[Paper Review] Landslide Segmentation with U-Net: Evaluating Different Sampling Methods and Patch Sizes

Lucas Pedrosa Soares, Helen Cristina Dias+1 more|arXiv (Cornell University)|Jul 13, 2020

Landslides and related hazardsEnvironmental Science29 references35 citations

TL;DR

This study assesses how patch size, sampling method, and input datasets affect landslide segmentation performance using U-net, RapidEye imagery, and an ALOS PALSAR-derived DEM. Random sampling and DEM augmentation improve generalization and accuracy.

ABSTRACT

Landslide inventory maps are crucial to validate predictive landslide models; however, since most mapping methods rely on visual interpretation or expert knowledge, detailed inventory maps are still lacking. This study used a fully convolutional deep learning model named U-net to automatically segment landslides in the city of Nova Friburgo, located in the mountainous range of Rio de Janeiro, southeastern Brazil. The objective was to evaluate the impact of patch sizes, sampling methods, and datasets on the overall accuracy of the models. The training data used the optical information from RapidEye satellite, and a digital elevation model (DEM) derived from the L-band sensor of the ALOS satellite. The data was sampled using random and regular grid methods and patched in three sizes (32x32, 64x64, and 128x128 pixels). The models were evaluated on two areas with precision, recall, f1-score, and mean intersect over union (mIoU) metrics. The results show that the models trained with 32x32 tiles tend to have higher recall values due to higher true positive rates; however, they misclassify more background areas as landslides (false positives). Models trained with 128x128 tiles usually achieve higher precision values because they make less false positive errors. In both test areas, DEM and augmentation increased the accuracy of the models. Random sampling helped in model generalization. Models trained with 128x128 random tiles from the data that used the RapidEye image, DEM information, and augmentation achieved the highest f1-score, 0.55 in test area one, and 0.58 in test area two. The results achieved in this study are comparable to other fully convolutional models found in the literature, increasing the knowledge in the area.

Motivation & Objective

• Assess the impact of patch size (32x32, 64x64, 128x128) on landslide segmentation accuracy.
• Evaluate sampling methods (random vs regular grid) for model generalization.
• Test the influence of data augmentation and DEM as an input channel on U-net performance.
• Compare model outcomes across two test areas with standard segmentation metrics.

Proposed method

• Train 288 U-net models with inputs from RapidEye imagery and ALOS PALSAR DEM.
• Patch inputs in three sizes (32x32, 64x64, 128x128) sampled via random or regular grid methods.
• Apply data augmentation (rotations and flips) using Albumentations.
• Evaluate models on two 1024x1024 test areas using precision, recall, f1-score, and mIoU.
• Use binary cross-entropy loss and Adam optimizer; train for 200 epochs with early saving based on validation loss.

Experimental results

Research questions

• RQ1 How do patch size and sampling method influence landslide segmentation accuracy with U-net?
• RQ2 Does adding a DEM channel improve model performance compared to using optical data alone?
• RQ3 Does data augmentation enhance model generalization across different land cover contexts?
• RQ4 Which combination of dataset, sampling method, and patch size yields the best balance of precision, recall, and IoU across test areas?

Key findings

• 128x128 random tiles with RapidEye+DEM+Augmentation achieved the highest F1-score: 0.55 (test area 1) and 0.58 (test area 2).
• Models with 32x32 tiles show higher recall but lower precision due to more false positives.
• 64x64 tiles yield intermediate results between 32x32 and 128x128 in generalization.
• DEM as an input channel improves accuracy across test areas.
• Random sampling outperforms regular grid sampling on average mIoU across both test areas.