A Benchmark Study of Segmentation Models and Adaptation Strategies for Landslide Detection from Satellite Imagery

Chen Chen, Md Kowsher, Weiwei Zhan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 08:37 UTC · model grok-4.3

classification 💻 cs.CV

keywords modelssegmentationlandslidedetectionfinetuningperformancedatasetfine-tuning

0 comments

The pith

Transformer-based models deliver strong landslide segmentation on satellite images, and parameter-efficient fine-tuning matches full fine-tuning accuracy while cutting trainable parameters by up to 95%.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests different AI models for spotting landslides from high-resolution satellite photos. Landslides are hard to map quickly after disasters, so reliable detection tools matter for response and planning. The authors evaluate convolutional neural networks that use filters to scan images, transformer models that break images into patches and attend across them, and large pre-trained foundation models. They run all models on the Globally Distributed Coseismic Landslide Dataset under the same training and testing rules. They also compare two adaptation approaches: updating every model parameter versus efficient methods such as LoRA and AdaLoRA that only adjust a small subset of parameters. Results indicate transformers outline landslide areas accurately. The efficient methods reach nearly the same accuracy while training far fewer parameters. The study further checks how well models hold up when tested on data that differs from the training distribution.

Core claim

Experimental results show that transformer-based models achieve strong segmentation performance, while parameter efficient finetuning reduces trainable parameters by up to 95% with comparable accuracy to full finetuning.

Load-bearing premise

That the GDCLD dataset and the fixed training/evaluation protocols provide a fair, representative testbed that allows direct comparison of CNN, transformer, and foundation models without hidden biases in data distribution or annotation quality.

read the original abstract

Landslide detection from high resolution satellite imagery is a critical task for disaster response and risk assessment, yet the relative effectiveness of modern segmentation architectures and finetuning strategies for this problem remains insufficiently understood. In this work, we present a systematic benchmarking study of convolutional neural networks, transformer based segmentation models, and large pre-trained foundation models for landslide detection. Using the Globally Distributed Coseismic Landslide Dataset (GDCLD) dataset, we evaluate representative CNN- and transformer-based segmentation models alongside large pretrained foundation models under consistent training and evaluation protocols. In addition, we compare full fine-tuning with parameter-efficient fine-tuning methods, including LoRA and AdaLoRA, to assess their performance efficiency tradeoffs. Experimental results show that transformer-based models achieve strong segmentation performance, while parameter efficient finetuning reduces trainable parameters by up to 95% with comparable accuracy to full finetuning. We further analyze generalization under distribution shift by comparing validation and held-out test performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A straightforward benchmark that shows transformers and LoRA work well for landslide segmentation on GDCLD, but adds little beyond applying existing tools to one more dataset.

read the letter

This paper runs a head-to-head comparison of CNN segmentation models, transformer models, and foundation models on the GDCLD landslide dataset, plus full fine-tuning versus LoRA and AdaLoRA. The central result is that transformers perform strongly while parameter-efficient methods cut trainable parameters by up to 95% with comparable IoU and F1 scores. They also check generalization on held-out test data to probe distribution shift. That is the useful piece for anyone who actually needs to deploy something for disaster monitoring.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Empirical benchmark relying on standard supervised segmentation training and the GDCLD dataset; no new free parameters, axioms, or invented entities beyond conventional ML practice.

pith-pipeline@v0.9.0 · 5468 in / 1064 out tokens · 36376 ms · 2026-05-10T08:37:14.450905+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 3 canonical work pages · 2 internal anchors

[1]

National strategy for landslide loss reduction,

J. W. Godt, N. J. Wood, A. Pennaz, C. M. Dacey, B. B. Mirus, L. N. Schaefer, and S. L. Slaughter, “National strategy for landslide loss reduction,” US Geological Survey, Tech. Rep., 2022

2022
[2]

Spatial and temporal analysis of a global landslide catalog,

D. Kirschbaum, T. Stanley, and Y . Zhou, “Spatial and temporal analysis of a global landslide catalog,”Geomorphology, vol. 249, pp. 4–15, 2015

2015
[3]

Landslides across the usa: occurrence, susceptibility, and data limitations,

B. B. Mirus, E. S. Jones, R. L. Baum, J. W. Godt, S. Slaughter, M. M. Crawford, J. Lancaster, T. Stanley, D. B. Kirschbaum, W. J. Burnset al., “Landslides across the usa: occurrence, susceptibility, and data limitations,”Landslides, vol. 17, no. 10, pp. 2271–2285, 2020

2020
[4]

Widespread initiation, reactiva- tion, and acceleration of landslides in the northern california coast ranges due to extreme rainfall,

A. L. Handwerger, E. J. Fielding, M.-H. Huang, G. L. Bennett, C. Liang, and W. H. Schulz, “Widespread initiation, reactiva- tion, and acceleration of landslides in the northern california coast ranges due to extreme rainfall,”Journal of Geophysical Research: Earth Surface, vol. 124, no. 7, pp. 1782–1797, 2019

2019
[5]

Variation in the frequency and char- acteristics of landslides in response to changes in forest cover and rainfall in japan over the last century: A literature review,

T. Sato and Y . Shuin, “Variation in the frequency and char- acteristics of landslides in response to changes in forest cover and rainfall in japan over the last century: A literature review,” Catena, vol. 249, p. 108639, 2025

2025
[6]

Harnessing geospatial artificial intelligence and deep learning for landslide inventory mapping: Advances, challenges, and emerging directions,

X. Chen, W. Li, C.-Y . Hsu, S. T. Arundel, and B. Higman, “Harnessing geospatial artificial intelligence and deep learning for landslide inventory mapping: Advances, challenges, and emerging directions,”Remote Sensing, vol. 17, no. 11, p. 1856, 2025

2025
[7]

Landslide hazard mapping with geospatial foundation models: Geographical generalizability, data scarcity, and band adaptability,

W. Li, S. Wang, H. Lee, C. Lu, S. Roy, R. Ramachandran, and C.-Y . Hsu, “Landslide hazard mapping with geospatial foundation models: Geographical generalizability, data scarcity, and band adaptability,” 2025. [Online]. Available: https: //arxiv.org/abs/2511.04474

work page arXiv 2025
[8]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

2015
[9]

Rethinking Atrous Convolution for Semantic Image Segmentation

L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethink- ing atrous convolution for semantic image segmentation. arxiv 2017,”arXiv preprint arXiv:1706.05587, vol. 2, p. 1, 2019

work page internal anchor Pith review arXiv 2017
[10]

Prithvi-eo-2.0: A versatile multi- temporal foundation model for earth observation applications,

D. Szwarcman, S. Roy, P. Fraccaro, O. E. G ´ıslason, B. Blu- menstiel, R. Ghosal, P. H. De Oliveira, J. L. de Sousa Almeida, R. Sedona, Y . Kanget al., “Prithvi-eo-2.0: A versatile multi- temporal foundation model for earth observation applications,” IEEE Transactions on Geoscience and Remote Sensing, 2025

2025
[11]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”ICLR, vol. 1, no. 2, p. 3, 2022

2022
[12]

AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning

Q. Zhang, M. Chen, A. Bukharin, N. Karampatziakis, P. He, Y . Cheng, W. Chen, and T. Zhao, “Adalora: Adaptive budget allocation for parameter-efficient fine-tuning,”arXiv preprint arXiv:2303.10512, 2023

work page internal anchor Pith review arXiv 2023
[13]

A globally distributed dataset of coseismic land- slide mapping via multi-source high-resolution remote sensing images,

C. Fang, X. Fan, X. Wang, L. Nava, H. Zhong, X. Dong, J. Qi, and F. Catani, “A globally distributed dataset of coseismic land- slide mapping via multi-source high-resolution remote sensing images,”Earth System Science Data, vol. 16, no. 10, pp. 4817– 4842, 2024

2024
[14]

Segformer: Simple and efficient design for semantic segmentation with transformers,

E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,”Advances in neural informa- tion processing systems, vol. 34, pp. 12 077–12 090, 2021

2021
[15]

Resunet-a: A deep learning framework for semantic segmen- tation of remotely sensed data,

F. I. Diakogiannis, F. Waldner, P. Caccetta, and C. Wu, “Resunet-a: A deep learning framework for semantic segmen- tation of remotely sensed data,”ISPRS Journal of Photogram- metry and Remote Sensing, vol. 162, pp. 94–114, 2020

2020
[16]

Deep high-resolution repre- sentation learning for visual recognition,

J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wanget al., “Deep high-resolution repre- sentation learning for visual recognition,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 10, pp. 3349–3364, 2020

2020
[17]

Unified perceptual parsing for scene understanding,

T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 418–434

2018
[18]

Swin-unet: Unet-like pure transformer for medical image segmentation,

H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-unet: Unet-like pure transformer for medical image segmentation,” inEuropean conference on computer vision. Springer, 2022, pp. 205–218

2022
[19]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015– 4026

2023
[20]

Scene parsing through ade20k dataset,

B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Tor- ralba, “Scene parsing through ade20k dataset,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, 2017, pp. 633–641

2017