LandslideFM: A Foundation Model Towards Unified Landslide Perception and Understanding

Introduction

Landslide hazards pose a serious threat to communities worldwide, demanding rapid and accurate detection as well as deeper contextual understanding of the affected terrain. Traditional approaches in this space have treated visual detection, segmentation, and scene interpretation as separate problems, often relying on large, fully‑labeled datasets that are costly to obtain. LandslideFM tackles these challenges head‑on by presenting the first unified foundation model that couples visual perception with language‑driven reasoning, enabling both precise landslide localization and richer situational insight.

A unified foundation model for landslides

LandslideFM builds on a two‑stage architecture. The first stage is a visual backbone trained with a Source‑Agreement Graduated Trust (SAGT) semi‑supervised framework. SAGT blends a modest set of annotated images with a much larger pool of unlabeled remote‑sensing data, gradually increasing trust in the model’s own predictions as training progresses. This approach reduces the reliance on exhaustive labeling while still achieving high‑quality segmentation of landslide regions.

The second stage integrates the visual output with a Vision‑Language Model (VLM) that has been adapted through parameter‑efficient fine‑tuning. The segmented landslide masks serve as explicit spatial cues for the VLM, guiding it to generate textual explanations, risk assessments, and answers to natural‑language queries about the scene. By linking the two modalities, LandslideFM moves beyond pixel‑level detection to provide contextual interpretations such as "the landslide is likely triggered by recent heavy rainfall" or "the slope angle exceeds the stability threshold."

LandslideVL: a benchmark for holistic evaluation

To evaluate this unified capability, the authors introduce LandslideVL, a benchmark that pairs high‑resolution imagery with curated textual prompts covering a range of tasks: binary detection, multi‑class segmentation, hazard rating, and open‑ended question answering. The benchmark draws from public datasets and newly collected field data, ensuring diverse geographic and environmental conditions. Experiments on LandslideVL demonstrate that LandslideFM consistently outperforms task‑specific CNNs on detection metrics and surpasses generic VLMs on reasoning tasks, highlighting the benefit of domain‑focused adaptation.

Practical implications and future directions

LandslideFM’s design offers several practical advantages for disaster‑response workflows:

• Reduced labeling burden: Semi‑supervised training leverages abundant unlabeled satellite or drone imagery, accelerating model deployment in data‑scarce regions.
• Unified output: Stakeholders receive both a precise landslide mask and an accompanying natural‑language summary, streamlining decision‑making for emergency managers.
• Scalable adaptation: Parameter‑efficient fine‑tuning allows the VLM component to be updated with new linguistic cues (e.g., local terminology) without retraining the entire model.

Future work could explore extending the framework to other geohazards such as floods or wildfires, integrating temporal sequences for early warning, and refining the trust‑graduation schedule to further improve robustness against noisy pseudo‑labels.

--- LandslideFM marks a significant step toward holistic, data‑efficient landslide analysis, marrying high‑resolution visual detection with language‑based insight. Its open‑source release invites the research community to build upon this foundation, fostering more resilient and informed responses to landslide disasters.