Multi-Temporal Land Cover Dataset for the Delmarva Peninsula (2000, 2002, 2005, 2009, 2016) at 30m Resolution

Abstract

Coastal regions of the eastern United States are increasingly vulnerable to saltwater intrusion and rising soil salinity, which can significantly alter land cover patterns. Monitoring these changes over time requires consistent, high-quality datasets. This work presents a multi-temporal, classified land cover dataset for the Delmarva Peninsula (14 coastal counties in Delaware, Maryland, and Virginia, USA) for five target years: 2000, 2002, 2005, 2009, and 2016. The datasets were produced using a Random Forest classifier trained on Continuous Change Detection and Classification (CCDC)-derived synthetic Landsat surface reflectance, and target variables dervied from a high-resolution (1m) NAIP–Landsat derived dataset. The classification distinguishes seven land cover classes: Bare Soil, Built, Farmland, Forest, Marsh, Other Vegetation, and Water. Overall testing accuracy reached 90.6% with a five-fold cross-validation accuracy of 90.2% ± 0.13%, demonstrating robust model performance. These datasets can support further studies on land use change, coastal resilience, and salinity-linked landscape dynamics.

Data Description

This dataset contains five integer-coded raster files, one for each study year. Each raster is a single-band GeoTIFF with 30m spatail resolution in which every pixel is classified into one of the following categories:

Files in this collection:

• LandCover_2000.zip
• LandCover_2002.zip
• LandCover_2005.zip
• LandCover_2009.zip
• LandCover_2016.zip

Each zipped folder has a TIFF file with associated suporting files. These TIFF files should be viewed with 'unique values'. For convenience, we have also provided a color file (Delmarva_7LandCover.clr) that can be imported into symbology while visualizing these datasets, thus ensuring a uniform color palette for all files. All datasets are georeferenced, GIS-ready, and compatible with open-source and proprietary software such as QGIS, ArcGIS Pro, R, and Python geospatial libraries.

Study Area

The Delmarva Peninsula spans 14 coastal counties across Delaware, Maryland, and Virginia, bounded by the Chesapeake Bay to the west and the Atlantic Ocean and Delaware Bay to the east. The region contains extensive farmland, forest, marshland, and other vegetation, with agriculture and aquaculture being major economic drivers. Soils vary from well-drained uplands to poorly drained lowlands, influencing hydrology and vegetation patterns.

Input Data

1. Synthetic Landsat Surface Reflectance

   • Generated using the Continuous Change Detection and Classification (CCDC) algorithm (Zhu & Woodcock, 2014).

   • Daily synthetic reflectance models at 30 m resolution created from all available, atmospherically corrected Landsat imagery.

   • Representative seasonal snapshots for January, April, July, and October extracted to capture intra-annual variability.

2. Reference Data for Training and Validation

   • NAIP–Landsat derived land cover dataset (Mondal et al., 2023) for 2016–2017, aggregated from 1 m to 30 m resolution.

   • Original NAIP imagery: four bands (red, green, blue, near-infrared), acquired during peak growing seasons.

   • Seven land cover classes used: Bare Soil, Built, Farmland, Forest, Marsh, Other Vegetation, Water.

3. Spectral Indices dervied from Landsat

   • Normalized Difference Vegetation Index (NDVI)

   • Enhanced Vegetation Index (EVI)

   • Two-band Enhanced Vegetation Index (EVI2)

   • Normalized Burn Ratio (NBR)

   • Tasseled Cap indices: Brightness, Greenness, Wetness

These indices enhance the spectral separability of vegetation, soil, and water classes and improve classification accuracy.

Method

Reference Point Selection

• Land cover classes from the NAIP–Landsat derived 1m dataset were filtered to pure 30 m pixels using a count threshold (= 900 matching 1 m pixels).

• Stability was assessed using CCDC-derived slope parameters for the red and near-infrared bands.

• Stable pixels were defined as having:

  • No change detection events in the year before or after the reference year.

  • Slopes between –0.005 and 0.005, indicating no significant spectral trend.

• Stratified random sampling was applied to select up to 2,000 stable points per class per state, yielding 40,110 reference points in total.

Feature Stack Creation

• For each year, a multiband image stack was created from synthetic surface reflectance for four seasonal dates.

• Original spectral bands, vegetation/water/soil indices, and Tasseled Cap outputs were combined into a single feature set.

Random Forest Classification

• Implemented in Python using scikit-learn.

• Hyperparameter tuning via grid search:

  • n_estimators: 50, 100, 200

  • max_depth: None, 10, 20, 30

  • min_samples_split: 2, 5, 10

  • min_samples_leaf: 1, 2, 4

• Optimal parameters:

  • n_estimators = 100

  • max_depth = None

  • min_samples_split = 2

  • min_samples_leaf = 1

• Accuracy assessment:

  • Hold-out test set (30% of points) → 90.6% overall accuracy

  • Five-fold cross-validation → 90.2% ± 0.13%

  • Out-of-bag (OOB) score → 90.3%

Map Production

• The final trained RF model was applied to each year’s CCDC feature stack to produce classified land cover maps.

• Output rasters were integer-coded according to the LULC scheme above.

Code Availability

Example scripts for preprocessing CCDC outputs and implementing the Random Forest classifier are available at:
https://github.com/Manan-prog/Landcover-classification-using-CCDC-NAIP