Selected work

Representative Papers

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations

Jun 24, 2026

This study addresses the lack of high-resolution, fine-scale quantification of urban tree biomass, which hinders the characterization of individual-tree heterogeneity. The authors propose a self-supervised dual-stream cross-attention network that fuses airborne LiDAR with near-infrared RGB imagery to generate semantic labels, enabling annotation-free crown delineation through multiscale watershed segmentation. Aboveground biomass is then estimated using species-specific allometric equations. The work introduces the first publicly available bitemporal non-forest tree biomass database and incorporates deep ensemble uncertainty maps to guide model refinement. On an independent test set, biomass predictions achieve R² values of 0.570–0.609. Applied to an 810 km² area in Ontario from 2018 to 2023, the approach reveals a net carbon stock increase of 39 Gg C, with localized densities reaching up to 140 Mg/ha.

0 citationsRead paper

Continuous biome representations from Earth observation embeddings

Jun 09, 2026

Traditional biome maps rely on discrete classifications, which struggle to capture the continuous gradients characteristic of ecological transition zones. This work proposes a continuous representation method leveraging embeddings from the Clay v1.5 foundation model trained on satellite imagery: a linear classifier is trained on its 1024-dimensional Earth observation embeddings to output softmax-normalized biome probability vectors, thereby preserving categorical semantics while explicitly modeling gradual transitions between biomes. Evaluated across six major Brazilian biomes, this continuous representation achieves a mean AUC of 0.618 in predicting species distributions—significantly outperforming discrete biome labels (AUC = 0.570)—and demonstrates consistently improved performance across varying distances from biome boundaries.

0 citationsRead paper

WATCH: Wide-Area Archaeological Site Tracking for Change Detection

May 04, 2026

This study addresses the challenge of precisely dating disturbance events at archaeological sites, which is hindered by subtle visual cues and sparse labels. Leveraging PlanetScope satellite imagery, the authors propose three complementary monthly change detection approaches: an unsupervised Temporal Embedding Distance (TED), a self-supervised change detection (SSCD) method, and a weakly supervised temporal localization model. For the first time, they integrate embeddings from six foundation models—including CLIP, GeoRSCLIP, and SatMAE—with handcrafted spectral-textural features and introduce a multi-strategy scoring mechanism. Evaluated on 1,943 sites in Afghanistan, the framework achieves a 55% exact-month recall when combining TED with SatMAE, while GeoRSCLIP and related models attain a 92.5% recall within a ±3-month window. Notably, SSCD demonstrates superior early-warning capability among the proposed methods.

0 citationsRead paper

A Multi-Agent Feedback System for Detecting and Describing News Events in Satellite Imagery

Apr 14, 2026

Existing approaches struggle to efficiently construct multi-temporal remote sensing event captioning datasets due to the high cost of manually identifying visible events and annotating corresponding image sequences. To address this challenge, this work proposes SkyScraper, an iterative multi-agent feedback system that automatically discovers and annotates remote sensing events by geocoding news articles, retrieving matching satellite image sequences, and generating descriptive image-text pairs. The proposed method substantially improves event discovery efficiency, yielding five times more events than conventional approaches, and enables the creation of the first large-scale multi-temporal remote sensing event captioning dataset, comprising 5,000 annotated sequences. This resource provides a critical foundation for interdisciplinary applications bridging remote sensing and news analysis.

0 citationsRead paper

Satellite-Based Detection of Looted Archaeological Sites Using Machine Learning

Feb 23, 2026

This study addresses the challenge of monitoring looting at archaeological sites in remote regions, a critical threat to cultural heritage preservation. The authors develop a scalable detection pipeline leveraging monthly PlanetScope mosaics and multi-temporal data from 1,943 sites in Afghanistan, augmented with spatial masks to isolate site-specific signals. They systematically evaluate the performance of end-to-end convolutional neural networks (CNNs), random forests, SatCLIP-V geovisual embeddings, and handcrafted spectral/textural features for looting identification. The work presents the first validation of ImageNet pretraining and spatial masking efficacy under domain shift in remote sensing contexts. The best-performing model achieves an F1 score of 0.926, substantially outperforming conventional approaches (F1 = 0.710), demonstrating that looting signatures are highly localized and that deep learning models possess superior discriminative capacity for this task.

0 citationsRead paper

Recent publications

Latest Papers

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations

Jun 24, 2026

This study addresses the lack of high-resolution, fine-scale quantification of urban tree biomass, which hinders the characterization of individual-tree heterogeneity. The authors propose a self-supervised dual-stream cross-attention network that fuses airborne LiDAR with near-infrared RGB imagery to generate semantic labels, enabling annotation-free crown delineation through multiscale watershed segmentation. Aboveground biomass is then estimated using species-specific allometric equations. The work introduces the first publicly available bitemporal non-forest tree biomass database and incorporates deep ensemble uncertainty maps to guide model refinement. On an independent test set, biomass predictions achieve R² values of 0.570–0.609. Applied to an 810 km² area in Ontario from 2018 to 2023, the approach reveals a net carbon stock increase of 39 Gg C, with localized densities reaching up to 140 Mg/ha.

0 citationsRead paper

Continuous biome representations from Earth observation embeddings

Jun 09, 2026

Traditional biome maps rely on discrete classifications, which struggle to capture the continuous gradients characteristic of ecological transition zones. This work proposes a continuous representation method leveraging embeddings from the Clay v1.5 foundation model trained on satellite imagery: a linear classifier is trained on its 1024-dimensional Earth observation embeddings to output softmax-normalized biome probability vectors, thereby preserving categorical semantics while explicitly modeling gradual transitions between biomes. Evaluated across six major Brazilian biomes, this continuous representation achieves a mean AUC of 0.618 in predicting species distributions—significantly outperforming discrete biome labels (AUC = 0.570)—and demonstrates consistently improved performance across varying distances from biome boundaries.

0 citationsRead paper

WATCH: Wide-Area Archaeological Site Tracking for Change Detection

May 04, 2026

This study addresses the challenge of precisely dating disturbance events at archaeological sites, which is hindered by subtle visual cues and sparse labels. Leveraging PlanetScope satellite imagery, the authors propose three complementary monthly change detection approaches: an unsupervised Temporal Embedding Distance (TED), a self-supervised change detection (SSCD) method, and a weakly supervised temporal localization model. For the first time, they integrate embeddings from six foundation models—including CLIP, GeoRSCLIP, and SatMAE—with handcrafted spectral-textural features and introduce a multi-strategy scoring mechanism. Evaluated on 1,943 sites in Afghanistan, the framework achieves a 55% exact-month recall when combining TED with SatMAE, while GeoRSCLIP and related models attain a 92.5% recall within a ±3-month window. Notably, SSCD demonstrates superior early-warning capability among the proposed methods.

0 citationsRead paper

A Multi-Agent Feedback System for Detecting and Describing News Events in Satellite Imagery

Apr 14, 2026

Existing approaches struggle to efficiently construct multi-temporal remote sensing event captioning datasets due to the high cost of manually identifying visible events and annotating corresponding image sequences. To address this challenge, this work proposes SkyScraper, an iterative multi-agent feedback system that automatically discovers and annotates remote sensing events by geocoding news articles, retrieving matching satellite image sequences, and generating descriptive image-text pairs. The proposed method substantially improves event discovery efficiency, yielding five times more events than conventional approaches, and enables the creation of the first large-scale multi-temporal remote sensing event captioning dataset, comprising 5,000 annotated sequences. This resource provides a critical foundation for interdisciplinary applications bridging remote sensing and news analysis.

0 citationsRead paper

Satellite-Based Detection of Looted Archaeological Sites Using Machine Learning

Feb 23, 2026

This study addresses the challenge of monitoring looting at archaeological sites in remote regions, a critical threat to cultural heritage preservation. The authors develop a scalable detection pipeline leveraging monthly PlanetScope mosaics and multi-temporal data from 1,943 sites in Afghanistan, augmented with spatial masks to isolate site-specific signals. They systematically evaluate the performance of end-to-end convolutional neural networks (CNNs), random forests, SatCLIP-V geovisual embeddings, and handcrafted spectral/textural features for looting identification. The work presents the first validation of ImageNet pretraining and spatial masking efficacy under domain shift in remote sensing contexts. The best-performing model achieves an F1 score of 0.926, substantially outperforming conventional approaches (F1 = 0.710), demonstrating that looting signatures are highly localized and that deep learning models possess superior discriminative capacity for this task.

0 citationsRead paper