposition: • Entropy/ decomposition (Ref. Cloude and Pottier, TGRS 1997) and three-component scattering decomposition (Ref. Freeman and Durden, TGRS 1998) • Wishart classification Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Samples of Convair 580 Images over Hinton Township 48 Range 22 ~ 9.66 x 9.66 KM2 Township 50 Range 23 Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts ~ 9.66 x 9.66 KM2 Conclusions To prepare for Radarsat-2, Envisat ASAR APP and Landsat ETM+ were acquired over Hinton. Image preprocessing and data fusion from multitemporal ASAR images were performed. A hierarchical logistic classification program, LOGIT, was implemented. LOGIT provided the non-parametric method of classification on remotely sensed data. LOGIT was applied on ASAR APP, ETM+, and various fused image sets. The classification results revealed that • Landsat ETM+: High classification accuracy (Average 88.9%) • Fused ASAR and ETM+: Average 92.4%; that is 3.5% higher than ETM+ average classification accuracy • Multitemporal and multi-polarization ASAR data set (winter and summer): Average 72.6% • Single-date dual-polarization ASAR: Average 66% (HV winter) that is better than other single-date dual-polarization acquisitions. More classification experiments on fused data sets, different band combinations, and multitemporal classification comparisons are under way. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts Acknowledgements We acknowledge the European Space Agency for providing Envisat ASAR data to Canadian Forest Service under the Category-1 Project 2481. We thank Tarin Resource Services Ltd. of Calgary for providing online search of orthorectified aerial photos. We appreciate the assistance of Steven Carey and Lionel Cai, co-op students at the Pacific Forestry Centre, Canadian Forest Service. We are most grateful for financial support from Natural Resources Canada, the Canadian Space Agency, and the Natural Sciences and Engineering Research Council. Natural Resources Canada Canadian Forest Service Ressources naturelles Canada Service canadien des forêts
© Copyright 2026 Paperzz