the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 02 Sep 2026
Automated landslide detection in SAR wrapped interferograms using a geomorphology-constrained YOLO CNN
Abstract. Slow-moving landslides pose significant hazards in mountain environments, requiring improved detection and monitoring capabilities. Traditional mapping is accurate but time-consuming, while multitemporal InSAR approaches are limited by data complexity and velocity constraints. Wrapped dual-pass DInSAR interferograms offer an alternative by preserving deformation signals without phase unwrapping, enabling detection across a wide range of movement rates.
We present a deep learning framework for the automated detection and classification of active slow landslides in Sentinel-1 wrapped SAR interferograms. The model uses a YOLO convolutional neural network ingesting wrapped phase, an InSAR reliability index, and a terrain morphometric attribute layer. We trained, validated, and tested the network on 2243 labelled DInSAR wrapped phase signals from expert geomorphological interpretation over a 1200 km² sector of the Northern Apennines (Italy), using interferograms from ascending and descending orbits generated with multiple temporal baselines between 6 and 30 days.
The network outputs bounding boxes with movement classification, achieving a mean Average Precision of 0.88 and an F1 score of 0.75. It successfully identifies deformation signals across multiple spatial scales, also in interferograms with low signal-to-noise ratio. Our results demonstrate the potential of wrapped DInSAR data combined with deep learning for efficient regional-scale landslide detection and inventory updating.
Status: open (until 14 Oct 2026)
-
RC1: 'Comment on egusphere-2026-4523', Anonymous Referee #1, 20 Sep 2026
reply
- Please replace the random tile-level split with spatially disjoint, temporally disjoint, and event-disjoint splits to eliminate strong leakage from adjacent or repeated observations of the same landslides.
- Please conduct leave-one-interferogram-out and leave-one-landslide-out validation, since random patches from the same scene can share phase texture, topography, acquisition geometry, and annotations.
- Please quantify the extent to which repeated observations of individual landslides occur across training, validation, and test partitions, and report results after grouping all instances from each physical landslide into one split only.
- Please evaluate generalization through an external test area with different lithology, terrain morphology, vegetation, climate, and SAR viewing geometry; a single-site test cannot support claims of framework transferability.
- Please provide a rigorous ablation study that isolates the contribution of wrapped phase, the InSAR sensitivity layer, and the geomorphometric PC1 layer, including statistical uncertainty for every ablation result.
- Please benchmark the customized YOLOv3 model against strong baselines, including phase-only YOLO, coherence-plus-phase input, modern one-stage detectors, segmentation approaches, and an expert-defined non-deep learning detection workflow.
- Please justify the use of raw wrapped phase as a numerical input by explicitly describing phase normalization, circular phase representation, handling of the discontinuity, and the physical validity of additive and multiplicative radiometric augmentation.
- Please report the complete training protocol, including initialization, pretraining status, batch size, learning rate schedule, loss weights, anchor dimensions, augmentation ranges, random seeds, hardware, inference time, and the exact criterion used to select the epoch and confidence threshold.
- Please report mAP across multiple IoU thresholds, including mAP@0.5:0.95, classwise confidence intervals, calibration analysis, and sensitivity to confidence threshold, because performance at only IoU =0.5=0.5=0.5 and a selected threshold of 0.81 is insufficiently informative.
- Please replace the qualitative assertion that many false positives are likely true missed landslides with an independent adjudication protocol, ideally involving blinded expert reassessment, field evidence where feasible, and a revised reference standard with inter-annotator agreement.
ReplyCitation: https://doi.org/10.5194/egusphere-2026-4523-RC1 -
CC1: 'Comment on egusphere-2026-4523', Xiaochuan Tang, 05 Oct 2026
reply
This manuscript proposes a YOLO-based framework for detecting slow-moving landslides from wrapped DInSAR interferograms by integrating three types of information: wrapped phase, an InSAR Sensitivity Layer, and geomorphological features. The use of object detection is reasonable because precise landslide boundaries are difficult to delineate from InSAR observations, and bounding boxes can reduce the uncertainty associated with pixel-level annotation. The integration of InSAR observability constraints and geomorphological information is a meaningful aspect of the study, as it provides both physical and terrain-based context for the interpretation of deformation signals. The reported results are promising, particularly the cases where the model appears to identify subtle deformation signals overlooked during expert interpretation. However, the manuscript requires major revision before publication. The following suggestions are provided to further improve the manuscript.
- A deeper geomorphological interpretation of the class-dependent performance is recommended. The manuscript would benefit from a deeper geomorphological interpretation of the class-dependent detection performance. In particular, the authors should discuss whether the differences among CF, CS, and FS are related to landslide kinematics, lithology, morphology, vegetation cover, slope orientation, or the spatial organization of the deformation field. Such an analysis would help demonstrate that the model is learning physically meaningful landslide signatures rather than merely image patterns.
- Confidence scores should be displayed for the predicted bounding boxes.The predicted bounding boxes shown in the figures currently provide mainly the predicted class. It would be useful to display the corresponding confidence score or class probability for each detection. This information would allow readers to better evaluate the reliability of individual predictions, especially for false-positive cases that may actually represent weak deformation signals overlooked by expert interpretation.
- Duplicate or highly overlapping bounding boxes should be further addressed.Some deformation signals appear to be represented by multiple overlapping predicted bounding boxes. This may artificially increase the number of false positives and can also complicate the interpretation of model performance. The authors should clarify whether non-maximum suppression (NMS), Soft-NMS, or a similar post-processing strategy was applied. If so, the relevant thresholds should be reported. If not, an appropriate bounding-box deduplication strategy should be considered.
- Quantitative comparisons with existing landslide detection methods are needed.The manuscript would be substantially strengthened by comparisons with existing methods. For example, the approaches mentioned in Introduction,i.e., previous approaches based on YOLO with phase-gradient maps, as well as FCN-, DeepLabV3-, U-Net-, and MB-Net-based methods using wrapped interferograms.
- An ablation study is strongly recommended.One of the principal methodological contributions of this manuscript is the integration of wrapped DInSAR phase, the InSAR Sensitivity Layer, and geomorphological information. However, the individual contribution of each component is not sufficiently quantified.An ablation experiment should therefore be conducted using, for example: wrapped DInSAR phase only; wrapped DInSAR phase + ISL; wrapped DInSAR phase + geomorphological information; wrapped DInSAR phase + ISL + geomorphological information. The resulting changes in mAP, precision, recall, and F1-score would provide direct evidence of the contribution of the InSAR physical-observability constraint and the geomorphological constraint.
- The figures should provide clearer geographic coordinates, which allows readers to better locate the detected landslides and facilitate verification and reproducibility of the case studies.
Citation: https://doi.org/10.5194/egusphere-2026-4523-CC1
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 239 | 0 | 3 | 242 | 0 | 0 |
- HTML: 239
- PDF: 0
- XML: 3
- Total: 242
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
