Bhogendra Mishra

Sensitivity Analysis for L-Band Polarimetric Descriptors and Fusion for Urban Land Cover Change Detection

A fully polarimetric synthetic aperture radar (PolSAR) image allows the generation of a number of polarimetric descriptors. These descriptors are sensitive to changes in land use and cover. Thus, the objective of this study is twofold: first, to identify the most effective descriptors for each change type and ascertain the best complementary pairs from the selected polarimetric descriptors; and second, to develop an information fusion approach to use the unique features found in each polarimetric descriptor to obtain a better change map for urban and suburban environments. The effectiveness of each descriptor was assessed through statistical analysis of the sensitivity index in selected areas and through change detection results obtained by using the supervised thresholding method. A good agreement was found between the statistical analysis and the performance of each descriptor. Finally, a polarimetric information fusion method based on the coupling of modified thresholding with a region-growing algorithm was implemented for the identified complementary descriptor pairs. The mapping accuracy, as measured by the Kappa coefficient, was improved by 0.09 (from 0.76 to 0.85) with a significant reduction of false and missing alarm rates compared to using single PolSAR images.

COUPLING OF THRESHOLDING AND REGION GROWING ALGORITHM FOR CHANGE DETECTION IN SAR IMAGES

In this research paper, we propose supervised and unsupervised change detection methodologies focused on the analysis of multitemporal Synthetic Aperture Radar (SAR) images. These approaches are based on three main steps: (1) a comparison of multitemporal image was carried out by normalized difference ratio (NDR) operator; (2) implementing a novel supervised or unsupervised thresholding and (3) generating the change map by coupling of thresholding along with a region growing algorithm. In the first step, two filtered multitemporal images were used to generate NDR image that was subjected to analysis. In the second step, by assuming a Gaussian distribution in the no-change area, we identified the pixel range that fits the Gaussian distribution better than any other range iteratively to detect the no-change area that eventually separates the change areas. In the supervised method, several sample no-change pixels were selected and the mean (μ) and the standard deviation (σ) were obtained. Then, μ ± 3σ was applied to select the best threshold values. Finally, a traditional thresholding algorithm was modified and implemented with the coupling of the region growing algorithm to consider the spatial information to generate the change map. The Gaussian distribution was assumed because it better fits the conditional densities of the no-change class in the NDR image. The effectiveness of the proposed methods was verified with the simulated images and the real images associated with geographical locations. The results were compared with the manual trial and error procedure (MTEP) and traditional unsupervised expectation-maximization (EM) method. Both proposed methods gave similar results with MTEP and significant improvement in Kappa coefficient in comparison to the traditional EM method was found in both geographical locations.

Generation of pseudo-fully polarimetric data from dual polarimetric data for land cover classification

A linear relationship among the HH, HV, and VV components of polarimetric synthetic aperture radar (SAR) data is studied. A regression model was developed to predict the real and imaginary parts of the VV polarimetric component from the HH and HV components in dual polarimetric SAR and the resulting dataset is called pseudo-fully polarimetric SAR data. Freeman-Wishart classification was applied to evaluate the preservation of scattering characteristics in the pseudo-fully polarimetric dataset. A kappa coefficient is 0.81 indicates very good agreement between the two classification results. An SVM was used for the land cover classification. Finally, post-processing was implemented to remove noise in the form of isolated pixels. A VNIR-2 optical data taken over the same area at nearly same time was used as ground truth data to assess the classification accuracy. The land cover classification result obtained from the SVM shows that using the pseudo-fully polarimetric data gives more than a 2% improvement of mean producers accuracy over dual polarimetric datasets.

SAR and optical data fusion for land use and cover change detection

This work presents a very simple but robust, synthetic aperture radar (SAR) and optical, data fusion framework for land use/cover change detection. The fusion was done with two indicators, namely the normalized difference ratio (NDR) and normalized difference vegetation index difference (NDVI difference) developed from multitemporal SAR and optical images respectively. A statistical analysis shows that the NDR and the NDVI difference have a consistent pattern in major land use/cover change classes. Thus, based on this pattern, a fusion approach was developed without altering the behavior of NDR with different types of changes. The effectiveness of the proposed fusion approach was evaluated through the change mapping with a manual trial and error thresholding approach. The results were compared with the results obtained from the optical and SAR images independently. The improvement of the results by making use of the the unique information from both, optical and SAR imagery, can be easily identified with a simple visual inspection. The accuracy assessment showed a significant improvement in overall detectability with the substantial decrease in false and missing alarms.

Development of method to automatically select passpoints for close range photogrammetry in dense urban areas

This paper proposes a methodology to select passpoints from images obtained using close-range photogrammetry. In this method, matches on stereo-pairs are extracted using Scale-Invariant Feature transform (SIFT) (Lowe, 2004), and erroneous matches are removed using Random Sample Consensus (RANSAC) (Fischler et al, 1981). At this stage, a lot of candidates of passpoints are available, and the number of points should be reduced to save computation time for external orientation. Therefore, limited number of passpoints are selected in terms of spatial dispersion by using the score obtained by SIFT. As a result, almost all erroneous matches could be removed and the validation result demonstrated that the accuracy of segment lengths was acceptable.

Estimation of enclosure index in urban areas using airborne LiDAR

This paper proposes a methodology to estimate an “enclosure index” in urban areas using airborne LiDAR. The index is defined as a ratio of occluded area to whole of area in azimuth angle-elevation angle space. The index can be applied to assess local landscape, and it is expected to estimate it in a wide area at low cost. The author examined the methodology to estimate it using airborne LiDAR data measured in last-pulse mode. The estimated index map was validated with the ground truth data, and the error was acceptable, approximately 3%. Even though the last-pulse mode data may underestimate actual digital surface model (DSM), it was found that the proposed methodology is effective to estimate the index in a wide area at low cost.

Unsupervised change detection in an urban environment using multitemporal polsar images

In this paper we address the problem of change detection in multi temporal polarimetric synthetic aperture radar (PolSAR) images of an urban environment. To compare PolSAR images acquired on two different dates, a change image was produced by using a modified ratio operator and normalized difference ratio operator. The amplitudes of three polarimetric components (HH, HV and VV) and the diagonal elements of the coherency matrix (T11, T22 and T33) from both dates were used to generate the change image. A thresholding algorithm based on histogram fitting was then implemented to automatically classify the change image into two classes: change and no-change. Experiments were carried out on two sets of multitemporal images acquired by ALOS PALSAR to confirm the effectiveness of the proposed unsupervised approach. The combination of the normalized difference ratio operator with the diagonal elements of the coherency matrix is better suited for change detection than any other combination.

Automatic thresholding for land cover change detection in SAR images

In this research, we propose an unsupervised change detection methodology for synthetic aperture radar (SAR) images. By analyzing the goodness of fit, it is found that pixels in no change area in the change image generated by normalized difference ratio (NDR) operator better fit with normal distribution. Based on this assumption, we identify the pixel range that fits the normal distribution better than any other range iteratively in the image and the range is defines as the threshold value. Finally, a region growing segmentation algorithm was modified to fit for the post processing in change detection. Experiments were carried out with all possible cases of changes: (i) double change, (ii) single change and (iii) no change to prove the effectiveness of the proposed methodology.

Land cover classification comparisons among dual polarimetric, pseudo-fully polarimetric, and fully polarimetric SAR imagery

In this paper, an approach is proposed that predicts fully polarimetric data from dual polarimetric data, and then applies selected supervised algorithm for dual polarimetric, pseudo-fully polarimetric and fully polarimetric dataset for the land cover classification comparison. A regression model has been developed to predict the complex variables of VV polarimetric component and amplitude independently using corresponding complex variables and amplitude in HH and HV bands. Support vector machine (SVM)is implemented for the land cover classification. Coherency matrix and amplitude were used for all dataset for the land cover classification independently.They are used to compare the data from different perspective. Finally, a post processing technique is implemented to remove the isolated pixels appeared as a noise. AVNIR-2 optical data over the same area is used as ground truth data to access the classification accuracy.The result from SVM indicates that the fully polarimetric mode gives the maximum classification accuracy followed by pseudo-fully polarimetric and dual polarimetric datasets using coherency matrix input for fully polarimetric image and pseudo-fully polarimetric image and covariance matrix input for dual polarimetric image. Additionally, it is observed that pseudo-fully polarimetric image with amplitude input does not show the significant improvement over dual polarimetric image with same input.