Subash Khanal, PhD

Broadly, I am interested in building scalable AI systems that bridge multiple modalities to address real-world challenges.

Email  /  CV  /  Google Scholar  /  Linkedin  /  Github

Research

My research focuses on multimodal self-supervised learning and generative AI. I build models that integrate satellite imagery, audio, and text, and curate large-scale datasets to support applications in geospatial AI, medicine, and ecology.

Selected Publications

For a complete list, see my Google Scholar profile.

We conduct the first systematic evaluation of segmentation and geospatial foundation models for global field boundary delineation using the Fields of The World (FTW) benchmark. We propose PRUE, combining a U-Net backbone, composite loss functions, and targeted data augmentations to achieve 76% IoU and 47% object-F1 on FTW, surpassing the previous baseline by 6% and 9% respectively.

A state-of-the-art soundscape mapping framework that leverages a Vision-Language Model (VLM) to enrich the semantic understanding of a location's soundscape, learns a shared codebook for fine-grained alignment, and enables retrieval-based, location-conditioned soundscape generation.

ProM3E is a probabilistic masked multimodal embedding model for any-to-any generation of multimodal representations for ecology. It learns to infer missing modalities given a few context modalities and proposes a novel cross-modal retrieval approach that mixes inter-modal and intra-modal similarities to achieve superior performance across all retrieval tasks.

SimLBR is a simple and efficient framework for fake image detection using Latent Blending Regularization (LBR). By learning a tight decision boundary around the real image distribution and treating the fake category as a sink class, it significantly improves cross-generator generalization, achieving up to +24.85% accuracy and +69.62% recall on the challenging Chameleon benchmark.

We introduce Radial Cross-Modal Embeddings (RCME), a framework for learning hierarchical vision-language representations that explicitly models transitivity-enforced entailment. Applied to the Tree of Life taxonomy, our model achieves state-of-the-art performance on hierarchical species classification and image-text retrieval tasks.

We introduce mixed-view panorama synthesis — generating a novel ground-level panorama conditioned on a satellite image and a set of nearby panoramas. Our diffusion-based model with geospatially-guided attention handles sparse or distant panorama inputs and consistently outperforms cross-view and same-view synthesis baselines.

We propose a novel retrieval-augmented strategy for multi-resolution geo-embeddings, called RANGE. Our method is based on the intuition that the visual features of a location can be estimated by aggregating visual features from multiple similar-looking locations.

TaxaBind is a suite of multimodal models useful for downstream ecological tasks covering six modalities: ground-level image, geographic location, satellite image, text, audio, and environmental features.

We introduce a language-driven approach for hierarchical species distribution modeling (SDM) that uses an LLM to encode species representations from taxonomic descriptions. This enables range map prediction at multiple taxonomic levels and zero-shot generalization to unseen species.

We develop a probabilistic, multi-scale, and metadata-aware embedding space that connects audio, text, and overhead imagery. This enables the creation of dynamic, multi-scale soundscape maps for any geographic region, along with uncertainty estimates for the mapping.

We propose Sat2Cap, a contrastive learning framework that aligns satellite imagery with fine-grained textual descriptions of locations. Trained on a novel large-scale dataset, it enables zero-shot geospatial mapping driven by free-form text queries.

GeoSynth is a diffusion-based model for synthesizing high-resolution satellite images with global style control via text prompts or geographic location, and image-driven layout control via OpenStreetMap data. It exhibits strong zero-shot generalization and produces diverse, geographically consistent satellite imagery.

BirdSAT unifies cross-view contrastive learning and masked autoencoders to jointly learn from paired ground-level bird images and satellite imagery. The resulting model supports both fine-grained bird species classification and geographic species distribution mapping.

We learn a tri-modal embedding space between audio, text and overhead imagery. This enables us to create soundscape maps over any geographic region, using either audio or textual queries.

We introduce CAT-XPLAIN, an inherently interpretable Vision Transformer that adds a causal selection token trained to identify the most causally significant image patches for the classification decision, eliminating the need for post-hoc explainers while maintaining strong task performance.

This website is modified from source code of John Barron's website.