Processed data and code for controlled analyses of salt marsh vegetation classification

Description

This archive contains the derived data and analysis code supporting a controlled evaluation of spatial resolution, sensor-related differences, spectral information, feature screening, and temporal enrichment in salt marsh vegetation classification. It includes results from five independent runs of ResNet50, ConvNeXt-Tiny, and Swin Transformer-Tiny on 1 m GF-2 imagery, together with five-run ConvNeXt-Tiny results for the subsequent GF-2 and Sentinel-2 scenarios. The archived numerical materials include run manifests, training histories, validation and test metrics, per-sample predictions, class-specific performance reports, confusion matrices, paired McNemar tests with Holm correction, deep embeddings, handcrafted feature matrices, cross-resolution similarity and separability metrics, and repeated LightGBM/TreeSHAP feature-screening results. Original GF-2 imagery is not redistributed because of data licensing restrictions. Sentinel-2 imagery is publicly accessible, and the corresponding acquisition dates, product information, band configurations, and preprocessing details are included in the dataset.

Steps to reproduce

Run all commands from the root directory of the archive. Python 3.12.13 was used for validation. Create a virtual environment and install the required packages: python -m venv .venv .venv/Scripts/python -m pip install -r 05_code/requirements.txt On macOS or Linux, use `python3` and `.venv/bin/python` instead. Run the following scripts to reproduce the principal numerical results: 05_code/reproduction/reproduce_run_level_metrics.py 05_code/reproduction/reproduce_mcnemar_tests.py 05_code/reproduction/reproduce_representation_metrics.py These scripts reproduce: 1. Five-run means and standard deviations for OA, Kappa, Macro-F1, Weighted-F1, and class-specific F1-scores; 2. Paired McNemar comparisons with Holm-adjusted p-values; and 3. Cross-resolution cosine similarity, class-centroid shifts, silhouette scores, and Davies–Bouldin indices for deep and handcrafted representations. The reproduced files are written to the `reproduced_results` directory and can be compared with the archived results in `01_per_run_results`, `02_confusion_matrices_and_statistical_tests`, and `03_feature_representation_analysis`. Minor floating-point differences may occur because the archived embedding files store finite decimal precision. The scripts reproduce numerical summaries from the archived derived data but do not retrain the networks or recreate full-area classification maps, because raw imagery, label rasters, and model checkpoints are not included.

Institutions

• Nanjing Forestry University

  Jiangsu, Nanjing

Categories

Licence