Source-linked AI summary
EarthLD: Towards Unified Open-World Landslide Understanding via Vision-Language Guided Diffusion Models
Yuanchao Su, Lianru Gao, Mengying Jiang, Jiangyi Chen, Jiaxin Cheng, Yicong Zhou
TL;DR
Global landslide understanding is difficult because landslides have irregular shapes, ambiguous spectral signatures, and substantial cross-sensor variation. EarthLD uses vision-language-guided diffusion and a harmonized open-world benchmark to unify landslide recognition and mapping. It consistently achieves the best performance in landslide identification and mapping while supporting cross-sensor transferability and large-scale monitoring.
Problem
Irregular morphology, similar-to-soil spectral signatures, and cross-sensor heterogeneity make automated landslide detection and mapping difficult in remote sensing imagery.
Method
EarthLD reframes landslide understanding as progressive denoising within a vision-language-guided diffusion framework and evaluates it using a harmonized global open-world benchmark.
Results
EarthLD consistently achieves the best performance in landslide identification and mapping across benchmark evaluations and demonstrates cross-modal transferability from optical to SAR imagery.
Takeaways & Limitations
EarthLD provides a unified framework for landslide recognition, range mapping, trigger estimation, counting, localization, and large-scale remote sensing monitoring.
Abstract
from arXiv · showhide
Landslides are widespread geological hazards, yet their automated detection and mapping in remote sensing imagery remain challenging because of their irregular morphology, ambiguous spectral signatures, and substantial domain shifts across imaging platforms. To overcome these challenges, we propose EarthLD, a vision-language-guided diffusion framework for open-world landslide understanding, enabling unified landslide recognition, mapping, and trigger interpretation. At its core, EarthLD formulates landslide understanding as a diffusion process that progressively infers the presence, spatial extent, and pixel-level boundaries of landslides from noisy latent representations. This probabilistic formulation enables the model to jointly perform image-level landslide recognition and mapping while characterizing predictive uncertainty. By integrating visual observations with contextual knowledge in the denoising process, EarthLD distinguishes diverse landslides from backgrounds, produces confidence-aware predictions for suspected regions, and maps landslide ranges. We additionally construct a global-scale open-world landslide benchmark by systematically harmonizing multiple publicly available remote sensing data collected by diverse institutions. Extensive experiments across regions, sensors, and triggering events demonstrate that EarthLD consistently outperforms existing landslide detection methods, highlighting its potential as a unified and robust solution for global geological-hazard monitoring and emergency response.
Introduction
Global landslide detection is hindered by irregular morphology, weakly distinctive spectral signatures, cross-sensor differences, and the limitations of existing approaches. EarthLD addresses these challenges by reframing detection as progressive noise reduction and combining vision-language guidance with a harmonized global benchmark.
- Challenges: Optical satellites remain essential for global landslide remote sensing because reliance solely on SAR/InSAR and ground-based sensors is insufficient.The passage specifically contrasts optical sensing with the costs and limitations of alternative sources.
- Challenges: Landslides lack consistent morphological and spectral characteristics, making them difficult to identify in multi-platform remote sensing imagery.Their signatures can closely resemble bare soil, limiting the benefit of multispectral or hyperspectral imagery.
- Challenges: Cross-sensor differences in grayscale, reflectance, and radiance increase the risk of false identification.These fine-grained discrepancies compound detection difficulty across imaging platforms.
- EarthLD: EarthLD reframes landslide detection as a step-by-step noise-reduction process to better capture irregular, hard-to-define shapes.The framework is presented as a unified open-world approach based on vision-language-guided diffusion models.
Proposed EarthLD
EarthLD unifies landslide recognition, spatial mapping, trigger estimation, and counting through vision-language-guided variational diffusion. It combines a global heterogeneous benchmark with detection, segmentation, and contextual reasoning modules for open-world remote-sensing imagery.
- Unified framework: EarthLD jointly performs landslide recognition, range mapping, trigger estimation, landslide counting, and geographic localization.Recognition is formulated as object detection, while mapping uses proposal-guided binary semantic segmentation.
- Mapping: The mapping branch classifies pixels as landslide foreground or background, with recognized boxes guiding delineation of irregular landslide boundaries.Foreground pixels in the resulting binary mask represent the mapped landslide extent.
- Recognition: The recognition branch predicts bounding boxes and objectness scores, retaining boxes above δdet as recognized landslide instances and using their count as the detected total.Bounding-box prediction is modeled with a variational diffusion process.
- Open-world benchmark: The benchmark harmonizes heterogeneous public imagery from 28 regions across six continents and 2011–2022 into a unified 25.1 GB repository.The sources span different sensors, orbital platforms, imaging geometries, resolutions, noise characteristics, and triggering conditions.
- Contextual reasoning: CLIP-based contextual reasoning integrates temporal, geographical, and inference-time trigger observations with visual features during landslide understanding.Trigger metadata may include rainfall or seismic measurements but excludes the ground-truth trigger label; cross-modal weights are computed from cosine similarity.
- Diffusion detection: Variational diffusion progressively models and denoises box representations, producing decoded boxes and objectness scores that define the final recognition output.The detection algorithm conditions proposal-wise diffusion on visual and metadata representations, then returns boxes satisfying the detection threshold.
Experiments
Experiments evaluate EarthLD for recognition, mapping, training behavior, ablation effects, regional generalization, SAR transfer, and large-scale imagery. EarthLD consistently achieves the best reported performance in landslide identification and mapping while supporting cross-modal and large-scale monitoring.
- Evaluation Metrics: Precision, Recall, and Average Precision are used to evaluate landslide recognition and mapping performance.Precision measures correctly predicted positives or pixel regions, while Recall measures successfully detected and segmented landslide targets.
- Training Visualization: Training progressively improves EarthLD’s feature representations for both object detection and semantic segmentation.Figure 3 reflects this enhancement through performance changes across training epochs.
- Ablation Study: Removing CLIP guidance or pre-training noticeably degrades EarthLD’s landslide identification and mapping performance.On unseen datasets, the ablated model without CLIP guidance and pre-training has markedly lower detection accuracy than the full model.
- Comparison with Baselines: EarthLD is compared with separate detection and semantic-segmentation baselines because it performs both tasks simultaneously.Detection baselines include Faster R-CNN, Sparse R-CNN, DAB-DETR, DINO, YOLOv11, and DiffusionDet; segmentation baselines include DeepLabV3+, UNet++, SegFormer, DINOv2, YOLO11s-seg, and Seg-Diffusion.
- Comparison with Baselines: EarthLD consistently achieves the best performance in both landslide identification and landslide mapping.The comparison covers detection and semantic-segmentation approaches, with regional evaluations using benchmark datasets from 10 global regions.
- Multimodal and Large-Scale Applications: EarthLD transfers to SAR imagery and handles large-scale remote sensing images while producing reports with event locations, spatial extents, and counts.Figure 8 presents SAR identification and mapping results, and Figure 9 presents large-scale imagery application.