Data and code for evaluating hyperparameter optimization in landslide susceptibility mapping in Xifeng District and Longxi County

Description

This dataset provides processed data, evaluation outputs, and source code supporting a unified framework for hyperparameter-optimized landslide susceptibility mapping in Xifeng District and Longxi County, Gansu Province, China. The study examined conditioning factor selection, optimization gains and efficiency, independent model and map evaluation, and spatial sensitivity across two contrasting loess regions. The modeling datasets contain 214 samples in Xifeng (107 landslide and 107 non-landslide samples) and 294 samples in Longxi (147 of each class). Fourteen candidate conditioning factors were considered in Xifeng and 15 in Longxi. Sample identifiers, class labels, training and test assignments, and cross-validation fold assignments are provided. Six models—logistic regression, support vector machine, random forest, multilayer perceptron, gated recurrent unit, and one-dimensional residual network—were compared using budgeted grid search, random search, the tree-structured Parzen estimator, and a genetic algorithm under a common budget of 50 completed trials. The repository includes multicollinearity diagnostics, candidate factor sets, Shapley additive explanations, optimization summaries for all 96 combinations of region, model, factor set, and optimization method, selected configurations, and available trial logs. Model evaluation files contain held-out predictions, receiver operating characteristic curve coordinates, statistical comparisons, repeated-training results, and out-of-fold predictions for random and spatial cross-validation. Map outputs comprise 24 continuous susceptibility rasters and 24 classified rasters, divided equally between independent evaluation and final mapping. They are accompanied by success rate and prediction rate curve coordinates, class-area statistics, and enrichment ratios with their reciprocal seed cell area indices. Independent evaluation uses maps generated by models fitted to training samples; prediction rate and enrichment measures use held-out landslides. Final maps were generated by refitting models to all regional samples. Their class areas describe regional susceptibility patterns, while full-inventory enrichment statistics are retained as archival descriptions. Source code, environment specifications, data dictionaries, and checksums support verification and reuse. Original sample coordinates and full-grid conditioning-factor inputs are restricted, and trained checkpoints are not included. Access prerequisites and reproduction steps are documented in the repository.

Steps to reproduce

1. Download the complete dataset and preserve its directory structure. Read 06_metadata/README.md and 06_metadata/data_dictionary.md. Verify indexed files using 06_metadata/SHA256_checksums.json. Keep an unchanged archive and run commands in a working copy. 2. From the dataset root, run: conda env create -f 06_metadata/environment.yml conda activate landslide-optuna Install the recorded PyTorch CUDA build as described in 06_metadata/requirements.txt. A GPU is recommended for deep learning. Search times depend on hardware. 3. Use source_code_submission_v1.0 or extract its ZIP. Follow README_code.md to create the required paths and copy samples from 01_modeling_samples. Preserve row order, values, and column names. HPO scripts recreate partitions using seeds 14 (Xifeng) and 17 (Longxi); compare them with the published Split labels. Data S2 records the selected configurations, including the corrected Longxi LR penalty, l1. 4. From source_code_submission_v1.0, run: python code/cal_vif_both.py --package-root .. This reports training-sample VIF without modifying files. Then run: python code/pooled_spatial_auc.py --package-root .. This recalculates pooled AUCs from deposited out-of-fold predictions and overwrites the two regional summary CSVs in the working copy. Use stored predictions without additional rounding. 5. To rerun HPO, use: python code/run_xifeng_seed14.py all python code/ml_optuna_longxi.py all These intensive runs cover six models, two factor sets, and four HPO methods per region, with 50 completed trials per combination. Both commands delete some existing results or models; use a fresh working project. Budgeted-grid candidates and checksums are in 03_hpo_records/Supplementary_Data_S1_budgeted_grid_candidates.zip. 6. Verify test AUCs and statistical comparisons from 04_model_evaluation. Follow 05_map_evaluation/README_SP_curves.md and map_source_manifest.json for map outputs. Integrate cumulative curve coordinates by the trapezoidal rule after converting percentages to fractions. In Data S5, independent ER equals held-out landslide percentage divided by class area percentage. Compute valid map area from class codes 1-5, excluding NoData. Final maps fitted to all regional samples and their class areas are separate descriptive outputs; ER_full is archival. 7. Deposited classified rasters use Jenks natural breaks in ArcGIS Pro, not the scripts' preliminary classifications. Original coordinates and full-grid factor inputs are restricted, and trained checkpoints are not included. The package supports numerical verification but cannot alone regenerate the complete spatial workflow. See README_code.md for prerequisites. Restricted inputs may be requested from the corresponding author, subject to approval by the Department of Natural Resources of Gansu Province.

Institutions

• Lanzhou Jiaotong University

  Gansu, Lanzhou

Categories

Licence