Landslide4Sense Benchmark for Landslide Detection

• Landslide4Sense is a benchmark that fuses 14-channel Sentinel-2 and topographic data, enabling pixel-level landslide detection via deep-learning segmentation.
• The dataset comprises 3,799 patches from four geographically distinct landslide events, supporting comparative analysis of segmentation models on imbalanced data.
• Its annotation pipeline combines rule-based extraction with manual refinement, providing precise, pixel-wise labels for robust and accurate landslide mapping.

Landslide4Sense is a reference benchmark for landslide detection from remote sensing that was introduced as a publicly available, pixel-labeled resource for comparing deep-learning segmentation models on fused optical and topographic imagery. In its original form, the benchmark contains 3,799 non-overlapping image patches of size $128 \times 128$ pixels, where each pixel covers $10 \times 10$ m on the ground and each patch is represented as a 14-channel tensor formed by stacking 12 Sentinel-2 bands with a digital elevation model and slope derived from ALOS PALSAR. The benchmark was assembled from four geographically distinct, event-specific landslide inventories and was designed to support the development, validation, and systematic comparison of methods for automatic landslide mapping and inventory updating (Ghorbanzadeh et al., 2022).

1. Benchmark definition and geographic coverage

The benchmark fuses multisource Earth-observation data. Its optical component is Sentinel-2 imagery with 12 spectral bands at 10 m, 20 m, and 60 m ground sampling, all resampled to 10 m in the released benchmark. Its topographic component consists of a 12.5 m ALOS PALSAR DEM and a derived slope layer, likewise resampled to 10 m. The final representation is therefore a 14-channel patch: 12 Sentinel-2 bands plus DEM and slope. The stated motivation for including topography is that optical data alone make accurate delineation of landslide borders difficult, whereas the added topographical information facilitates border detection (Ghorbanzadeh et al., 2022).

Region	Event and date	Trigger context
Iburi–Tobu, Hokkaido, Japan	September 2018	earthquake + Typhoon Jebi
Kodagu District, Karnataka, India	August 2018	extreme monsoon rainfall
Rasuwa District, Bagmati, Nepal	April 2015	Mw 7.8 Gorkha earthquake
Western Taitung County, Taiwan	August 2009	Typhoon Morakot

Every released patch is $128 \times 128$ pixels, and the original benchmark release comprises 3,799 such patches. The data are intended for landslide-versus-non-landslide semantic segmentation at pixel level, not merely scene classification. Later competition material preserves the same core 14-layer representation and the same four-event public training set, but adds hidden evaluation splits for leaderboard-based comparison (Ghorbanzadeh et al., 2022).

2. Annotation pipeline, preprocessing, and evaluation regimes

The labeling process is explicitly two-stage. First, an OBIA pre-segmentation workflow computes pre-/post-event indices, performs multi-resolution segmentation, and applies rule-based extraction of candidate landslide polygons. Second, each polygon is manually checked against external sources such as Google Earth and existing inventories, then refined so that every pixel is labeled as either “landslide” or “non-landslide.” This makes Landslide4Sense a pixel-level benchmark rather than a weakly labeled scene archive (Ghorbanzadeh et al., 2022).

The original benchmark paper defines a geographic split protocol rather than a random split. For each study area, the mosaicked extent is divided into four geographic “quarters”; one quarter from each area, totaling 959 patches, is used for training, and the other three quarters, totaling 2,840 patches, are used for testing. Evaluation is based on the standard segmentation metrics

$$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$

with Intersection-over-Union, when reported, defined as

$$\mathrm{IoU} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}.$$

Cloud-free Sentinel-2 acquisition was ensured via Google Earth Engine, and all bands, including DEM and slope, were resampled to 10 m. No further radiometric normalization is prescribed in the benchmark release, although later users commonly apply min–max scaling or per-band mean–std normalization (Ghorbanzadeh et al., 2022).

A second, competition-oriented evaluation regime was introduced in the 2022 Landslide4Sense competition. In that setting, the public benchmark comprises 3,799 labeled training patches, while 245 validation and 800 test patches remain unlabeled for participants; no very-high-resolution or ancillary data beyond the 14 bands are permitted. This establishes that Landslide4Sense has been used under more than one formal protocol, and reported scores from different papers are not automatically commensurate unless the split definition is stated (Ghorbanzadeh et al., 2022).

Later studies further emphasize the dataset’s severe pixel imbalance. One Geo-Foundational Model study characterizes the training data as containing approximately 2% positive pixels, with an imbalance ratio of approximately $49:1$ background to landslide; another reports that only approximately 2.3% of all pixels are landslide. This suggests that loss design, threshold calibration, and sampling strategy are central methodological variables on Landslide4Sense rather than secondary implementation details (Vu, 12 Jun 2026).

3. Original baseline study and comparative segmentation results

The benchmark paper evaluates 11 state-of-the-art semantic-segmentation architectures trained from scratch under the original quarter-based protocol: FCN-8s, PSPNet, ContextNet, DeepLab-v2, DeepLab-v3+, LinkNet, FRRN-A, FRRN-B, SQNet, U-Net, and ResU-Net. The reported test-set results are as follows (Ghorbanzadeh et al., 2022).

Model	Recall (%)	Precision (%)	F1-score (%)
PSPNet	52.03	61.55	56.39
ContextNet	49.29	70.77	58.11
DeepLab-v2	63.68	60.80	62.21
DeepLab-v3+	62.11	69.91	65.78
FCN-8s	63.05	68.66	65.73
LinkNet	67.02	66.76	66.89
FRRN-A	64.40	76.57	69.96
FRRN-B	76.16	64.93	70.10
SQNet	66.69	74.20	70.24
U-Net	62.17	79.91	69.94
ResU-Net	67.71	76.08	71.65

ResU-Net is the best overall model in this baseline comparison, with Recall $67.71\%$, Precision $76.08\%$, and F1-score $71.65\%$. The paper’s qualitative analysis distinguishes several error modes. PSPNet and ContextNet exhibit high precision but low recall, which is associated with missing small or subtle landslides. FRRN-B and the DeepLab variants detect most landslide pixels but incur false alarms. ResU-Net is reported as achieving the best balance, and its qualitative maps minimize both false negatives and false positives relative to the other tested architectures. Feature-map visualization further indicates that shallow layers preserve fine boundary details while deeper layers capture abstract contextual structure, which is presented as a justification for U-shaped and skip-connected designs in this task (Ghorbanzadeh et al., 2022).

These baseline findings gave Landslide4Sense a concrete role beyond data release: it became a controlled benchmark for studying the recall–precision trade-off under extreme foreground sparsity and for testing whether multiscale encoders, skip connections, and topographic fusion materially improve pixel-level delineation.

4. Competition outcomes and later architectural developments

The 2022 Landslide4Sense competition extended the benchmark into a public evaluation ecosystem. The final ranking was determined by pixel-wise F1 on hidden test patches, while Precision, Recall, and IoU were additionally exposed during validation. The competition paper reports that the baseline U-Net achieved F1 $=57.82\%$ on validation and $10 \times 10$0 on test. The top-ranked solutions then pushed hidden-test performance into the mid-70% range through combinations of Swin Transformer, SegFormer, U-Net-style decoders, hard example mining, self-training, mix-up augmentation, and domain-adaptive normalization (Ghorbanzadeh et al., 2022).

The winning Kingdrone solution used an ensemble of three models, including a U-Decoder paired with Swin Transformer and EfficientNetV2 encoders, plus a SegFormer branch and a small MLP head for hard examples. Its reported design choices include “scale promotion,” which upsamples $10 \times 10$1 patches to $10 \times 10$2, “separate normalization” with per-domain statistics for training versus validation/test domains, Lovász-hinge loss, Online Hard Example Mining, soft cross-entropy for pseudo labels, and progressive self-training. Tanmlh, the third-place solution, used teacher–student self-training, Monte Carlo dropout for uncertainty estimation, mix-up on both labeled and pseudo-labeled samples, and DenseCRF post-processing. The competition paper reports hidden-test scores of $10 \times 10$3 for Kingdrone, $10 \times 10$4 for Tanmlh, and $10 \times 10$5 for the Special Prize solution Sklgp (Ghorbanzadeh et al., 2022).

Subsequent papers repurposed the benchmark for different transfer-learning and architecture questions. A 2023 study based on U-Net feature engineering, residual-convolutional layers, multi-head attention, multi-resolution outputs, and a combined Focal+IoU loss reports F1 $10 \times 10$6 and mIoU $10 \times 10$7 on the Landslide4Sense development set, improving its own proposed U-Net baseline and the challenge baseline under that study’s evaluation setting (Le et al., 2023). A 2024 transfer-learning study, “TransLandSeg,” uses only RGB bands extracted from the benchmark, resizes inputs to $10 \times 10$8, and reports that adaptive transfer learning on SAM trains only 1.3% of SAM’s parameters while achieving Precision $10 \times 10$9, Recall $128 \times 128$0, F1-score $128 \times 128$1, and MIoU approximately $128 \times 128$2 under a withheld-test protocol for that paper (Hou et al., 2024).

Because these later reports rely on different splits, preprocessing choices, input subsets, or evaluation settings, they should be read as protocol-specific results rather than direct replacements for the original ResU-Net baseline or the official competition leaderboard. A plausible implication is that Landslide4Sense evolved from a fixed benchmark into a broader experimental platform spanning leaderboard evaluation, cross-validation on the public development set, and transfer-learning studies based on partial-channel inputs.

5. Foundation models, feature selection, and the problem of input design

A prominent later line of work asks whether foundation-model priors or engineered channel selection improve Landslide4Sense performance under heavy imbalance. In “Clay-CNN Hybrids,” Clay v1.5 is introduced as a Geo-Foundational Model pretrained by masked autoencoding on more than 70 million Earth-observation patches. The study compares three strategies on Landslide4Sense: a Clay-only encoder with multi-scale residual terrain fusion, a U-Net backbone augmented with Clay semantic context at the bottleneck, and a standard U-Net baseline. The reported test F1 values are $128 \times 128$3 for the Clay-only backbone, $128 \times 128$4 for the U-Net baseline, and $128 \times 128$5 for the hybrid U-Net+Clay model over three seeds. The same paper also reports that default $128 \times 128$6 thresholds cause large F1 drops and that a validation-selected threshold of approximately $128 \times 128$7 is used for the best hybrid model (Vu, 12 Jun 2026).

Methodologically, that study argues against using the foundation model as a standalone replacement for multiscale CNN segmentation. Clay processes the 12 optical bands and produces a $128 \times 128$8 feature map, whereas the hybrid architecture injects Clay features into a U-Net bottleneck that still retains four skip connections from a 14-band encoder. Fine-tuning is performed by a two-stage Low-Rank Adaptation procedure: the Clay backbone is frozen during early decoder training and later augmented with LoRA adapters of rank $128 \times 128$9, inserted into every transformer block’s Q/K/V projections and MLP layers. The authors’ interpretation is explicit: pretrained spectral representations are helpful when added as auxiliary context, but they underperform skip-connected CNN decoders when used alone for precise landslide delineation (Vu, 12 Jun 2026).

A different 2026 study addresses Landslide4Sense from the perspective of feature redundancy rather than model scale. “Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data” augments the raw 14 channels with 16 engineered spectral and structural indices, yielding a 30-channel candidate pool. Using Sequential Forward Floating Selection with a lightweight U-Net++ proxy, it reports that an 8-channel subset—B3, B4, B5, B8, B11, B13, B14, and Band 21 (grayscale composite)—achieves test F1 $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$0, compared with $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$1 for the raw 14-band configuration, $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$2 for the full 30-channel configuration, and $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$3 for a 23-channel high-dimensional baseline. The paper explicitly interprets this as evidence of the Hughes Phenomenon and reports that the 8-band configuration reduces input-preprocessing cost by approximately 73% relative to the 30-band model (Ahmad et al., 10 May 2026).

Taken together, these studies frame Landslide4Sense as a benchmark not only for architecture search but also for representational questions: whether global pretrained semantics complement local convolutional detail, which channels contribute nonredundant information, and how threshold calibration interacts with rare-event segmentation. This suggests that the benchmark has become a useful stress test for the spectral–topographic design of landslide models, rather than only a scoreboard for end-to-end F1.

6. Accessibility, applications, and ambiguity of the name

The benchmark is openly distributed at https://www.iarai.ac.at/landslide4sense, where the release includes Sentinel-2 plus ALOS PALSAR $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$4 patches in tif format, pixel-wise ground-truth masks, training/testing splits, pretrained model weights and training scripts for all 11 original architectures, and example evaluation notebooks. The associated 2022 challenge remains accessible through the “Future Development Leaderboard” at https://www.iarai.ac.at/landslide4sense/challenge/, where submissions are automatically scored and ranked in real time (Ghorbanzadeh et al., 2022, Ghorbanzadeh et al., 2022).

The documented use cases are broad but technically specific. They include developing or benchmarking fully supervised, semi-supervised, or self-supervised segmentation models; transfer learning for rapid landslide detection in novel cloud-free scenes; updating multi-temporal landslide inventories; integration into hazard-assessment workflows in remote sensing and GIS environments; and research on multi-source data fusion for natural-hazard detection. In the competition paper, the benchmark is also positioned as a setting for studying generalization to unseen geographies, severe class imbalance, small-object morphology, and inter-scene distribution shift (Ghorbanzadeh et al., 2022, Ghorbanzadeh et al., 2022).

A recurring source of confusion is that the string “LandSlide4Sense” is used in some later summaries for systems that are not the 2022 remote-sensing benchmark. One such use refers to an ASSED-based, end-to-end, real-time landslide detection system that fuses streaming social-media sources with high-confidence physical feeds through a Redis-backed pub/sub pipeline and reports detection of 519 unique 2.5′-grid cells by December 2018 versus 149 for a static keyword classifier, along with F1 $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$5 after four years under concept drift (Suprem et al., 2019). Another refers to a GSM/GPRS wireless sensor-network pilot in Karsingsa, Itanagar, with rain gauge, in-place inclinometer, tilt-meter, piezometer, and crack-meter inputs, reporting a data delivery rate of $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$6, uptime of $$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}, \qquad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}, \qquad \mathrm{F1} = 2 \cdot \frac{\mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$7, and median end-to-end latency of approximately 90 s (Rath et al., 2013). A third refers to an InSAR-based forecasting workflow near Mud Creek that constructs a multilayer network from DEM and satellite deformation data and reports a community-persistence rise beginning approximately 56 days before collapse (Desai et al., 2022).

These systems address landslide monitoring, event detection, or forecasting, but they are methodologically distinct from Landslide4Sense as a multisource, pixel-labeled benchmark for deep-learning segmentation. Distinguishing the benchmark from these unrelated pipelines is important when interpreting reported metrics, because a benchmark F1 on pixel-wise satellite segmentation, a patch-level classification score, a social-sensor event-detection F1, and a geotechnical early-warning latency are not interchangeable quantities.