From Landslide Reports to Training Masks

1. Problem & Constraint

Landslides are among the most destructive natural hazards in mountainous regions, yet building accurate detection models requires large, well-annotated datasets that are expensive and time-consuming to produce. Reported incident records provide a starting point, but raw incident zones are far too large for manual inspection — their median area is approximately 37.2 km² with a median longest side of approximately 6.94 km.

The core challenge this work addresses: how do we go from reported incidents to reviewable, geospatially consistent candidate packages with pixel-level annotations?

Instead of asking analysts to search complete incident zones, the workflow presents ranked regions supported by optical change, SAR change, terrain plausibility, and segment-level statistical evidence. This preserves human judgment while making review more focused and reproducible.

2. Multi-Source Observations

Input Data

The workflow fuses four complementary data sources, each providing distinct information about the landscape before and after an event:

Why Multi-Source?

Optical and SAR observations respond differently to vegetation, moisture, shadow, layover, and land-cover change. Combining them reduces false positives from cloud cover (which affects optical but not SAR) and from surface moisture changes (which affect SAR but may not represent landslide geometry). Terrain slope and aspect provide plausibility constraints — landslides are more likely on steep slopes with specific orientations.

3. Detection & Annotation Pipeline

Incident-Driven Pipeline

The pipeline proceeds through five stages: incident ingestion, multi-source image acquisition, change detection with candidate generation, QGIS manual review, and segmentation annotation. Each stage is tracked with persistent processing records and deterministic incident identifiers.

1. 01
   Ingest

   Normalize BIPAD incident records and assign deterministic identifiers.

2. 02
   Align

   Acquire before/after observations and reproject optical, SAR, and terrain layers to one grid.

3. 03
   Propose

   Detect change, fuse modalities, and rank geometrically plausible candidate regions.

4. 04
   Review

   Classify candidates in QGIS using before/after imagery, terrain, and change context.

5. 05
   Annotate

   Draw pixel-level boundaries for confirmed landslides while retaining provenance.

Change Detection

For each incident, the workflow applies IR-MAD (Iteratively Reweighted Multivariate Alteration Detection) to the co-registered optical and SAR imagery. The change maps are combined through geometric optical-SAR fusion and subjected to terrain filtering using slope and aspect constraints.

Statistical Significance & Filtering

Bonferroni correction at the incident level limits false candidate generation. Hysteresis thresholding separates strong change cores from weaker surrounding change. Morphological cleanup removes isolated pixels and fills small gaps. Geometric screening applies area and elongation filters to remove implausible objects (too small, too elongated, or otherwise inconsistent with landslide morphology).

Manual Review in QGIS

Each generated candidate is examined in QGIS using the before-event image, after-event image, slope context, and candidate mask. The reviewer classifies the candidate as a landslide or non-landslide change. Confirmed landslides are then selected for pixel-level segmentation annotation.

Segmentation Annotation

For confirmed landslides, pixel-level segmentation masks are drawn so that the visible landslide extent, rather than only the candidate bounding region, can be used for supervised segmentation-model development. Annotations are retained with incident identifiers and supporting imagery to preserve traceability.

4. Quality Control

Quality control is enforced through multiple mechanisms:

5. Results

Current Data Volumes

Coverage Analysis

The coverage analysis below was generated from a slightly earlier snapshot containing 247 landslide and 5,594 non-landslide labels—one fewer in each class than the current operational tally above. The incident catalog contains 4,138 reports spanning February 16, 2018 to May 20, 2026. Keeping the snapshot distinction explicit prevents exploratory statistics from being mistaken for live counts.

Geographic Distribution

Elevation Distribution

Labeled landslides occur at lower elevations than non-landslide labels, with a broader high-elevation tail among non-landslide samples.

Temporal Distribution

The dataset is strongly concentrated in the monsoon season. Of the 4,138 reported incidents, 3,524 (85.2%) occur during the monsoon, and 201 of the 247 labeled landslides (81.4%) fall in the same season.

Physiographic Distribution

6. Discussion

Value of Incident-Driven Approach

The workflow contributes a continuously extensible dataset-construction process rather than a fixed benchmark. Its principal value is the traceable connection between a reported incident, the multi-source evidence used for candidate generation, the analyst's decision, and the final segmentation annotation. This structure supports later error analysis and enables the dataset to grow without discarding provenance.

Class Imbalance Is Informative

The 4.2% landslide rate among reviewed candidates confirms that temporal change alone is not equivalent to landslide occurrence. The retained hard negatives — changes caused by vegetation, agriculture, roads, rivers, illumination, moisture, and sensor effects — are essential for training models that must distinguish landslides from visually similar disturbances.

Coverage Gaps

The coverage analysis reveals geographic, temporal, elevation, and physiographic imbalance. Labeled landslides are concentrated in the monsoon season and Middle Hills, with no labeled sample in Madhesh or the Terai band. These imbalances should be considered when splitting the data and interpreting model performance because a random split could overrepresent dominant regions while providing weak evidence of generalization to underrepresented environments.

7. Limitations

Coordinate uncertainty. BIPAD coordinates may describe affected settlements rather than exact failure locations, introducing spatial offset between the reported incident and the actual landslide.

Temporal smoothing. Quarterly Planet mosaics may smooth or temporally mix short-lived surface changes, potentially missing rapid post-event recovery or obscuring the exact timing of change.

Modality-specific biases. Optical and SAR observations respond differently to vegetation, moisture, shadow, layover, and land-cover change. Reprojection and resampling can introduce boundary uncertainty, while geometric fusion may suppress a true change that is strongly visible in only one modality.

Statistical assumptions. The chi-square aggregation assumes an approximate statistical model that may not hold uniformly across terrain and acquisition conditions.

Incomplete processing. Only 730 of 4,138 incidents have completed candidate detection. Manual labels remain class-imbalanced and subject to analyst interpretation. Segmentation annotations are available for 164 of 248 confirmed landslide candidates, leaving part of the positive collection without pixel-level boundaries.

The workflow establishes a traceable data foundation. Model performance, operational recall, and generalization to underrepresented regions remain open questions.

8. Future Work

• Process the remaining 3,408 incidents through the full pipeline.
• Complete outstanding segmentation masks for all 248 confirmed landslide candidates.
• Resolve the 170 non-landslide labels outside province polygons through coordinate or boundary review.
• Measure inter-annotator agreement to quantify label quality.
• Compare detections with independent landslide inventories for external validation.
• Evaluate baseline classification and segmentation models using province-aware, elevation-aware, and temporally separated validation splits.
• Address the strong geographic, temporal, and physiographic imbalance through stratified sampling or targeted data collection in underrepresented regions (Madhesh, Terai, off-season events).

9. Conclusion

The system converts broad BIPAD reports into a traceable sequence of aligned observations, ranked change candidates, human decisions, and pixel-level masks. It has processed 730 incidents and produced 164 segmentation annotations without treating automated change maps as ground-truth landslide labels.

The strongest outcome is the workflow itself: it makes sparse-event labeling more focused and reproducible while preserving difficult negative examples. The next evidence threshold is model evaluation with geographically and temporally separated splits—not a larger raw annotation count alone.