2.1.

Flood Inundation Area Extraction

There are several advantages of using RADARSAT-2 in flood monitoring: (1) a left- and right-looking sensor, (2) short revisit time, (3) multi-image mode, and (4) multipolarization. The selective look direction is particularly effective in flood management.²¹ In particular, the polarimetric mode is very useful in monitoring the vegetated areas.²² These features also make it well suited for rapid disaster response, monitoring, mitigation, and evaluation.

Flooded areas appear dark in SAR images due to specular reflection.³ After preprocessing, including geometric correction, we classified the images to extract the inundated areas using temporal data acquired during the flood. For the image filtering, we used an enhanced frost filter with a window size of $7\times 7$, which can reduce noise and also preserve detailed information. Combined with human interpretation, the highly efficient decision tree method was applied to rapidly extract the inundated areas. The resulting images were then imported to ArcGIS to extract the inundated area changes. The processing flow chart used is shown in Fig. 1.

Fig. 1

Flow chart of flood monitoring using RADARSAT-2 data.

2.2.

PolInSAR Tree Height Extraction

PolInSAR is a rapidly evolving technique and has been applied in many fields, such as forest height inversion and topographic mapping.^23–25 However, for a diverse and complex land surface, it is very difficult to precisely ascribe backscattering characteristics. In order to simplify the scattering model, some assumptions are necessary for PolInSAR applications. At the same time, the errors of incidence angle, slant range, and baseline length will also affect the accuracy of PolInSAR. The critical element is how to accurately estimate the locations of the effective scattering phase centers,²⁶ which depend on the system parameters, such as wavelength and polarization, as well as the physical and geometrical parameters of the scatterers.

Five different scattering mechanisms of PolInSAR occur when microwaves pass through the forest above ground level.²⁷ The random volume over ground (RVoG) model provides an equation to describe the relationship between volume coherence and the parameters of the forest, such as height and extinction.⁴ For a given forest and interferometric configuration, a forest with high extinction and tall trees may have the same volume coherence as a forest with lower extinction and shorter trees, which is the height and extinction ambiguity. Therefore, the PolInSAR technique is not appropriate for parameter inversion of these two kinds of forest.

In the RVoG model, the volume coherence increases proportionately with the volume height but inversely to the extinction. However, some studies have demonstrated the low dependence of height inversion on the mean extinction coefficient.²⁸ Additionally, for high extinction, the variation of vegetation height will also have no obvious effect on the volume coherence for the small penetration depth of the radar waves. In this study, we used two scenes of fully polarimetric data acquired by RADARSAT-2 over Tiantan Park in Beijing to estimate forest height.

2.3.

Alluvial Fan Surface Mapping

Many geomorphological features related to the alluvial fan units of different ages can be extracted from SAR images, especially PolSAR images. These features’ parameters can be used to differentiate between different alluvial fan units. Three steps were adopted for this study.

Cloude decomposition expresses the target scattering vector $\overrightarrow{k}$ in terms of five parameters under a reciprocity assumption, i.e.,

Some key parameters, such as scattering type parameter alpha ($\alpha$), coherency entropy (H), anisotropy (A), target orientation angle ($\beta$), and target phase angle (${\mathrm{\Phi }}_i$, $i=1,2,3$) have been introduced from the eigenvalue decomposition of the Hermitian target coherency matrixes. H is a global measure of the distribution of components in the scattering process and can be used to indicate the complexity and randomness of scattering types. A high alpha in the intact region indicates that only one dominant scattering type occurs there. Alpha shows the scattering type of observed targets. Recent studies have shown that the Cloude alpha scattering type ambiguities may occur for certain scatterers.³⁰ Touzi et al. proposed a roll-invariant target decomposition method to solve this problem.³¹ Some roll-invariant parameters can be extracted by using this method, such as symmetric scattering type ${\alpha }_s$ and its phase $\mathrm{\Phi }{\alpha }_s$, among others.

In this paper, RADARSAT-2 fine mode PolSAR data were acquired over the Shule River alluvial fan and the temporal-spatial distribution of the fans were studied.

2.4.

Support Vector Machine-Based Impervious Urban Surface Extraction

Recently, urban impervious surfaces were mostly estimated from optical imagery. However, there have been some open problems, including

To resolve these limitations, RADARSAT-2 full PolSAR data (C-band) were used to investigate the potential of extracting urban impervious surfaces in dense urban areas. In our research, a robust algorithm, a support vector machine (SVM), was used for estimating impervious surfaces. SVMs have been used for image classification^32–35 and impervious surface estimation^36,37 in remote sensing studies in recent years. Because too many PolSAR-identified features will result in data redundancy and reduce classification accuracy, the optimization of features is an important step for extracting impervious surfaces from PolSAR data. The various polarimetric features were selected through a stepwise process. Different classification accuracies were derived when different PolSAR features were put into the SVM classifier. While reducing the classification accuracy, the features with an adverse effect on classification were removed, and when improving the classification accuracy, the features with a positive effect on classification were retained. Then, the PolSAR features that improved the classification accuracy were further statistically analyzed using the sample points of seven land use types. After statistical analysis of these features, the final polarimetric features were optimally selected for estimating the impervious surfaces. Figure 2 shows a detailed flow chart of the optimized feature selection process for PolSAR processing.

Fig. 2

Detailed flow chart of the optimized feature selection process.

3.1.

Flood Disaster Monitoring and Evaluation with RADARSAT-2 Imagery

China often suffers from disasters caused by floods, with huge losses every year. The regions around Dongting Lake and Poyang Lake are the two most seriously flood-prone areas in China.

Dongting Lake, one of the largest and shallowest lakes in China, is a flood basin of the Yangtze River in northeastern Hunan Province. Poyang Lake is also a flood basin of the Yangtze in northern Jiangxi Province.

On June 21, 2010, the eastern bank of the Fuhe River (the upstream of Poyang Lake) collapsed in Fuzhou City, Jiangxi Province. One hundred thousand people were under the threat of flooding. Center for Earth Observation and Digital Earth (CEODE), Chinese Academy of Sciences scientists developed emergency satellite observation plans to monitor dynamic changes and evaluate disaster losses using RADARSAT-2 imagery during the cloudy and rainy weather. In this paper, fine, ultrafine, and ScanSAR narrow mode RADARSAT-2 data were used for flood monitoring due to their large areal extent and the polarization information provided by these three modes of data.

The RADARSAT-2 images covering the Changkai embankment and the city of Fuzhou were acquired on June 24, 26, and 27, 2010, and the images covering the submerged area around Dongting Lake were acquired on June 27 and July 2, 2010. The data are summarized in Table 1.

Table 1

The RADARSAT-2 images used for flood monitoring.

Only the SAR all-weather, day and night imaging capability made this dynamic flood monitoring possible. Due to the specular reflection of radar waves from water, the inundated area can be easily identified. After classifying the RADARSAT-2 images using a decision tree method, the inundated areas were extracted and then converted to vector format in ArcGIS. Figures 3Fig. 4Fig. 5 to 6 show the inundated area. By comparing the images from June 24 and 27 (Figs. 3 and 4), one can see that the flood area had decreased noticeably, but remained at some $107.69{\mathrm{km}}^2$. The flooded area between June 27 and July 2 (Figs. 5 and 6) around Dongting Lake increased by $29.32{\mathrm{km}}^2$.

Fig. 3

Flooded area around Changkai dyke on June 24, 2010.

Fig. 4

Flooded area around Changkai dyke on June 27, 2010.

Fig. 5

Flooded area around Dongting Lake on June 27, 2010.

Fig. 6

Flooded area around Dongting Lake on July 2, 2010.

After extracting the inundated land, we incorporated population information and land use data, and were able to evaluate the impact of the flood on human life and property.

The land cover layer was overlain on the inundation layer to define the intersection between these two layers. The inundation statistics were generated from the intersection layer. Table 2 shows the statistics of the inundated area, grouped by land cover type, for the city of Linchuan District and Dongxiang County, which are the two districts of Fuzhou City that were inundated. The statistics show that the paddy fields in the plain suffered the most severe flooding, with an area up to 77.49 and $16.10{\mathrm{km}}^2$ for Linchuan and Dongxiang County, respectively. Table 3 shows the population affected by the flood, with 214,500 people and 51,500 families affected. From these statistics, we classified this disaster as severe.

Table 2

Flooded area by land cover type for Linchuan District and Dongxiang County of Fuzhou City.

Table 3

Affected population for Linchuan District and Dongxiang County of Fuzhou City.

Due to the GEC, extreme weather events occur more frequently and are more disastrous.³⁸ Flood disasters, as one of the most severe, can be an indication of GEC. On June 21, 2010, rare heavy rains caused an extraordinary flood in Fuhe River, resulting in the collapse of the right bank. Using RADARSAT-2 data, we monitored the flood continuously.

3.2.

Estimation of Forest Height with PolInSAR

Forests respond significantly to environmental change. These changes may include increased ${\mathrm{CO}}_2$, global warming, and associated precipitation changes. Another category of change is habitat and population fragmentation. Forests are important sources and sinks of carbon, which is a predominant factor in GEC.

For homogeneous forests, the mean extinction coefficient can be retrieved based on some known tree heights, and this coefficient can characterize other regions of the forest. Then in the three-stage inversion process,³⁹ the known mean extinction coefficient can be used to estimate the forest height. Meanwhile, the look-up table (LUT) is simplified to one dimension (Fig. 7). For temporal decorrelation, some volume coherent points are outside the two-dimensional LUT, and the nearest point with zero extinction coefficients in the LUT will be searched for forest height inversion [Fig. 7(a)]. However, in the one-dimensional LUT, we can search the volume coherence based on the minimum phase difference, which can partly reduce the effect of decorrelation [Fig. 7(b)].

Fig. 7

Two-dimensional look-up table simplified to one dimension.

The performance of this approach is demonstrated using fully polarimetric data acquired by RADARSAT-2 over Tiantan Park in Beijing. There are sparse cypress trees in this park, and despite the relatively short wavelength, the radar signal can still penetrate the forest layer. The results indicate that the PolInSAR overestimates the tree height and the RADARSAT-2 is still a potential data for the PolInSAR applications (Fig. 8) in large area tree height quick mapping.

Fig. 8

(a) Forest height inversion in Tiantan Park using polarimetric synthetic aperture radar interferometry (PolInSAR). (b) Comparison between the ground measurement and PolInSAR.

3.3.

Spatiotemporal Distribution of the Shule River Alluvial Fan

There are many alluvial fans in the northern foot of Qilian Mountain, located in northwestern China. The largest one is the Shule River alluvial fan. Covering an area of $\sim 2400{\mathrm{km}}^2$, it is part of the super alluvial fan in the front of the mountains (Fig. 9). Its scale and location make it an ideal study area for remote sensing alluvial fan research.

Fig. 9

Location of the Shule River alluvial fan (Google map).

RADARSAT-2 fine mode PolSAR data acquired for the Shule River alluvial fan over the location is shown in the red rectangle in Fig. 9. The Pauli basis RGB image is shown in Fig. 10, where red, green, and blue represent HH-VV, 2HV, and HH+VV, respectively. Combined with the covering and cutting features of alluvial fan surfaces derived from SAR data and field survey data, it is generally considered that about five alluvial fan units from the Quaternary period are exposed in the study area. Three formed in the Late Pleistocene, Q3a, Q3b, and Q3c, respectively, as shown in Fig. 10, while the other two alluvial fan units formed in the Holocene, Q4a, and Q4b.

Fig. 10

RADARSAT-2 polarimetric SAR image of the Shule River alluvial fan.

Research results of an analysis of the relationship between the characteristic parameters and the alluvial fan units demonstrate that the following polarimetric parameters are sensitive to alluvial fan units of different ages: backscattering coefficients with HH, HV, and VV polarization states; the coherency entropy (H), and the anisotropy (A), derived using Cloude’s decomposition method. These are shown in Figs. 11(a) and 11(b). These parameters follow an increasing trend with increasing age in the Holocene period and then decrease quickly with increasing exposure time in the Late Pleistocene, while the highest value occurs in the alluvial fan unit, which was probably formed near the end of Late Pleistocene (Q3c).

Fig. 11

(a) Plot of backscattering coefficients for each alluvial fan unit. (b) Plot of H and A for each alluvial fan unit. (c) Spatial distribution of the Shule River alluvial fan.

Based on these characteristic parameters, the maximum likelihood classification method can be used to map the spatial distribution of the different alluvial fan units. The result is shown in Fig. 11(c) and indicates that the alluvial fan units can be effectively delimited using SAR imagery.

Alluvial fans in arid and semiarid regions provide an important record of neotectonics and climate change, which reveal the spatiotemporal evolution of the regional environment. SAR remote sensing, which is sensitive to geomorphological features, plays an important role in mapping alluvial fan surfaces.

3.4.

Impervious Surfaces

In our work, a total of 45 PolSAR features were extracted from RADARSAT-2 full polarimetric data (HH/HV/VH/VV) by the PolSARpro and the software developed by the authors. A case study of a dense urban area (Beijing, China, Fig. 12) was conducted by applying an SVM algorithm to RADARSAT-2 full PolSAR data. Two sample sites were selected to demonstrate this potential of SAR data. One site is in Beijing’s Olympic Park [Fig. 13(c)]; the other is in Beijing’s Imperial Palace [Fig. 13(d)]. The final polarimetric features are selected through a stepwise process (Fig. 1) listed in Table 4.

Fig. 12

Impervious surface extracted from the RADARSAT-2 full PolSAR data.

Fig. 13

Impervious surface maps derived from RADARSAT-2 data using the support vector machine (SVM) method in four selected sample sites corresponding to the four areas [(a), (b), (c), and (d)] in the figure. (a) Railroad estimation. (b) Asphalt road extraction. (c) and (d) Extraction of impervious surfaces covered by tree crowns.

Table 4

The final polarimetric features selected for classification through a stepwise process.

Figure 12 shows the impervious surfaces extracted from the RADARSAT-2 full PolSAR data, using the SVM method, in Beijing. In Fig. 12, the right image shows the percent impervious surfaces (PIS) with values ranging from 0 to 1, while the left image shows the percent impervious surfaces reclassified into six categories: high PIS (with values $>0.9$), medium PIS (with values between 0.5 and 0.9), low PIS (with values $<0.5$), bare soil, vegetation (forest and crop), and water. Figure 13 shows impervious surface maps derived from RADARSAT-2 data using the SVM method in four selected sample sites corresponding to the four areas [(a), (b), (c), and (d)] in Fig. 12. In Fig. 13, sample sites (a) and (b) highlight estimated railroads and asphalt roads, respectively, while sites (c) and (d) represent the extraction of impervious surfaces covered by tree crowns.

Figures 12 and 13 indicate the following:

In the PIS classification image, some crop classes were misclassified to bare soil in the image derived from RADARSAT-2 data (Fig. 13). That is a result of RADARSAT-2 PolSAR (C-band) having better canopy penetration as, on March 8, 2009, the main crop in the study area was small, short winter wheat, and C-band penetrates the wheat stalks and identifies bare soil. In a selected sample area [Fig. 13(a)], because of stronger backscatter in the SAR image, the railroads can be extracted from the full PolSAR data. In Fig. 13(c), the two left photos were taken on the ground, and in Fig. 13(d), the left image is a WorldView image (acquired on February 11, 2009) with a spatial resolution of 0.5 m. The right maps in Figs. 13(c) and 13(d) are urban impervious surface maps derived from RADARSAT-2 data using the SVM method. The impervious surfaces covered by tree crowns can be easily extracted from the RADARSAT-2 PolSAR data in sparse tree areas, while, because of the limited penetration, it is difficult for PolSAR data to extract the impervious surfaces covered by tree crowns in dense tree areas. Therefore, PolSAR imagery has enormous potential for extracting impervious surfaces covered by tree crowns in wooded areas, especially where there is sparse cover.

For RADARSAT-2 data, as a significant part of urban impervious surfaces, the entire road area (such as asphalt roads and railroads) might be either occluded or covered by layover from adjacent buildings or trees. Hence, roads in an along-track orientation sometimes cannot be seen at all in dense urban scenes. In summary, urban roads derived from RADARSAT-2 data are basically fully extracted; nonetheless, some gaps are visible in a number of roads [Fig. 13(b)]. Intrinsic characteristics of SAR imagery, i.e., layover and shadow effects, become a severe problem for extracting impervious surfaces, especially in dense urban areas. In our research, those characteristics were not considered.

To assess the accuracy of impervious surface estimation, a total of 100 sites of $100\mathrm{m}\times 100\mathrm{m}$ were randomly sampled. For each sample, the impervious surfaces were digitized on corresponding high-quality WorldView images with a spatial resolution of 0.5 m. Finally, three statistical indicators—root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination ($R^2$)—were used to evaluate the accuracy of the impervious surfaces percentage (ISP) estimation.

The results summarized in Table 5 show that RADARSAT-2 imagery allows for high accurate estimation. RADARSAT-2 yields results with an RMSE of 13.27%, an MAE of 9.98%, and an $R^2$ of 0.8421. Figure 14 shows the scatter plots of the accuracy assessment. This result also indicates that RADARSAT-2 has enormous potential for extracting impervious surfaces.

Table 5

Accuracy assessment of ISP estimation results.

Fig. 14

Accuracy assessment of impervious surface estimation.

Monitoring and managing the urban environment is an increasingly important issue across a range of disciplines. The environmental impacts of urbanization are mostly from different impervious surfaces. In recent years, impervious surfaces have become an important indicator in assessing urban ecological conditions because they are related to many environmental problems, such as water quality, stream health, the urban heat island effects, etc. Therefore, impervious surface-related environmental issues, as well as increased concerns about rapid urbanization worldwide, have resulted in a surge of research interest in impervious surfaces.