Research

A Controlled Benchmark of Visual State-Space Backbones with Domain-Shift and Boundary Analysis for Remote-Sensing Segmentation

IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2026 Washington, DC, USA Accepted

Nichula Wasalathilaka, Dineth Perea, Oshadha Samarakoon, Buddhi Wijenayake, Roshan Godaliyadda, Vijitha Herath, Parakrama Ekanayake

This work presents a controlled benchmark of visual state-space model backbones for remote-sensing semantic segmentation. Instead of changing multiple pipeline components, the study fixes the decoder, training protocol, feature interface, loss, preprocessing, and evaluation setup while varying only the encoder backbone. The benchmark compares VMamba, MambaVision, and Spatial-Mamba on LoveDA and ISPRS Potsdam, analyzing in-domain accuracy, cross-domain robustness, boundary sensitivity, and practical efficiency. The results show that robustness under domain shift is backbone-dependent and asymmetric, while boundary delineation remains a major failure mode.

Mitigating Long-Tail Bias in LoveDA via Prompt-Controlled Diffusion Augmentation

IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2026 Washington, DC, USA Accepted

Buddhi Wijenayake, Nichula Wasalathilaka, Roshan Godaliyadda, Vijitha Herath, Parakrama Ekanayake, Vishal M. Patel

This work addresses long-tail class imbalance in LoveDA semantic segmentation using prompt-controlled diffusion augmentation. The framework first generates domain-aware semantic layouts with class-ratio control using discrete diffusion, then translates those layouts into photorealistic remote-sensing images using Stable Diffusion with ControlNet guidance. By mixing the generated Urban and Rural image-label pairs with real training data, the method improves segmentation performance across multiple backbones, with gains concentrated on minority classes and cross-domain generalization.

Mamba-FCS: Joint Spatio-Frequency Feature Fusion, Change-Guided Attention, and SeK Inspired Loss for Enhanced Semantic Change Detection in Remote Sensing

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Published

Buddhi Wijenayake, Athulya Ratnayake, Praveen Sumanasekara, Roshan Godaliyadda, Parakrama Ekanayake, Vijitha Herath, Nichula Wasalathilaka

Mamba-FCS is a semantic change detection framework for remote sensing imagery built on a Visual State Space Model backbone. It introduces joint spatio-frequency feature fusion to combine spatial and frequency-domain cues, a Change-Guided Attention module to connect binary and semantic change detection, and a SeK-inspired loss to improve learning under class imbalance. Experiments on SECOND and Landsat-SCD show improved semantic change detection performance, especially for boundary refinement and rare land-cover transitions.

Precision Spatio-Temporal Feature Fusion for Robust Remote Sensing Change Detection

2025 IEEE 19th International Conference on Industrial and Information Systems (ICIIS) Kandy, Sri Lanka Published

Buddhi Wijenayake, Athulya Ratnayake, Praveen Sumanasekara, Nichula Wasalathilaka, Mathivathanan Piratheepan, Roshan Godaliyadda, Mervyn Ekanayake, Vijitha Herath

This work improves remote sensing binary change detection by extending ChangeMamba with precision fusion blocks for fine-grained temporal comparison, an enhanced decoder for preserving local detail, and a combined Cross Entropy, Lovasz, and Dice loss for class-imbalanced change masks. The method was evaluated on SYSU-CD, LEVIR-CD+, and WHU-CD, showing stronger change detection performance across multiple metrics.

Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification

MICCAI 2026 Strasbourg, France Accepted

Nichula Wasalathilaka, Abhijit Das, Imran Razzak, Dwarikanath Mahapatra

Research conducted at MBZUAI, Abu Dhabi · Supervised by Prof. Imran Razzak & Prof. Dwarikanath Mahapatra

Graph-of-Differences introduces an anatomy-structured framework for medical image re-identification. Instead of comparing arbitrary visual tokens, the method represents each image as an anatomy graph with named anatomical nodes, establishes soft node correspondence across image pairs, and aligns anatomy-matched differences with the global backbone difference signal. This makes retrieval decisions more auditable by producing explanations over named anatomical structures rather than unstable pixel-level heatmaps. Experiments on fundus and chest X-ray benchmarks show improved Rank-1 and mAP over a strong frozen-backbone baseline, with additional zero-shot external transfer gains.

PROTON: Prototype-based Test-time Online OOD Detection for Medical Vision-Language Models

MICCAI 2026 Strasbourg, France Accepted

Abhijit Das, Nichula Wasalathilaka, Yifan Lu, Adinath Dukre, Dwarikanath Mahapatra, Shadab Khan, Imran Razzak

Research conducted at MBZUAI, Abu Dhabi · Supervised by Prof. Imran Razzak, Prof. Dwarikanath Mahapatra & Prof. Shadab Khan

Medical VLMs enable zero-shot clinical image classification, yet reliably detecting out-of-distribution (OOD) inputs at deployment remains an open problem. PROTON (PROtotype-based Test-time ONline OOD detection) is a lightweight post-hoc module that maintains an online prototype bank from high-confidence test predictions and adaptively fuses prototype distance with MCM scoring via stream-level variance statistics — requiring no model modification, training data, or prompt engineering. The approach exploits a structural mismatch overlooked by static scoring methods: covariate-shifted inputs are indistinguishable from in-distribution samples in softmax space, yet occupy distinct regions in the VLM's embedding space. On the full-spectrum ophthalmology benchmark (FLAIR + FIVES), PROTON improves MCM by +23.9 AUROC points on covariate shift, +8.8 on semantic shift, and +8.1 on far-OOD — the only zero-shot method to improve across all three shift types without hierarchical prompts or labeled data.

RespGeomLib: A Reproducible Parametric Engine for Generating Analysis-Ready Human Airway Lumen Geometry

IEEE Mercon 2026 Accepted

Nichula Wasalathilaka, Parakrama Ekanayaka, Roshan Godaliyadda

CT-derived airway models support pulmonary morphometry and airflow simulation, but are often limited by distal scan resolution and the need for substantial cleanup near bifurcations. Procedural alternatives are reproducible, yet many rely on stitched tubular primitives that introduce non-smooth junctions and poorly defined open boundaries. RespGeomLib is a reproducible parametric engine for generating analysis-ready human airway lumen surfaces from compact YAML specifications. The framework combines port-based assembly with implicit smooth-min junction blending to produce seamless junctions, while avoiding full-tree voxelization through analytic segments and local implicit extraction around bifurcations. RespGeomLib yields cleaner junctions than a Boolean/stitch baseline and is substantially faster and more memory-efficient than whole-tree global implicit extraction. The engine further supports morphometry-guided tree generation, controlled synthetic airway variants, and CFD-ready export with stable airflow simulation.

Project website: nichula01.github.io/Respgeomlib