Shaunak De

Modifying the Yamaguchi Four-Component Decomposition Scattering Powers Using a Stochastic Distance

Model-based decompositions have gained considerable attention after the initial work of Freeman and Durden. This decomposition, which assumes the target to be reflectionsymmetric, was later relaxed in the Yamaguchi et al. decomposition with the addition of the helix parameter. Since then, many decomposition have been proposed where either the scattering model was modified to fit the data or the coherency matrix representing the second-order statistics of the full polarimetric data is rotated to fit the scattering model. In this paper, we propose to modify the Yamaguchi four-component decomposition (Y4O) scattering powers using the concept of statistical information theory for matrices. In order to achieve this modification, we propose a method to estimate the polarization orientation angle (OA) from full-polarimetric SAR images using the Hellinger distance. In this method, the OA is estimated by maximizing the Hellinger distance between the unrotated and the rotated T33 and the T22 components of the coherency matrix [T]. Then, the powers of the Yamaguchi four-component model-based decomposition (Y4O) are modified using the maximum relative stochastic distance between the T33 and the T22 components of the coherency matrix at the estimated OA. The results show that the overall double-bounce powers over rotated urban areas have significantly improved with the reduction of volume powers. The percentage of pixels with negative powers have also decreased from the Y4O decomposition. The proposed method is both qualitatively and quantitatively compared with the results obtained from the Y4O and the Y4R decompositions for a Radarsat-2 C-band San-Francisco dataset and an UAVSAR L-band Hayward dataset.

Selection of Suitable Window Size for Speckle Reduction and Deblurring using SOFM in Polarimetric SAR Images

Classification performance of PolSAR data, when used without speckle reduction is insufficient for most applications. Thus, speckle filtering becomes an essential preprocessing step. In this study we evaluate the effectiveness of different popular speckle filters and analyse their effects on the classification accuracy. We have used L-band and C-band fully polarimetric dataset acquired over Mumbai, India. The Wishart supervised classifier algorithm is used for classification of the filtered and unfiltered data. Boxcar, Refined Lee, Lopez, IDAN, Improved Sigma and sequential filters are analysed for the improvement in classification accuracy. Further we also evaluate the effect of window size on classification accuracy in order to be able to select appropriate window for speckle suppression. Boxcar and Refined Lee filters are used to test the effect of speckle filtering on classification with varying moving window size. Boxcar filter is widely used in the SAR application domain owing to it’s simplicity. However, the indiscriminate averaging of the Boxcar filter causes a resolution loss in the vicinity of sharp edges and point targets in the image. To overcome this, we have applied Kohonens Self-Organizing Feature Map (SOFM) algorithm to deblurr the image and improve edge and target preservation performance.

Comparative analysis of classification accuracy for RISAT-1 compact polarimetric data for various land-covers

The launch of RISAT-1 Indian remote sensing satellite on 26th April 2012, made it possible to collect hybrid polarimetric data from a space-borne sensor. The RISAT-1 C-band compact polarimetry data acquired over Mumbai is analyzed and assessed for classification of various land features and also compared with other fully polarimetric spaceborene SAR data sets. For better comparison, RISAT-1 C-band and RADARSAT-2 C-band simulated compact polarimetric data is classified and compared.

Urban classification using PolSAR data and deep learning

The urban classification of PolSAR images is made difficult by the characteristic of a rotated target to exhibit volume scattering. In this paper we use a deep learning technique in conjunction with some statistical parameters to learn to classify urban areas irrespective of the rotation. The learning algorithm was trained to differentiate urban from non-urban areas and was able to achieve a 8.5834% validation accuracy and 6.554% test accuracy.

Local Contrast Based Adaptive SAR Speckle Filter

In this paper, we propose an adaptive filtering technique for Synthetic Aperture Radar (SAR) images. A new windowing technique is introduced where the total window is divided into five equal sized overlapping sub-windows. The pixel to be filtered is a part of each of these sub-windows. A weighted mean of all sub-windows is computed for the pixel under consideration. The weights are accounted from a measure of heterogeneity calculated for each sub-windows. The filter is able to adapt automatically and adjust the speckle suppression strength based on local statistics. This allows the filter to preserve edges while strongly suppressing speckle over homogeneous areas. The proposed filter was compared with some well known SAR filtering techniques in terms of speckle suppression and edge preservation ability. Several experiments were performed on datasets acquired from both air-borne and space-borne SAR platforms. Some well known indices were used for quantitative comparison with other filters. Among the filters compared, the proposed filter shows good speckle suppression ability while still exhibiting reasonable edge preservation ability.

A Novel Technique Based on Deep Learning and a Synthetic Target Database for Classification of Urban Areas in PolSAR Data

The classification of urban areas in polarimetric synthetic aperture radar (PolSAR) data is a challenging task. Moreover, urban structures oriented away from the radar line of sight pose an additional complexity in the classification process. The characterization of such areas is important for disaster relief and urban sprawl monitoring applications. In this paper, a novel technique based on deep learning is proposed, which leverages a synthetic target database for data augmentation. The PolSAR dataset is rotated by uniform steps and collated to form a reference database. A stacked autoencoder network is used to transform the information in the augmented dataset into a compact representation. This significantly improves the generalization capabilities of the network. Finally, the classification is performed by a multilayer perceptron network. The modular architecture allows for easy optimization of the hyperparameters. The synthetic target database is created and the classification performance is evaluated on an L-band airborne UAVSAR dataset and L-band space-borne ALOS-2 dataset acquired over San Francisco, USA. The proposed technique shows an overall accuracy of

Variable importance and random forest classification using RADARSAT-2 PolSAR data

91.3{\%}

Predicting the popularity of instagram posts for a lifestyle magazine using deep learning

. An improvement over state-of-the-art techniques is achieved, especially in urban areas rotated away from the radar line of sight.

An unsupervised hidden Markov random field based segmentation of polarimetric SAR images

In this paper we have classified Polarimetric Synthetic Aperture Radar (PolSAR) data using the Random Forest (RF) classifier. The variables were ranked using the mean decrease in accuracy permutation method for each terrain class. RADARSAT-2 (RS-2) data acquired over Mumbai, India was used in this study. This technique is able to efficiently classify the dataset, as well as rank the parameters used in that classifier.

Snow wetness estimation from dual polarimetric coherent TerraSAR-X data

In this paper we use a Deep Neural Network (DNN) trained on data collected from the visual media-sharing social platform Instagram account of a popular Indian lifestyle magazine to predict the popularity of future posts. This predicted popularity of the post can be used to decide advertising rates and measure performance metrics important for publishing strategy decisions. The DNN primarily uses growth rate in subscriber base, tags associated with the post, time of day when the post was made, day of the week, color descriptors of the image, time between current and previous post, popularity of previous post as features for prediction. This covers majority of the causes of variation in popularity. Mini-batch gradient descend method is used to learn the weights in DNN and the objective function is cross-entropy. The network performs acceptable for real world applications and tolerances are within acceptable limits for the application.