Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data
Wetlands are an important ecosystem for many vital functions such as groundwater recharge, flood control, water quality improvement, and to mitigate erosion. Monitoring and mapping wetlands on a large scale is becoming increasingly more important, and satellite remote sensing provides a practical ap...
| Autor principal: | |
|---|---|
| Formato: | H2 |
| Lenguaje: | Inglés |
| Publicado: |
SLU/Dept. of Forest Resource Management
2012
|
| Materias: |
| _version_ | 1855570792720367616 |
|---|---|
| author | Martin, Jennifer |
| author_browse | Martin, Jennifer |
| author_facet | Martin, Jennifer |
| author_sort | Martin, Jennifer |
| collection | Epsilon Archive for Student Projects |
| description | Wetlands are an important ecosystem for many vital functions such as groundwater recharge, flood control, water quality improvement, and to mitigate erosion. Monitoring and mapping wetlands on a large scale is becoming increasingly more important, and satellite remote sensing provides a practical approach. This study examines the potential for using multi-beam Radarsat-2 C-band polarimetric SAR, Landsat-5 TM, and DEM data for classifying wetland and non-wetland classes in a forested watershed in Ontario, Canada. It investigates the influence of incidence angle, leaf presence and moisture conditions in the classification of SAR images. The images were classified using two classification methods: the Maximum Likelihood Classifier and Random Forests classifier. Lastly, SAR polarimetric variables and decompositions were investigated for their usefulness in classification.
Fourteen Radarsat-2 Fine Quad (FQ) SAR images were acquired from October 2010 to November 2011 at different incidence angles but with the same orbit-descending pass (west-looking direction). The images were paired according to the beam mode (FQ4 and FQ22/27), leaf presence (off and on) and moisture (wet/dry) conditions. The FQ image pair which gave the best classification overall accuracy (76.3%) using the Maximum Likelihood classification was those from the two FQ22/27 images acquired under leaf-off and dry conditions. When the FQ images were classified together with five optical bands of a Landsat image, the classification accuracy was higher for all classes as well as for the overall accuracy (94.4%). When the FQ images were combined with the Landsat image and slope, overall accuracy improved only slightly from the FQ and Landsat combination (95.4%).
With the Random Forests classification, the best overall accuracy was obtained with the combination of the FQ 22/27 image pair acquired under leaf-off and dry image conditions, Landsat and slope (98.7%), followed closely by the FQ pair and Landsat combination (98.6%). When all FQ images were used as input to the Random Forests classification, this also produced high cross-validation overall accuracies (98.3%), indicating that while Landsat does add accuracy FQ images can give comparable accuracies if the right dates and conditions are chosen. A benefit of using Random Forests is the ability to rank band importance in image classification. From this it was determined that using multiple FQ images with leaf-off conditions were preferred. As for the other conditions, a mix of incidence angles, moisture conditions, and polarizations were important for classification. The incoherent target decompositions were the most important polarimetric variable in the classification, while the only other parameter indicated as important from both classifications was the orientation angle for the maximum of the completely polarized component.
In future studies, it may be of interest to test the combination of multi-date polarimetric variables and decompositions parameters together with all polarizations (HH, HV, VH, and VV). So far, we classified only two types of wetlands (closed and open). Further studies are needed to test the Random Forests classifier for classifying the wetlands into more detailed classes (bog, fen, marsh, swamp, etc.). Lastly, future studies should test the results found here using independent evaluation data to assess the accuracy.
|
| format | H2 |
| id | RepoSLU5127 |
| institution | Swedish University of Agricultural Sciences |
| language | Inglés |
| publishDate | 2012 |
| publishDateSort | 2012 |
| publisher | SLU/Dept. of Forest Resource Management |
| publisherStr | SLU/Dept. of Forest Resource Management |
| record_format | eprints |
| spelling | RepoSLU51272012-12-13T10:37:05Z Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data Martin, Jennifer Radarsat Landsat wetlands classification land cover multi-soruce data remote sensing Random Forests Maximum Likelihood Classifier Wetlands are an important ecosystem for many vital functions such as groundwater recharge, flood control, water quality improvement, and to mitigate erosion. Monitoring and mapping wetlands on a large scale is becoming increasingly more important, and satellite remote sensing provides a practical approach. This study examines the potential for using multi-beam Radarsat-2 C-band polarimetric SAR, Landsat-5 TM, and DEM data for classifying wetland and non-wetland classes in a forested watershed in Ontario, Canada. It investigates the influence of incidence angle, leaf presence and moisture conditions in the classification of SAR images. The images were classified using two classification methods: the Maximum Likelihood Classifier and Random Forests classifier. Lastly, SAR polarimetric variables and decompositions were investigated for their usefulness in classification. Fourteen Radarsat-2 Fine Quad (FQ) SAR images were acquired from October 2010 to November 2011 at different incidence angles but with the same orbit-descending pass (west-looking direction). The images were paired according to the beam mode (FQ4 and FQ22/27), leaf presence (off and on) and moisture (wet/dry) conditions. The FQ image pair which gave the best classification overall accuracy (76.3%) using the Maximum Likelihood classification was those from the two FQ22/27 images acquired under leaf-off and dry conditions. When the FQ images were classified together with five optical bands of a Landsat image, the classification accuracy was higher for all classes as well as for the overall accuracy (94.4%). When the FQ images were combined with the Landsat image and slope, overall accuracy improved only slightly from the FQ and Landsat combination (95.4%). With the Random Forests classification, the best overall accuracy was obtained with the combination of the FQ 22/27 image pair acquired under leaf-off and dry image conditions, Landsat and slope (98.7%), followed closely by the FQ pair and Landsat combination (98.6%). When all FQ images were used as input to the Random Forests classification, this also produced high cross-validation overall accuracies (98.3%), indicating that while Landsat does add accuracy FQ images can give comparable accuracies if the right dates and conditions are chosen. A benefit of using Random Forests is the ability to rank band importance in image classification. From this it was determined that using multiple FQ images with leaf-off conditions were preferred. As for the other conditions, a mix of incidence angles, moisture conditions, and polarizations were important for classification. The incoherent target decompositions were the most important polarimetric variable in the classification, while the only other parameter indicated as important from both classifications was the orientation angle for the maximum of the completely polarized component. In future studies, it may be of interest to test the combination of multi-date polarimetric variables and decompositions parameters together with all polarizations (HH, HV, VH, and VV). So far, we classified only two types of wetlands (closed and open). Further studies are needed to test the Random Forests classifier for classifying the wetlands into more detailed classes (bog, fen, marsh, swamp, etc.). Lastly, future studies should test the results found here using independent evaluation data to assess the accuracy. SLU/Dept. of Forest Resource Management 2012 H2 eng https://stud.epsilon.slu.se/5127/ |
| spellingShingle | Radarsat Landsat wetlands classification land cover multi-soruce data remote sensing Random Forests Maximum Likelihood Classifier Martin, Jennifer Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title | Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title_full | Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title_fullStr | Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title_full_unstemmed | Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title_short | Mapping wetland areas on forested landsacpes using Radarasat-2 and Landsat-5 TM data |
| title_sort | mapping wetland areas on forested landsacpes using radarasat-2 and landsat-5 tm data |
| topic | Radarsat Landsat wetlands classification land cover multi-soruce data remote sensing Random Forests Maximum Likelihood Classifier |