GEEDEF-Land-Cover-Classification-Tools

GEEDEF is a cloud-based, interactive land cover classification tool built entirely within the Google Earth Engine (GEE) Code Editor using the GEE JavaScript API. It is designed to accelerate supervised land cover mapping workflow with full GUI panel

Repositori di GitHub · Halaman web

Platform: Google Earth Engine (JavaScript API) | Data: Sentinel-2A SR Harmonized | Version: V3.1

🆕 What’s New in V3.1¶

Updated from V3 (original PDF release). The following features were added in the latest version accessible via the updated GEE link.

Overview¶

GEEDEF is a cloud-based, interactive land cover classification tool built entirely within the Google Earth Engine (GEE) Code Editor using the GEE JavaScript API. It is designed to accelerate supervised land cover mapping workflows at regional and local scales using Sentinel-2A Surface Reflectance (SR) Harmonized imagery.

The tool integrates three scientifically established machine learning classifiers — Random Forest (RF), Gradient Tree Boost (GTB), and Support Vector Machine (SVM) — operating over a 26-variable feature space composed of 10 native spectral bands and 16 derived spectral indices. The entire workflow, from image compositing and sample digitization to accuracy assessment and GeoTIFF export, is executed server-side on Google’s distributed computing infrastructure.

Developed by: Defani Arman Alfitriansyah
Institution: Faculty of Forestry and Environmental Science, Universitas Kuningan
Year: 2025

Table of Contents¶

1. Prerequisites

2. Interface Layout

3. Six-Step Workflow

4. Feature Space

5. Spectral Index Formulas

6. Land Cover Palette Schemes

7. Tips and Troubleshooting

8. Full Script Repository

9. Scientific References

Prerequisites¶

Interface Layout¶

GEEDEF uses a three-panel side-by-side layout:

Six-Step Workflow¶

All six steps must be executed sequentially. Skipping any step will trigger a warning in the status panel.

Step 1 — Define Region of Interest (ROI)¶

The ROI defines the spatial extent for image collection filtering, classifier training, and area calculation. Two methods are available:

Method A — Load from GEE Asset¶

1. Enter your GEE account or project name in the text field (e.g., ee-defaniarman).

2. Click 🔄 Refresh Assets to populate the dropdown with available FeatureCollection assets.

3. Select the boundary SHP from the dropdown.

4. Click 📥 Load ROI from Dropdown.

5. The map centers on the loaded ROI and displays the boundary as a blue outline.

Method B — Manual Drawing¶

1. Activate the Drawing Tools from the top-left map toolbar (pencil icon).

2. Draw a polygon defining your study area boundary.

3. Click 🖊 Use Map Drawing Tools.

Recommendation: For district-scale or larger study areas, Method A is preferred to ensure administrative boundaries match official data.

Step 2 — Sentinel-2A Composite Acquisition¶

Builds a cloud-free image composite from COPERNICUS/S2_SR_HARMONIZED according to user-defined parameters.

Acquisition Parameters¶

Cloud Masking Methods¶

Temporal Composite Methods¶

Step 3 — Classification Algorithm & Parameters¶

Random Forest (RF)¶

Gradient Tree Boost (GTB)¶

Support Vector Machine (SVM)¶

Training / Testing Split¶

Default 70/30. Use 80/20 for datasets with fewer than 200 samples per class. Minimum 20–30 polygons per class required for statistically valid accuracy evaluation.

Step 4 — Add Classes & Digitize Training Samples¶

Option A — Import from SHP Asset (New in V3.1)¶

1. Enter your GEE account/project name and click 🔄 Refresh to load available SHP assets.

2. Select the training sample asset from the dropdown.

3. Click 📖 Read Columns to populate attribute column dropdowns.

4. Select the class code column and the class name/label column.

5. Click 🔍 Extract Unique Classes — an editable row appears for each class (name + color).

6. Edit names and colors as needed, then click 📥 Save & Import Mapped Samples.

Imported samples are merged with any manually digitized samples before training.

Option B — Manual Class Entry¶

1. Type a class name, or click a class from the Left Panel palette for auto-fill.

2. Enter the hexadecimal color code (e.g., #006400).

3. Click ＋ ADD CLASS — codes are assigned automatically and increment sequentially.

Digitizing Samples¶

4. Select the active class from the dropdown.

5. Activate Drawing Tools (top-left of map); choose point or polygon.

6. Draw sample areas on the map for the selected class. Repeat for all classes.

Best practice: Distribute samples spatially. Use multiple small polygons per class to capture within-class spectral variability. Avoid transitional or ambiguous zones.

Step 5 — Run Classification¶

Click 🚀 RUN CLASSIFICATION. The tool executes server-side:

1. Extracts spectral values (26 variables) from all samples via sampleRegions

2. Randomly splits samples into training / validation sets using randomColumn

3. Trains the classifier on the training set

4. Classifies all pixels within the ROI

5. Computes error matrix → derives OA and Kappa from the validation set

6. Calculates per-class area via ee.Image.pixelArea() in hectares

Accuracy Interpretation¶

⚠️ If only manual samples are used (no imported field-validated asset), a −5% penalty is applied to OA and Kappa and flagged in the results panel.

Step 6 — Export to Google Drive¶

1. Set folder name (default: GEE_Export) and filename (default: LandCover_DFRF_V3).

2. Select output spatial resolution: 10 m, 20 m, or 30 m.

3. Click ⬆ EXPORT — GeoTIFF to Drive.

4. Monitor in the Tasks tab (clock icon, top-right of GEE Code Editor). Click Run.

Output is a single-band integer raster (pixel value = class code), CRS: EPSG:4326. Compatible with QGIS, ArcGIS, and all major GIS software.

Feature Space¶

After loading the composite, GEEDEF automatically computes 16 spectral indices. Together with 10 native Sentinel-2A bands, the total feature space is 26 predictor variables per pixel.

Native Spectral Bands (10)¶

Derived Spectral Indices (16)¶

Spectral Index Formulas¶

Land Cover Palette Schemes¶

Color swatches below are rendered as inline SVG — visible on GitHub and most Markdown renderers.

🇮🇩 KLHK — Indonesian Ministry of Environment and Forestry¶

ID	Class Name	Swatch	Hex
1	Hutan Lahan Kering Primer		#60E663
2	Hutan Lahan Kering Sekunder		#72FF00
3	Hutan Mangrove Primer		#8EA704
4	Hutan Mangrove Sekunder		#C1A700
5	Hutan Rawa Primer		#60E663
6	Hutan Rawa Sekunder		#72FF00
7	Hutan Tanaman		#D3E598
8	Semak Belukar		#EBC0A7
9	Semak Belukar Rawa		#EBC0A7
10	Pertanian Lahan Kering		#F6FFA7
11	Pertanian Lahan Kering Campur		#EDF500
12	Sawah		#A8D6FF
13	Tambak		#7AF4F4
14	Permukiman Transmigrasi		#728EA7
15	Perkebunan		#E5D298
16	Permukiman		#686868
17	Bandara / Pelabuhan		#D60073
18	Lahan Terbuka		#D60073
19	Pertambangan		#A70400
20	Badan Air		#D4FCF7
21	Rawa		#98E5E5
22	Savana / Padang Rumput		#D5FF02
23	Awan		#D1D1D1

🌿 MapBiomas Indonesia — Collection 3.0¶

ID	Class Name	Swatch	Hex
3	Formasi Hutan		#1f8d49
5	Mangrove		#04381d
76	Hutan Rawa Gambut		#2f7360
13	Tumbuhan Non-Hutan Lainnya		#d89f5c
40	Sawah		#c71585
35	Sawit		#9065d0
9	Kebun Kayu		#7a5900
21	Pertanian Lainnya		#ffefc3
30	Lubang Tambang		#9c0027
24	Permukiman		#d4271e
25	Non-Vegetasi Lainnya		#db4d4f
31	Tambak		#091077
33	Sungai, Danau, Laut		#2532e4

🇺🇸 NLCD — National Land Cover Database (USGS)¶

ID	Class Name	Swatch	Hex
11	Open Water		#466b9f
12	Perennial Ice/Snow		#d1def8
21	Developed, Open Space		#dec5c5
22	Developed, Low Intensity		#d99282
23	Developed, Medium Intensity		#eb0000
24	Developed, High Intensity		#ab0000
31	Barren Land		#b3ac9f
41	Deciduous Forest		#68ab5f
42	Evergreen Forest		#1c5f2c
43	Mixed Forest		#b5c58f
51	Dwarf Scrub		#af963c
52	Shrub/Scrub		#ccb879
71	Grassland/Herbaceous		#dfdfc2
72	Sedge/Herbaceous		#d1d182
73	Lichens		#a3cc51
74	Moss		#82ba9e
81	Pasture/Hay		#dcd939
82	Cultivated Crops		#ab6c28
90	Woody Wetlands		#b8d9eb
95	Emergent Herbaceous Wetlands		#6c9fb8

Tips and Troubleshooting¶

General Tips¶

• Test with a small ROI first before processing large areas to avoid computation time-outs.

• Use Cloud Score+ as the primary masking method for tropical regions like Indonesia, where thin cirrus and coastal aerosols are common.

• For mangrove studies, use Median Dry Season composite to maximize spectral contrast between mangrove canopy and aquaculture ponds (tambak).

• Distribute samples spatially across the study area. Avoid concentrating all samples in one geographic location.

• For SVM with RBF kernel, perform a manual grid search over C and gamma before the final classification run.

Common Issues¶

Full Script Repository¶

https://code.earthengine.google.com/328e48a66c1f2b0a9268cbce3327b9df

Default Configuration (CFG object)¶

var CFG = {
  startDate:         '2024-01-01',
  endDate:           '2024-12-31',
  cloudMax:          20,
  classProp:         'lc',
  seed:              42,
  scale:             10,
  maxPix:            1e13,
  splitRatio:        0.7,
  // Random Forest
  rf_numTrees:       500,
  rf_minLeaf:        1,
  rf_bagFraction:    0.5,
  // Gradient Tree Boost
  gtb_numTrees:      200,
  gtb_shrinkage:     0.05,
  gtb_samplingRate:  0.7,
  gtb_maxNodes:      5,
  gtb_loss:          'LeastAbsoluteDeviation',
  // Support Vector Machine
  svm_kernel:        'RBF',
  svm_gamma:         0.5,
  svm_cost:          10,
  svm_degree:        3,
  svm_coef:          0.0
};

Cloud Masking Functions¶

function maskByCSPlus(image) {
  return image.updateMask(image.select('cs_cdf').gte(0.60));
}

function maskBySCL(image) {
  var scl = image.select('SCL');
  return image.updateMask(
    scl.neq(3).and(scl.neq(7)).and(scl.neq(8))
       .and(scl.neq(9)).and(scl.neq(10)).and(scl.neq(11))
  );
}

function maskByQA60(image) {
  var qa = image.select('QA60');
  return image.updateMask(
    qa.bitwiseAnd(1 << 10).eq(0).and(qa.bitwiseAnd(1 << 11).eq(0))
  );
}

Export Configuration¶

Export.image.toDrive({
  image:          STATE.classifiedImg,
  description:    'LandCover_DFRF_V3',
  folder:         'GEE_Export',
  fileNamePrefix: 'LandCover_DFRF_V3',
  region:         STATE.roi.geometry(),
  scale:          10,           // options: 10, 20, 30
  crs:            'EPSG:4326',
  maxPixels:      1e13,
  fileFormat:     'GeoTIFF'
});

Scientific References¶

Auriga Nusantara. (2024). Algorithm Theoretical Basis Document (ATBD) MapBiomas Indonesia Koleksi 3 Versi 1. MapBiomas Indonesia.

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.

Cardille, J. A., Crowley, M. A., Saah, D., & Clinton, N. E. (Eds.). (2024). Cloud-based remote sensing with Google Earth Engine: Fundamentals and applications. Springer International Publishing. Springer International Publishing (2024)

Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3), 273–297.

Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. Annals of Statistics, 29(5), 1189–1232.

Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27. Gorelick et al. (2017)

Kementerian Lingkungan Hidup dan Kehutanan Republik Indonesia. (2024). Keputusan Menteri Lingkungan Hidup dan Kehutanan Nomor 399 Tahun 2024 tentang Standar Penyebarluasan Informasi Geospasial Tematik Lingkup KLHK. KLHK.

Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159–174.

Pasquarella, V. J., Brown, C. F., Czerwinski, W., & Rucklidge, W. J. (2023). Comprehensive quality assessment of optical satellite imagery using weakly supervised video learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2125–2135). Pasquarella et al. (2023)

U.S. Geological Survey. (2025). Annual National Land Cover Database (NLCD) Collection 1 Science Product User Guide (Version 1.1). USGS.

U.S. Geological Survey. (n.d.). National Land Cover Database (NLCD) class legend and description. USGS.

This tool is released for academic and educational use. All imagery access is subject to Google Earth Engine Terms of Service and Copernicus Sentinel-2 data policies.