Source-linked AI summary

GeBDA: Building Damage Assessment as Text-Based Sequence Prediction

Olivier Dietrich, Krishna Sapkota, Konrad Schindler, Genady Beryozkin

arXiv:2608.28567v1cs.CV

TL;DR

Building damage assessment needs scalable localization and grading after disasters, yet GeoVLMs have not demonstrated the full joint task with standard autoregressive decoding. GeBDA fine-tunes Gemma to generate quantized building boxes and damage labels from bi-temporal imagery and a prompt, achieving promising localization and classification results while retaining representation and generalization limitations.

  • Problem

    Manual damage mapping is too slow to scale, while existing GeoVLMs simplify dense BDA or delegate localization instead of jointly grading every building.

  • Method

    GeBDA fine-tunes a general-purpose Gemma VLM to autoregressively generate a variable-length sequence of quantized building bounding boxes and damage labels from bi-temporal images and a prompt.

  • Results

    GeBDA is competitive in localization and classification on optical imagery; rasterized localization reaches F_loc1 78.34 on xBD and 81.23 on Bright.

  • Takeaways & Limitations

    The results support using a general-purpose VLM for end-to-end BDA, with strong localization and promising damage classification from post-event optical images.

  • Takeaways & Limitations

    Plain-text coordinate encoding is not token-efficient, compute-heavy fine-tuning limits backbone scaling, and benchmark splits share disaster events.

Abstract

from arXiv · show

Conventionally, Building Damage Assessment (BDA) is tackled either with dedicated network architectures or by fine-tuning geospatial image foundation models. In this work, we ask whether a general-purpose Vision-Language Model (VLM) can localize buildings and grade their damage through autoregressive sequence generation alone. We cast BDA as predicting a variable-length set of bounding boxes, each specified by its coordinates and a damage label. Our preliminary implementation, based on the open Gemma model, achieves promising damage mapping results from only bi-temporal satellite images and a suitable text prompt.

1 Introduction

Building damage mapping is important for disaster response, but manual inspection does not scale. GeBDA tests whether a general-purpose VLM can jointly localize buildings and classify damage through single-pass autoregressive prediction.

  • Rapid building damage mapping supports disaster response and recovery, while manual inspection is prohibitively slow and unscalable.
  • GeoVLMs struggle with dense BDA localization because autoregressive decoders must generate long sequences of precise coordinates and may drift or hallucinate.
  • Existing GeoVLM approaches simplify BDA or delegate localization, leaving joint single-pass localization and grading of every building undemonstrated.
  • GeBDA fine-tunes a Gemma model to directly produce building locations and damage classes from bi-temporal images and a task-specific prompt.
  • Preliminary results suggest GeBDA is competitive in both localization and classification using optical images.

2 Related Work

BDA traditionally uses specialist dense-prediction networks, while GeoVLM research has not demonstrated full joint localization and grading. GeBDA follows an end-to-end generative alternative based on serialized spatial predictions.

  • Specialist BDA networks process pre- and post-disaster image pairs to output dense masks that localize and grade buildings.
  • Generalist VLM localization commonly uses external decoders, plain-text bounding-box serialization, or dedicated coordinate tokens.
  • Sequence-based localization preserves end-to-end generation, but autoregressive formulations share a critical failure mode in dense environments.
  • Existing GeoVLMs simplify full BDA through the dense-localization bottleneck rather than jointly localizing and grading every building.
  • GeBDA’s formulation transforms building polygons into quantized, serialized bounding-box strings with damage labels and predicts one target sequence.

3 Method

GeBDA reformulates joint building localization and damage assessment as autoregressive instance-sequence prediction. It serializes quantized box coordinates and damage classes, fine-tunes Gemma, and evaluates both instance and rasterized outputs.

  • BDA is formulated as instance-based prediction rather than pixel-level classification to match the autoregressive next-token operation of VLMs.
  • 3.1 Problem Formulation: Given pre- and post-disaster images and a prompt, the model generates a discrete sequence representing all scene buildings after converting polygons to boxes.
  • 3.1 Problem Formulation: Coordinates are uniformly quantized to a 1000 × 1000 grid before serialization into text alongside each building’s damage class.
  • 3.1 Problem Formulation: The target sequence concatenates all building strings, tokenizes them with the native tokenizer, and appends an <EOS> token.
  • 3.1 Problem Formulation: Plain-text coordinates preserve Gemma’s pretrained numerical representation and implicit geometric priors, but use tokens less efficiently.
  • Training and Inference: GeBDA combines a Transformer backbone with a SigLIP 2 vision encoder, freezes the vision encoder, and fine-tunes the remaining model end-to-end.
  • Evaluation: Localization matches predicted and ground-truth boxes with Hungarian assignment at IoU 0.5, while classification computes F1 only on correctly matched instances.
  • Evaluation: Predictions are rasterized into dense masks for comparison with pixel-based models, while an Oracle quantifies the geometric ceiling from box representation.

4 Experiments and Results

GeBDA is evaluated on xBD and Bright using instance-based and rasterized pixel-based comparisons, with qualitative analyses of localization, classification, and zero-shot generalization. Results show strong localization, competitive optical-image classification, substantial degradation on Bright SAR imagery, and improved Bright classification when jointly trained.

  • Datasets: GeBDA is trained and evaluated on xBD and a subset of Bright with available instance labels, excluding unclassified buildings and those smaller than 16 pixels.xBD contains intact, minor-damage, major-damage, and destroyed classes; Bright contains intact, damaged, and destroyed classes.
  • Experimental setup: Each 1024×1024 image is cropped into 512×512 patches, limiting sequence length while enabling up to 200 buildings per patch.The configuration uses up to 19 tokens per building and a vision-encoder token budget of 280.
  • Localization: Predicted and ground-truth building counts show near-perfect correlation, although dense patches can produce clusters of tiny false-positive boxes and ambiguous boundaries can lead to multiple boxes for one building.These artifacts help explain why qualitative localization appears stronger than instance-based metrics.
  • Damage classification: Fcls1 drops from 76.62 on xBD to 48.99 on Bright, which the authors attribute to the modality gap in Bright’s post-event SAR imagery.The frozen RGB-pretrained vision encoder struggles to extract robust discriminative features from the SAR data.
  • Cross-dataset training: Joint training on xBD and Bright yields no improvement overall except a +22.6% increase in Bright Fcls1.The authors hypothesize that xBD transfers modality-agnostic priors and grounding in shared optical pre-event imagery to Bright.

5 Discussion & Conclusion

GeBDA frames building damage assessment as autoregressive sequence prediction, showing strong optical-image localization and promising damage classification while exposing dense-scene, SAR, and representation limitations. Future work targets more efficient representations, parameter-efficient scaling, spatially aligned optimization, and cross-disaster generalization.

  • GeBDA frames BDA as autoregressive sequence prediction and achieves strong localization and promising damage classification with post-event optical images.The localization capability is tempered by the bounding-box representation.
  • A frozen vision encoder fails to extract discriminative features from post-event SAR imagery, motivating vision-encoder fine-tuning.
  • Dense patches can trigger coordinate-generation loops, likely because training caps sequences at 200 buildings, while heterogeneous tiles can produce a single predicted damage class.
  • Plain-text bounding-box coordinates are token-inefficient, and fully supervised fine-tuning limits backbone scaling; larger models and LoRA are suggested alternatives.The authors identify Gemma4-26B with a 500M-parameter vision encoder as a potentially beneficial but currently impractical scale-up.
  • Future work will investigate reinforcement learning for spatial metrics such as bounding-box IoU and test generalization across disasters, whose splits currently share disaster events.
Loading 2608.28567v1…