GEEDEF — Land Cover Classification Tools

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.

#	New Feature	Description
1	📥 Import Training Samples from SHP Asset	Fully new sub-panel in Step 4 — load existing training data from a GEE FeatureCollection asset (uploaded SHP), auto-read attribute columns, extract unique classes, map class codes and names, and merge with manually digitized samples
2	📖 Read Columns	Dedicated button to read and populate attribute column dropdowns from the selected SHP asset — eliminates manual column name entry
3	🔍 Extract Unique Classes	Scans the selected SHP asset for all unique numeric class codes and renders an editable row per class (name + color) before import
4	🔄 Dual Asset Refresh	Separate asset refresh inputs for both the ROI boundary and the training sample SHP — previously only one input field existed
5	⚠️ Accuracy Penalty Flag	If only manual samples are used (no field-validated asset imported), the tool applies a −5% correction to OA and Kappa and displays a warning note — makes the reliability of the assessment transparent
6	🌐 Bilingual UI (EN / ID)	Full interface toggle between English and Bahasa Indonesia via EN / ID buttons on the map — all labels, buttons, placeholders, and status messages switch dynamically
7	📋 Hex Color Copy Box	Each class in the palette panel now renders an editable hex code textbox next to the class button — allows direct copy without needing to memorize or retype the color code

---

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

Requirement	Details
GEE Account	Active account at earthengine.google.com
Internet Connection	Stable connection required for server-side computation
ROI Boundary	As a GEE FeatureCollection asset, or drawn manually on the map
Domain Knowledge	Basic understanding of land cover concepts and satellite image interpretation

---

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

Parameter	Description	Default
Date Range	Start and end date in YYYY-MM-DD format	2024-01-01 to 2024-12-31
Cloud Max (%)	Maximum CLOUDY_PIXEL_PERCENTAGE per scene	20%
Cloud Masking	Enable/disable pixel-level cloud masking	Enabled
Composite Method	Temporal aggregation strategy	Median

Cloud Masking Methods

Method	Description	Recommendation
Cloud Score+	Uses cs_cdf layer; masks pixels with cs_cdf < 0.60	✅ Recommended — highest accuracy (Pasquarella et al., 2023)
SCL (ESA)	Removes SCL classes: cloud shadow (3), unclassified (7), medium cloud (8), high cloud (9), thin cirrus (10), snow/ice (11)	Suitable for most tropical regions
QA60 (Bitmask)	Uses bit 10 (opaque cloud) and bit 11 (cirrus cloud)	Simplest but least precise

Temporal Composite Methods

Method	Description	Best Use
Median (Default)	Pixel-wise median across all scenes in date range	Annual mapping; stable and representative
Median Dry Season	Selects lowest NDVI quartile (25% driest scenes)	Spectral contrast between open land and vegetation
Median Wet Season	Selects highest NDVI quartile (25% greenest scenes)	Maximizing vegetation response

---

Step 3 — Classification Algorithm & Parameters

Random Forest (RF)

Parameter	Range	Default	Description
numTrees	50–1000	500	Number of decision trees
minLeafPopulation	1–20	1	Minimum samples per terminal leaf
bagFraction	0.3–1.0	0.5	Proportion of samples per tree

Gradient Tree Boost (GTB)

Parameter	Range	Default	Description
numTrees	50–500	200	Number of boosting iterations
shrinkage	0.01–0.30	0.05	Learning rate
samplingRate	0.30–1.0	0.7	Proportion of samples per iteration
maxNodes	2–10	5	Maximum depth per tree
loss	—	LeastAbsoluteDeviation	Loss function

Support Vector Machine (SVM)

Parameter	Range	Default	Description
kernel	RBF / LINEAR / POLY / SIGMOID	RBF	Kernel function
cost (C)	0.1–100	10	Regularization parameter
gamma	0.001–2.0	0.5	Influence range (RBF, POLY, SIGMOID)
degree	1–10	3	Polynomial degree (POLY only)
coef0	0.0–5.0	0.0	Independent coefficient (POLY, SIGMOID)

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

Metric	Threshold	Note
Overall Accuracy (OA)	> 85%	Generally accepted for regional land cover studies
Kappa Coefficient	> 0.80	"Very Strong" agreement (Landis & Koch, 1977)

⚠️ 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)

Band	Name	Wavelength	Resolution
B2	Blue	~490 nm	10 m
B3	Green	~560 nm	10 m
B4	Red	~665 nm	10 m
B5	Red Edge 1	~705 nm	20 m
B6	Red Edge 2	~740 nm	20 m
B7	Red Edge 3	~783 nm	20 m
B8	NIR	~842 nm	10 m
B8A	NIR Narrow	~865 nm	20 m
B11	SWIR 1	~1610 nm	20 m
B12	SWIR 2	~2190 nm	20 m

Derived Spectral Indices (16)

Index	Full Name	Primary Application
NDVI	Normalized Difference Vegetation Index	Vegetation density
EVI2	Enhanced Vegetation Index 2	Canopy structure; reduces atmospheric noise
CAI	Cellulose Absorption Index	Bare soil and litter discrimination
NDWI	Normalized Difference Water Index	Leaf water content
GCVI	Green Chlorophyll Vegetation Index	Chlorophyll estimation
SAVI	Soil-Adjusted Vegetation Index	Sparse vegetation over bare soil
PRI	Photochemical Reflectance Index	Canopy light use efficiency
HALL_COVER	Hall Forest Cover Index	Forest canopy cover estimation
IRECI	Inverted Red-Edge Chlorophyll Index	Red-edge based chlorophyll
NDRE	Normalized Difference Red Edge	Vegetation health and stress
MNDWI	Modified Normalized Difference Water Index	Open water body mapping
NDMI	Normalized Difference Moisture Index	Land moisture
NDBI	Normalized Difference Built-up Index	Urban and built-up area
NBR	Normalized Burn Ratio	Burned area / forest degradation
GVS	Green Vegetation Signal	Green vegetation fraction
NDFI	Normalized Difference Fraction Index	Forest degradation and fragmentation

---

Spectral Index Formulas

$$NDVI = \frac{\rho_{B8} - \rho_{B4}}{\rho_{B8} + \rho_{B4}}$$

$$EVI2 = 2.5 \times \frac{\rho_{B8} - \rho_{B4}}{\rho_{B8} + 2.4 \times \rho_{B4} + 1}$$

$$CAI = \frac{\rho_{B12}}{\rho_{B11}} \qquad NDWI = \frac{\rho_{B8} - \rho_{B11}}{\rho_{B8} + \rho_{B11}} \qquad GCVI = \frac{\rho_{B8}}{\rho_{B3}} - 1$$

$$SAVI = \frac{1.5 \times (\rho_{B8} - \rho_{B4})}{\rho_{B8} + \rho_{B4} + 0.5} \qquad PRI = \frac{\rho_{B2} - \rho_{B3}}{\rho_{B2} + \rho_{B3}}$$

$$HALL_COVER = -0.017\rho_{B4} - 0.007\rho_{B8} - 0.079\rho_{B12} + 5.22$$

$$IRECI = \frac{\rho_{B7} - \rho_{B4}}{\rho_{B5} / \rho_{B6}} \qquad NDRE = \frac{\rho_{B8A} - \rho_{B5}}{\rho_{B8A} + \rho_{B5}} \qquad MNDWI = \frac{\rho_{B3} - \rho_{B11}}{\rho_{B3} + \rho_{B11}}$$

$$NDMI = \frac{\rho_{B8} - \rho_{B11}}{\rho_{B8} + \rho_{B11}} \qquad NDBI = \frac{\rho_{B11} - \rho_{B8}}{\rho_{B11} + \rho_{B8}} \qquad NBR = \frac{\rho_{B8} - \rho_{B12}}{\rho_{B8} + \rho_{B12}}$$

$$GVS = \frac{100 \times NDVI}{100|NDVI| + 100|NBR| + 1} \qquad NDFI = \frac{GVS - (1 - NDVI)}{GVS + (1 - NDVI)}$$

---

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

Error / Issue	Solution
"Computation timed out"	Reduce ROI extent; increase scale to 20 or 30 m; shorten the date range
"Not enough pixels in region"	Use 10 m scale or enlarge the sample polygon area
OA very low (< 70%)	Check sample placement; avoid transitional zones; add more samples; merge spectrally similar classes
Class missing from output	Ensure all classes have drawn polygons before clicking Run Classification
Error loading ROI from asset	Verify full asset path; asset must be FeatureCollection; check sharing/access settings
Map layer not appearing	Open Layer Manager (top-right of map); enable all layers; zoom out if ROI is small

---

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. https://doi.org/10.1007/978-3-031-26588-4

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. https://doi.org/10.1016/j.rse.2017.06.031

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). https://doi.org/10.1109/CVPRW59228.2023.00206

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.