Selected work

Representative Papers

Towards Seamless Lunar Mosaics: Deep Radiometric Normalization for Cross-Sensor Orbital Imagery Using Chandrayaan-2 TMC Data

Apr 28, 2026

This study addresses the challenge of radiometric inconsistency in multi-mission lunar orbital imagery caused by variations in illumination, sensors, and imaging conditions, which hinders seamless mosaicking. The work proposes the first cross-mission radiometric normalization framework based on conditional generative adversarial networks (cGANs), mapping Chandrayaan-2 Terrain Mapping Camera (TMC) data—augmented with SELENE imagery—as inputs to the Lunar Reconnaissance Orbiter Camera Wide Angle Camera (LROC WAC) reference standard. The approach employs a U-Net generator and a PatchGAN discriminator, combined with a patch-based training scheme and an overlap-aware inference strategy to preserve large-scale structural continuity while effectively eliminating visible seams. Experimental results demonstrate that the method significantly outperforms conventional histogram matching in terms of SSIM, PSNR, and RMSE metrics, substantially improving tonal uniformity and structural consistency across multi-source lunar images.

0 citationsRead paper

DEM Refinement and Validation on the Lunar Surface Using Shape-from-Shading with Chandrayaan-2 OHRC Imagery

Apr 19, 2026

Existing sub-meter lunar digital elevation models (DEMs) suffer from insufficient detail due to limitations in stereo imaging baselines. This study addresses this constraint by proposing a novel framework that treats shape-from-shading (SfS) as an independent source of topographic information, leveraging high-resolution imagery from the Chang’e-2 Orbiter High-Resolution Camera (OHRC) to refine DEMs without relying on stereo constraints. Through a three-stage smoothness-weight parameter sweep and systematic sensitivity analysis, the method significantly improves the statistical accuracy of surface slope estimates, reveals previously unresolved fine-scale crater morphologies, and elucidates the spatially variable impact of large incidence angle differences and local image coverage on enhancement quality.

0 citationsRead paper

Deep Learning Aided Vision System for Planetary Rovers

Mar 26, 2026

This work addresses the challenges of insufficient real-time perception accuracy and low-fidelity offline 3D reconstruction for planetary rovers operating in complex terrains by proposing a lightweight visual system that integrates real-time neural depth estimation with offline monocular depth reconstruction. The system combines CLAHE image enhancement, YOLOv11n object detection, a custom distance estimation network, and the Depth Anything V2 model to generate high-fidelity dense point clouds, which are subsequently fused using Open3D. Experimental results demonstrate a median depth error of only 2.26 cm within the 1–10 meter range and achieve a balanced trade-off between detection precision and recall on lunar grayscale imagery, thereby validating the method’s effectiveness in metric accuracy, geometric detail preservation, and computational efficiency.

0 citationsRead paper

Beyond Spherical geometry: Unraveling complex features of objects orbiting around stars from its transit light curve using deep learning

Sep 18, 2025

This work addresses the ill-posedness of reconstructing non-spherical celestial body geometries from transit light curves. We propose an end-to-end reconstruction method integrating Fourier-based elliptical harmonic decomposition with deep learning. Specifically, arbitrary 2D contours are represented as elliptical harmonic series, and a dedicated neural network is trained on synthetic light curves generated by the Yuti simulator to directly map photometric signals to shape parameters. Experiments demonstrate high-fidelity recovery of low-order elliptical contours—including global morphology and orientation—while higher-order features exhibit reliable scale estimation but inherent degeneracies in eccentricity and orientation. Crucially, this study provides the first systematic characterization of the theoretical limits of geometric information recoverable from transit photometry. The framework establishes a new paradigm for characterizing non-spherical astrophysical objects—such as exomoons, fragmented asteroids, and ringed bodies—based solely on light-curve observations.

0 citationsRead paper

Population Estimation using Deep Learning over Gandhinagar Urban Area

Sep 16, 2025

Traditional censuses suffer from high costs, long intervals, and heavy reliance on manual labor. To address these limitations, this study proposes a deep learning–based urban population estimation method leveraging multi-source remote sensing data. Specifically, we integrate high-resolution satellite imagery, digital elevation models (DEMs), and vector administrative/building boundary layers into a hybrid CNN-ANN architecture: convolutional neural networks (CNNs) extract fine-grained morphological and textural features for building classification, while artificial neural networks (ANNs) fuse topographic and spatial contextual information to estimate resident population per individual building. Evaluated on a dataset of 48,000 buildings in Gandhinagar, India, the method achieves a building classification F1-score of 0.9936 and estimates the city’s total population as 278,954—within acceptable error bounds. This approach significantly enhances the efficiency, scalability, and spatial resolution of population mapping, enabling low-cost, high-frequency, fine-grained demographic intelligence for resource allocation and smart city development.

0 citationsRead paper

Recent publications

Latest Papers

Towards Seamless Lunar Mosaics: Deep Radiometric Normalization for Cross-Sensor Orbital Imagery Using Chandrayaan-2 TMC Data

Apr 28, 2026

This study addresses the challenge of radiometric inconsistency in multi-mission lunar orbital imagery caused by variations in illumination, sensors, and imaging conditions, which hinders seamless mosaicking. The work proposes the first cross-mission radiometric normalization framework based on conditional generative adversarial networks (cGANs), mapping Chandrayaan-2 Terrain Mapping Camera (TMC) data—augmented with SELENE imagery—as inputs to the Lunar Reconnaissance Orbiter Camera Wide Angle Camera (LROC WAC) reference standard. The approach employs a U-Net generator and a PatchGAN discriminator, combined with a patch-based training scheme and an overlap-aware inference strategy to preserve large-scale structural continuity while effectively eliminating visible seams. Experimental results demonstrate that the method significantly outperforms conventional histogram matching in terms of SSIM, PSNR, and RMSE metrics, substantially improving tonal uniformity and structural consistency across multi-source lunar images.

0 citationsRead paper

DEM Refinement and Validation on the Lunar Surface Using Shape-from-Shading with Chandrayaan-2 OHRC Imagery

Apr 19, 2026

Existing sub-meter lunar digital elevation models (DEMs) suffer from insufficient detail due to limitations in stereo imaging baselines. This study addresses this constraint by proposing a novel framework that treats shape-from-shading (SfS) as an independent source of topographic information, leveraging high-resolution imagery from the Chang’e-2 Orbiter High-Resolution Camera (OHRC) to refine DEMs without relying on stereo constraints. Through a three-stage smoothness-weight parameter sweep and systematic sensitivity analysis, the method significantly improves the statistical accuracy of surface slope estimates, reveals previously unresolved fine-scale crater morphologies, and elucidates the spatially variable impact of large incidence angle differences and local image coverage on enhancement quality.

0 citationsRead paper

Deep Learning Aided Vision System for Planetary Rovers

Mar 26, 2026

This work addresses the challenges of insufficient real-time perception accuracy and low-fidelity offline 3D reconstruction for planetary rovers operating in complex terrains by proposing a lightweight visual system that integrates real-time neural depth estimation with offline monocular depth reconstruction. The system combines CLAHE image enhancement, YOLOv11n object detection, a custom distance estimation network, and the Depth Anything V2 model to generate high-fidelity dense point clouds, which are subsequently fused using Open3D. Experimental results demonstrate a median depth error of only 2.26 cm within the 1–10 meter range and achieve a balanced trade-off between detection precision and recall on lunar grayscale imagery, thereby validating the method’s effectiveness in metric accuracy, geometric detail preservation, and computational efficiency.

0 citationsRead paper

Beyond Spherical geometry: Unraveling complex features of objects orbiting around stars from its transit light curve using deep learning

Sep 18, 2025

This work addresses the ill-posedness of reconstructing non-spherical celestial body geometries from transit light curves. We propose an end-to-end reconstruction method integrating Fourier-based elliptical harmonic decomposition with deep learning. Specifically, arbitrary 2D contours are represented as elliptical harmonic series, and a dedicated neural network is trained on synthetic light curves generated by the Yuti simulator to directly map photometric signals to shape parameters. Experiments demonstrate high-fidelity recovery of low-order elliptical contours—including global morphology and orientation—while higher-order features exhibit reliable scale estimation but inherent degeneracies in eccentricity and orientation. Crucially, this study provides the first systematic characterization of the theoretical limits of geometric information recoverable from transit photometry. The framework establishes a new paradigm for characterizing non-spherical astrophysical objects—such as exomoons, fragmented asteroids, and ringed bodies—based solely on light-curve observations.

0 citationsRead paper

Population Estimation using Deep Learning over Gandhinagar Urban Area

Sep 16, 2025

Traditional censuses suffer from high costs, long intervals, and heavy reliance on manual labor. To address these limitations, this study proposes a deep learning–based urban population estimation method leveraging multi-source remote sensing data. Specifically, we integrate high-resolution satellite imagery, digital elevation models (DEMs), and vector administrative/building boundary layers into a hybrid CNN-ANN architecture: convolutional neural networks (CNNs) extract fine-grained morphological and textural features for building classification, while artificial neural networks (ANNs) fuse topographic and spatial contextual information to estimate resident population per individual building. Evaluated on a dataset of 48,000 buildings in Gandhinagar, India, the method achieves a building classification F1-score of 0.9936 and estimates the city’s total population as 278,954—within acceptable error bounds. This approach significantly enhances the efficiency, scalability, and spatial resolution of population mapping, enabling low-cost, high-frequency, fine-grained demographic intelligence for resource allocation and smart city development.

0 citationsRead paper