02EarthEmbeddings - Earth Embeddings Lab

User manual, version 0.1.1 (2026-11-16). QGIS 3.28 to 4.x. GPL-2.0-or-later. Yusuf Eminoğlu (Geophilo).

1. What the plugin does

02EarthEmbeddings works with the Google AlphaEarth Foundations Satellite Embedding dataset in QGIS. Every 10 m pixel of every year from 2017 to 2025 carries one 64-dimensional unit vector x^∈R64\hat{\mathbf x} \in \mathbb{R}^{64}, ∥x^∥=1\lVert \hat{\mathbf x} \rVert = 1, that summarises a year of multi-sensor observation (Brown et al. 2025). Mapping then becomes a matter of comparing vectors: places that look alike have a large cosine x^⊤y^\hat{\mathbf x}^{\top}\hat{\mathbf y}.

The plugin reads the public Cloud-Optimised GeoTIFFs without an Earth Engine account and carries a study from the embeddings to results you can defend: example-based similarity search, label-free spherical clustering with model-selection diagnostics, few-label classification with spatial block cross-validation and an area of applicability, design-based accuracy and area estimates with 95 % confidence intervals, multi-year change with calibrated thresholds, polygon embeddings for planning units, and a self-contained HTML report that names every formula and reference. The same chain runs in two ways: the lab dock (page by page, with charts and styled layers) and the Processing provider (one algorithm per step plus a whole-chain algorithm, for models and batch runs). Only qgis.core, GDAL, NumPy and PyQt are needed.

Why it exists: the QGIS tools that use the embeddings each cover part of a study. EMBED-CD does year-over-year change; AlphaEarth Toolbox needs scikit-learn for clustering and classification and reports random k-fold accuracy, which overstates accuracy on spatially autocorrelated pixels; eo_inference_qgis runs other foundation models; GeoAI includes a downloader inside a deep-learning stack; Earth Engine plugins need an account. Outside QGIS, aef-loader and the R package spatial-tools/alphaearth only search and retrieve data. 02EarthEmbeddings covers the chain from data to defended answer in one place.

2. Install

  1. QGIS: Plugins > Manage and Install Plugins, search 02EarthEmbeddings, install; or Install from ZIP with a release zip.
  2. A toolbar button and the menu Plugins > 02EarthEmbeddings open the lab; the Processing toolbox lists the algorithms under 02EarthEmbeddings.
  3. Optional packages: scikit-learn enables the random-forest classifier; pyarrow is a second way to read the Parquet tile index when GDAL has no Parquet driver. The dock shows the pip command for both (it never runs a process itself). Everything else works without them.

3. Data, units and working CRS

Dataset. AlphaEarth Foundations Satellite Embedding V1, annual, 2017 to 2025, 10 m, 64 channels A00 to A63, CC BY 4.0. Sources: Google Cloud (official; anonymous HTTPS), the Source Cooperative mirror (same relative paths and identical values), or a local folder laid out as <year>/<zone>/<file> with an aef_index file (GeoPackage, Parquet or CSV). The tile index is read with GDAL's Parquet driver, with pyarrow, or from the local index, and cached per source, year and 1-degree cell.

Reading. Each channel is a signed byte qcq_c de-quantised and normalised by equation (1):

xc=sign⁡(qc)(qc127.5)2,qc∈[−127,127],x^=x∥x∥(1)x_c = \operatorname{sign}(q_c)\left(\frac{q_c}{127.5}\right)^{2},\quad q_c \in [-127, 127],\qquad \hat{\mathbf{x}} = \frac{\mathbf{x}}{\lVert \mathbf{x} \rVert} \tag{1}

Tiles are stored bottom-up (positive pixel height) in UTM zones. They are read only through gdal.Warp, which turns them north-up and mosaics zones into one working CRS. Nodata is −128-128 in the cube and NaN (or a declared value) in float outputs, never 0 or −1-1.

Working CRS. The UTM zone of the AOI centre, or the AOI CRS when it is projected in metres and keep CRS is on. All areas are in square metres, hectares or km2^2; pixel sizes are 10, 20, 40 or 80 m (any size of 10 m or coarser in Processing). Vector inputs in any CRS (including EPSG:4326) are reprojected to the cube CRS before sampling.

Estimate before download. The Data page and Estimate download show pixels, tiles, zones, missing years and the size B=w⋅h⋅64⋅0.82⋅nyearsB = w \cdot h \cdot 64 \cdot 0.82 \cdot n_{\mathrm{years}} bytes before anything is fetched. An AOI is limited to 25×10625 \times 10^{6} pixels per year (about 50×5050 \times 50 km at 10 m); above that a coarser pixel size is suggested.

The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind. The sentence is written into every layer's metadata, the report and the Help page.

4. The lab dock, page by page

The lab is a dock on the right with a dark icon rail of eight pages and Help. Long steps run in QGIS background tasks with a progress bar and Cancel; the dock never freezes and no network read runs on the GUI thread. Results go to a 02EarthEmbeddings layer group with a subgroup per run. Settings (source, cache folder, AOI mode, pixel size) are kept in the QGIS settings, never in project files.

Data

Area of interest from the canvas extent, a polygon layer or a rectangle drawn on the map; year chips 2017 to 2025; pixel-size chips 10 / 20 / 40 / 80 m; source chips (Google Cloud, Source Cooperative, local folder or raster layers). Check connection queries the index and shows its tier; the CRS badge shows the working CRS and why. Estimate fills the estimate card (pixels, size, tiles, zones, warnings), then Fetch cube downloads. Open existing cube reads a folder with cube.json or 64-band GeoTIFFs; a raster in geographic degrees (an Earth Engine export in EPSG:4326, say), in non-metre units or in Web Mercator is reprojected on the fly into the UTM zone of its centre (nearest neighbour, whole-metre pixels) so areas and block sizes are in metres, and the status line says so; Load demo cube (synthetic) writes the demo of section 5. KPI cards: years, area in km2^2, valid share and the mean vector norm ∥x∥‾\overline{\lVert \mathbf x \rVert} (expected ≈1.00\approx 1.00; a value far from 1 means the rasters are not AlphaEarth embeddings or were rescaled). The PCA colour composite is added automatically.

Similar

Positive and negative examples from map clicks (two map tools, markers coloured by role, a list with remove buttons) or from point or polygon layers. Mode chips mean / max / contrast (equations (3) to (5)); threshold chips Otsu / top % / fixed with a live histogram and a draggable threshold line; minimum mapping unit. KPIs: threshold, matched area in ha, regions, best score; a ranked region list zooms the map on click.

Clusters

kk range, sample size and seed. Find k draws the silhouette, Calinski-Harabasz and Davies-Bouldin curves with the chosen kk highlighted; Run clustering maps the clusters. Cluster cards show the colour, share, area, coherence Rˉ\bar R and the nearest cluster; clusters can be renamed in place and the names carry into the report.

Classify

Training layer and class field, or quick labels (pick a class, click the map). Samples-per-class bars warn about imbalance. Classifier chips k-NN / prototype / logistic / random forest (greyed out with the install command when scikit-learn is missing); validation chips (spatial blocks, random or both), block size and folds. When every sample falls in one block the run stops and names a remedy (smaller blocks, samples further apart or random validation); when there are fewer blocks than folds, the folds are reduced to the number of blocks. KPIs: OA (blocks), OA (random), the optimism gap Δopt\Delta_{\mathrm{opt}}, macro-F1 and the share inside the AOA. A confusion-matrix heat map; probability, margin, DI and AOA layers. The report (and the Processing class table) gives the mapped area of every class and the part of it outside the AOA, so a class that absorbs an unknown land cover shows up by its high outside share.

Accuracy

Choose the class raster, the total nn and the minimum per class, then Create sample. Size n for this precision sets the total from a target standard error S(O^)S(\hat{O}) of the overall accuracy and an expected user's accuracy UiU_i for every class, with the mapped shares WiW_i of the class raster (equation (21)): S(O^)=0.01S(\hat{O}) = 0.01 means about ±2\pm 2 points at 95 %. A labelling helper steps through the points (previous, next, zoom) and writes the chosen class to ref_class, with progress. Estimate areas fills a table of mapped area versus estimated area ±\pm 95 % CI, UA and PA with intervals, OA ±\pm CI, and a bar chart with intervals. Every mapped class needs labelled points: a class whose stratum has none gets no estimate (an empty cell, not 0), the estimate covers only the labelled strata, and the status line gives the mapped area left out.

Change

Two or more years, threshold chips (Otsu, MAD, stable reference), an optional stable-mask layer and the minimum mapping unit. KPIs: changed area (ha and %), threshold, the year with most change; a distance histogram; distance, change-year and change-region layers and a before/after PCA pair. Both images of the pair use one PCA basis and one colour stretch, fitted on the two years pooled, so a pixel whose embedding did not change keeps its colour and a colour difference marks a change.

Zones

Polygon layer and id field, year, minimum valid share. KPIs: zones and mean coherence. Pick a reference zone to write and colour sim_to_ref; Export similarity matrix (CSV) writes the zone-by-zone cosine matrix.

Report

Writes the HTML report: data provenance with the tile list, attribution and licence, every method used with its equation (typeset with KaTeX, offline) and reference, the KPIs, tables and charts of each run, and the caveats. Export GPKG and rasters copies every result raster to a folder and writes the vectors into one GeoPackage.

Help

A short guide, method references, the data licence and attribution, the cache folder size with Clear cache, and links to this manual and the repository.

5. Worked example with the demo cube

The demo cube is synthetic, so every result can be checked against its planted truth: 400×300400 \times 300 pixels of 10 m in EPSG:32635 (12 km2^2), years 2022 to 2024 (the dock's demo; the Processing demo defaults to 2022 and 2023), five classes plus one novel class, a 60×4060 \times 40 pixel block (24 ha) that changes class in the last year (2024 in the dock), and an 8×88 \times 8 nodata corner in the last year. Class 4 is made confusable with class 3 (cos⁡(μ3,μ4)=0.35\cos(\boldsymbol\mu_3, \boldsymbol\mu_4) = 0.35).

  1. Data: press Load demo cube (synthetic). The KPI cards show 3 years, 12.0 km2^2 and a mean norm close to 1.00; the PCA composite shows the Voronoi patches.
  2. Similar: click ten pixels inside one patch, keep mean and Otsu, run. The match covers that class: on the demo, precision and recall against the truth are 1.0. Add negatives on the confusable class and switch to contrast: the match stays exact.
  3. Clusters: Find k over 3 to 7; the silhouette peaks at k=5k = 5 (about 0.87). Run clustering with k=5k = 5: the adjusted Rand index against the known classes is 1.0.
  4. Classify: use the demo training points (field class, 10 per class), k-NN, block size 500 m. OA (blocks) is 1.0 on the demo; the AOA flags the novel class (which has no training points) as outside, and about 0.5 % of known-class pixels.
  5. Accuracy: create a sample of 150 points with at least 20 per class, label it (on the demo, from the truth raster), estimate: the estimated class areas sum to the mapped area and OA is reported with its interval.
  6. Change: years 2022 to 2024, Otsu: the changed area is 24.0 ha, the change year is 2024 inside the block, 2024 is the year with most change, and the nodata corner stays nodata.
  7. Zones: draw or load two polygons, run, pick one as reference: it gets sim_to_ref = 1.
  8. Report: write the report; it carries the SYNTHETIC note on every section.

The same example in the Python console with Processing:

import processing
d = processing.run("zero2earthembeddings:demo_cube",
                   {"YEARS": "2022, 2023", "OUTPUT_FOLDER": "C:/tmp/demo", "TRAINING": "TEMPORARY_OUTPUT"})
r = processing.run("zero2earthembeddings:earth_embeddings_lab",
                   {"CUBE": "C:/tmp/demo", "YEAR": 2022, "K": 0, "TRAINING": d["TRAINING"],
                    "CLASS_FIELD": "class", "BLOCK_SIZE": 500, "CHANGE": True,
                    "OUTPUT_FOLDER": "C:/tmp/demo_lab", "REPORT": "C:/tmp/demo_lab/report.html"})
print(r["K_USED"], r["OA_BLOCKS"], r["CHANGED_HA"])   # 5 1.0 24.0 on the demo

6. Processing algorithms

Provider zero2earthembeddings, 13 algorithms in seven groups. Every algorithm opens the cube through the same staged core API as the dock (a cube folder with cube.json, or 64-band year rasters whose year is read from the file name), copies vector inputs and reprojects them to the cube CRS, writes rasters with GDAL (nodata and band names kept) and vector or table outputs with real field types (integers, doubles, text; NULL where a value is undefined). Outputs are not styled by the algorithms. Year 0 means the latest year of the cube. Parameters marked advanced are under Advanced parameters.

Synthetic demo cube Data

Id: zero2earthembeddings:demo_cube

Writes a deterministic synthetic cube with planted truth (marked SYNTHETIC - not AlphaEarth data in layer names, metadata and the report): a Voronoi layout of five land-cover classes plus one novel class that has no training points, a 60×4060 \times 40 pixel block that changes class between the last two years, and an 8×88 \times 8 nodata corner in the last year. Each pixel is its class prototype plus Gaussian noise of standard deviation σ\sigma (default σ=0.35\sigma = 0.35), normalised and quantised with the inverse of equation (1), so it is read exactly like AlphaEarth data.

ParameterDescriptionTypeDefaultNotes
YEARSYearsText2022, 2023
WIDTHWidth (pixels)Number400advanced min 64, max 4000
HEIGHTHeight (pixels)Number300advanced min 64, max 4000
NOISENoise (standard deviation)Number0.35advanced min 0, max 2
SEEDSeedNumber7advanced min 0
OUTPUT_FOLDERCube folderFolder output-
TRAININGTraining points (synthetic)Vector / table output-
TRUTHTruth classes (synthetic)Raster output-optional

Outputs. Cube folder; training points (field class, 10 per known class); truth class raster (Int16, −1-1 = nodata); the id of the novel class.

Result keys in the dictionary returned by processing.run: NOVEL_CLASS - novel class id (absent from training).

Estimate download (AlphaEarth tiles) Data

Id: zero2earthembeddings:estimate_download

Looks up the tiles of the source's index that cover the area of interest and returns the grid in the working CRS, the number of tiles, the UTM zones and the download estimate B=w⋅h⋅64⋅0.82⋅nyearsB = w \cdot h \cdot 64 \cdot 0.82 \cdot n_{\mathrm{years}} bytes (0.82 is the measured compressed size of one Int8 band value). Nothing is downloaded. Years or tiles missing from the index are listed in the warnings.

ParameterDescriptionTypeDefaultNotes
EXTENTArea of interestExtent-
YEARSYears (for example 2024, or 2018-2024)Text2024
RESOLUTIONPixel size (m, 10 or coarser)Number10min 10, max 10000
SOURCEData sourceChoiceGoogle Cloud (official)Options: Google Cloud (official); Source Cooperative mirror; Local folder (<year>/<zone>/<file> + index)
LOCAL_FOLDERLocal folder (local source only)File / folder-optional
KEEP_CRSKeep the AOI CRS when it is projected in metresBooleannoadvanced

Outputs. Pixels per year, width, height, tiles, estimated bytes, working EPSG code, warnings.

Result keys in the dictionary returned by processing.run: PIXELS - pixels per year; WIDTH - width (pixels); HEIGHT - height (pixels); TILES - tiles; EST_BYTES - estimated download (bytes); WORKING_EPSG - working CRS (EPSG code); WARNINGS - warnings.

Fetch AlphaEarth embeddings Data

Id: zero2earthembeddings:fetch_embeddings

Reads the planned tiles and writes one north-up 64-band Int8 GeoTIFF per year (aef_<year>_<res>m.tif, bands A00 to A63, nodata −128-128, DEFLATE, 512-pixel tiles) and cube.json. AlphaEarth tiles are stored bottom-up (positive pixel height); they are read only through gdal.Warp, which turns them north-up and mosaics several UTM zones into the working CRS. A missing year or tile becomes nodata, never zeros. Pixel sizes above 10 m read Google's averaged overviews. The output folder is the Cube folder of every other algorithm.

ParameterDescriptionTypeDefaultNotes
EXTENTArea of interestExtent-
YEARSYears (for example 2024, or 2018-2024)Text2024
RESOLUTIONPixel size (m, 10 or coarser)Number10min 10, max 10000
SOURCEData sourceChoiceGoogle Cloud (official)Options: Google Cloud (official); Source Cooperative mirror; Local folder (<year>/<zone>/<file> + index)
LOCAL_FOLDERLocal folder (local source only)File / folder-optional
KEEP_CRSKeep the AOI CRS when it is projected in metresBooleannoadvanced
OUTPUT_FOLDERCube folderFolder output-
PCAPCA colour composite (latest year)Raster output-optional

Outputs. Cube folder; years written; width and height; optional PCA composite of the latest year.

Result keys in the dictionary returned by processing.run: YEARS_WRITTEN - years written; WIDTH - width (pixels); HEIGHT - height (pixels).

PCA colour composite Data

Id: zero2earthembeddings:pca_rgb

Principal components (equation (2)) of a random sample of valid unit vectors; the first three scores of every pixel are stretched between their 2nd and 98th percentiles and written as red, green and blue (values 1 to 255, 0 = nodata). Similar embeddings get similar colours. The explained-variance shares ρ1,ρ2,ρ3\rho_1, \rho_2, \rho_3 are stored in the raster metadata item PCA_EXPLAINED.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
N_SAMPLEPCA sample size (pixels)Number50000advanced min 100
SEEDSeedNumber42advanced min 0
OUTPUTPCA colour compositeRaster output-

Outputs. 3-band Byte raster; explained-variance shares of PC1 to PC3.

Id: zero2earthembeddings:similarity_search

Scores every pixel by cosine similarity to example locations, with the mean prototype (equation (3)), the nearest example (equation (4)) or the contrast with negative examples (equation (5)). Points contribute the pixel under them and polygons all their valid pixels. The score is cut by Otsu's threshold (equation (6)), a top percent, or a fixed value; for the contrast score Otsu is applied to the pixels with s>0s > 0 only, because s≤0s \le 0 already prefers the negative prototype. Patches smaller than the minimum mapping unit are removed (8-connected sieve) and the rest are polygonised and ranked by mean score.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
POSITIVESPositive examplesVector features-
NEGATIVESNegative examplesVector features-optional
MODESimilarityChoiceMean prototypeOptions: Mean prototype; Nearest example (max); Contrast with negative examples
THRESHOLDThreshold ruleChoiceOtsuOptions: Otsu; Top percent; Fixed value
TOP_PERCENTTop percent (top-percent rule)Number1min 0.001, max 100
FIXEDFixed threshold (fixed rule)Number-optional min -2, max 2
MMUMinimum mapping unit (pixels)Number8min 0
SCORESimilarity scoreRaster output-
MASKMatch maskRaster output-
REGIONSMatched regionsVector / table output-

Outputs. Score (Float32, NaN = nodata); match mask (Byte 0/1, 255 = nodata); region polygons with rank, area_ha, mean_score, max_score; threshold, matched area (ha), number of regions.

Result keys in the dictionary returned by processing.run: THRESHOLD_VALUE - threshold; AREA_HA - matched area (ha); N_REGIONS - regions.

Choose the number of clusters Clusters

Id: zero2earthembeddings:choose_k

Fits spherical k-means (equation (8)) for every kk in the range on one random sample and returns the cosine silhouette (equation (9), on a 4 000-pixel subsample), the Calinski-Harabasz and Davies-Bouldin indices on chordal distance (equation (10)) and the objective JJ. The chosen kk has the largest silhouette; ties go to the smaller kk.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
K_MINSmallest kkNumber2min 2, max 50
K_MAXLargest kkNumber10min 3, max 50
N_SAMPLESample size (pixels)Number20000min 100
SEEDSeedNumber42advanced min 0
OUTPUTkk diagnosticsVector / table output-

Outputs. Table k, silhouette, ch, db, objective; the chosen kk.

Result keys in the dictionary returned by processing.run: CHOSEN_K - chosen k.

Label-free clusters (spherical k-means) Clusters

Id: zero2earthembeddings:spherical_clusters

Spherical k-means (equation (8)) with k-means++ seeding fitted on a random sample of valid pixels; every pixel is then assigned to the most similar centroid, 512 x 512 pixels at a time. With k=0k = 0 the number of clusters is chosen as in Choose the number of clusters. The profile table gives the pixel count, share and area of each cluster, its coherence Rˉ\bar R (equation (20) over the members) and the nearest other cluster with its cosine.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
KNumber of clusters (0: choose by silhouette)Number0min 0, max 50
K_MINSmallest kk (when choosing)Number2advanced min 2, max 50
K_MAXLargest kk (when choosing)Number10advanced min 3, max 50
N_SAMPLESample size (pixels)Number100000min 100
SEEDSeedNumber42advanced min 0
MMUMinimum mapping unit (pixels, 0: none)Number0advanced min 0
OUTPUTClustersRaster output-
PROFILESCluster profilesVector / table output-

Outputs. Cluster raster (Byte 0…k−10 \dots k-1, 255 = nodata); profile table cluster, pixels, share, area_ha, coherence, nearest, nearest_cos; kk and the silhouette.

Result keys in the dictionary returned by processing.run: K_USED - k; SILHOUETTE - silhouette (cosine).

Few-label classification Classification

Id: zero2earthembeddings:classify

Trains a classifier on the embeddings under labelled points (or random valid pixels inside labelled polygons) and maps every pixel: k nearest neighbours with similarity-weighted votes (equation (11), k=min⁡(k,nmin⁡)k = \min(k, n_{\min}) with nmin⁡n_{\min} the smallest class count), nearest prototype with a softmax (equation (12)), multinomial logistic regression (equation (13)) or a random forest (equation (14), only when scikit-learn is installed). Accuracy is estimated twice from the same samples: spatial block cross-validation (square blocks of the block size assigned to folds by a seeded greedy balance of sample counts, every sample of a block in the same fold) and random k-fold, so the optimism gap of equation (15) is visible. Quantity and allocation disagreement (equation (16)) replace kappa. The area of applicability (equation (17)) marks pixels whose embedding is farther from the training data than training samples are from each other across folds.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
TRAININGTraining points or polygonsVector features-
CLASS_FIELDClass fieldField-
CLASSIFIERClassifierChoicek nearest neighbours (cosine)Options: k nearest neighbours (cosine); Nearest prototype; Multinomial logistic regression; Random forest (needs scikit-learn)
KNN_Kkk of k-NNNumber5advanced min 1, max 100
BLOCK_SIZEBlock size of the spatial cross-validation (m)Number1000min 1
N_FOLDSFoldsNumber5min 2, max 20
SEEDSeedNumber42advanced min 0
AOAArea of applicabilityBooleanyes
PER_POLYGONPixels per training polygonNumber50advanced min 1
OUTPUTClassesRaster output-
PROBABILITYWinning-class probabilityRaster output-optional
MARGINMargin (top-1 minus top-2)Raster output-optional
DIDissimilarity indexRaster output-optional
AOA_MASKArea of applicability (1 = inside)Raster output-optional
CLASSESClass tableVector / table output-
CONFUSIONConfusion matrix (block CV)Vector / table output-

Outputs. Class raster (Byte class index 0…K−10 \dots K-1, 255 = nodata) with the class table value, class, n_train; winning-class probability; margin (top-1 minus top-2 probability); dissimilarity index; AOA mask (1 inside, 0 outside); block-CV confusion matrix (rows = reference); OA (blocks), OA (random), optimism gap, macro-F1, AOA share.

Result keys in the dictionary returned by processing.run: OA_BLOCKS - overall accuracy, block CV; OA_RANDOM - overall accuracy, random CV; OPTIMISM_GAP - optimism gap (random - blocks); MACRO_F1 - Macro-F1, block CV; AOA_SHARE - share of pixels inside the AOA.

Accuracy sample (stratified random) Accuracy and area

Id: zero2earthembeddings:accuracy_sample

Draws a stratified random sample of pixel centres on the classes of a map, the design of Olofsson et al. (2014): every class gets at least the minimum nmin⁡n_{\min}, and the rest of the total nn is shared out in proportion to the mapped area. With TARGET_SE above 0 the total is computed from equation (21) instead of N_TOTAL. Fill the empty ref_class field with the reference class of each point (the same codes as the map), then run Area estimate.

ParameterDescriptionTypeDefaultNotes
INPUTClass rasterRaster layer-
N_TOTALTotal sample sizeNumber300min 2
MIN_PER_CLASSMinimum per classNumber50min 1
SEEDSeedNumber42advanced min 0
TARGET_SETarget standard error of OA (0: use the total sample size)Number0advanced min 0, max 0.2
EXPECTED_UAExpected user's accuracy (for the target SE)Number0.85advanced min 0.5, max 0.99
OUTPUTAccuracy sampleVector / table output-

Outputs. Points sample_id, map_class, empty ref_class in the map CRS; number of points.

Result keys in the dictionary returned by processing.run: N_POINTS - sample points; N_TOTAL_USED - total sample size used.

Area estimate and accuracy (Olofsson) Accuracy and area

Id: zero2earthembeddings:area_estimate

Design-based estimates (equation (18)) from the labelled sample: the error matrix in area proportions p^ij\hat p_{ij}, the area A^j\hat A_j of each class with its 95 % confidence interval, and overall, user's and producer's accuracy with their intervals. Points with an empty reference, outside the map or on nodata are skipped and counted; a stratum with fewer than two labelled points has no variance and is noted in the log.

ParameterDescriptionTypeDefaultNotes
INPUTClass rasterRaster layer-
REFERENCELabelled sample pointsVector features-
REF_FIELDReference class fieldFieldref_class
OUTPUTArea estimatesVector / table output-

Outputs. Table per class: map_area_ha, area_ha, area_ci95_ha, users, users_ci95, producers, producers_ci95, n; OA, its 95 % half-width, labelled points used.

Result keys in the dictionary returned by processing.run: OA_CI95 - overall accuracy, 95 percent half-width; N_LABELLED - labelled points used.

Change between years Change

Id: zero2earthembeddings:change_detection

The cosine distance dtd_t between consecutive years (equation (19)). The distance raster holds max⁡tdt\max_t d_t; it is thresholded with Otsu (equation (6)), with the median plus kk MAD (equation (7)), or with the 99th percentile of dd inside a stable-mask polygon layer. Patches below the minimum mapping unit are removed and the rest are polygonised. The change year is yt∗+1y_{t^{*}+1}; it is NaN outside the change mask, so an unchanged pixel has no year. Annual embeddings show that change happened, not when within the year, and an unusual season can raise dd.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARSYears (empty: all years of the cube)Text-optional
THRESHOLDThreshold ruleChoiceOtsuOptions: Otsu; Median + k MAD; 99th percentile inside a stable-mask layer
MAD_Kkk of the MAD ruleNumber3min 0.1, max 20
STABLE_MASKStable-mask polygons (stable rule)Vector features-optional
MMUMinimum mapping unit (pixels)Number10min 0
DISTANCEChange distanceRaster output-
CHANGE_YEARChange yearRaster output-optional
MASKChange maskRaster output-optional
REGIONSChange regionsVector / table output-

Outputs. Distance (Float32, max⁡tdt\max_t d_t); change year (Float32, NaN outside the mask); change mask (Byte 0/1, 255 = nodata); regions area_ha, mean_d, max_d, year_from, year_to; threshold, changed area (ha and percent), the year with most change.

Result keys in the dictionary returned by processing.run: THRESHOLD_VALUE - threshold; CHANGED_HA - changed area (ha); CHANGED_PCT - changed share (percent of valid pixels); YEAR_MOST_CHANGE - year with most change.

Polygon embeddings (zones) Zones

Id: zero2earthembeddings:zone_embeddings

The polygon embedding of equation (20): the normalised mean e^\hat{\mathbf e} of the valid unit vectors inside each polygon (fields e00 to e63) and the coherence Rˉ∈[0,1]\bar R \in [0, 1], 1 for a uniform zone. Each polygon is rasterised on its own, so overlapping polygons keep all their pixels. The valid share is the number of valid pixels over the polygon area in pixels; below the minimum the embedding and coherence are NULL. With a reference zone, sim_to_ref =e^⊤e^ref= \hat{\mathbf e}^{\top}\hat{\mathbf e}_{\mathrm{ref}}. The CSV matrix holds the cosine between every pair of zones.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
INPUTZones (polygons)Vector features-
ID_FIELDZone id fieldField-optional
MIN_VALID_SHAREMinimum valid shareNumber0.5min 0, max 1
REFERENCEReference zone (id value; empty: none)Text-optional
OUTPUTZone embeddingsVector / table output-
MATRIXSimilarity matrix (CSV)File output-optional

Outputs. Zones with e00 to e63, n_px, valid_share, coherence, sim_to_ref; optional CSV matrix; zones with an embedding and the mean coherence.

Result keys in the dictionary returned by processing.run: ZONES_VALID - zones with an embedding; MEAN_COHERENCE - mean coherence.

Earth Embeddings Lab (full workflow and report) Workflow

Id: zero2earthembeddings:earth_embeddings_lab

The whole chain on one cube through the staged core API, then the HTML report: the PCA composite, clusters (kk chosen by silhouette when k=0k = 0, range 2 to 10), and, when their inputs are given, similarity search from the positive examples (mean prototype, Otsu), classification with block cross-validation and the area of applicability, change between all years of the cube (two or more) and zone embeddings. Every raster is written to the output folder; the report lists every method used with its equation and reference.

ParameterDescriptionTypeDefaultNotes
CUBECube folder (with cube.json)File / folder-optional
CUBE_LAYERSor 64-band year rastersRaster layers (several)-optional
YEARYear (0: the latest year of the cube)Number0min 0, max 2100
KNumber of clusters (0: choose by silhouette)Number0min 0, max 50
POSITIVESPositive examples (optional)Vector features-optional
TRAININGTraining layer (optional)Vector features-optional
CLASS_FIELDClass fieldField-optional
BLOCK_SIZEBlock size of the spatial cross-validation (m)Number1000min 1
CHANGEChange between the years of the cubeBooleanyes
ZONESZones (optional)Vector features-optional
ZONE_ID_FIELDZone id fieldField-optional
OUTPUT_FOLDEROutput folderFolder output-
REPORTHTML reportFile output-

Outputs. Output folder with every raster; the HTML report; kk, OA (blocks), changed area, the stages run.

Result keys in the dictionary returned by processing.run: K_USED - k; OA_BLOCKS - overall accuracy, block CV; CHANGED_HA - changed area (ha); STAGES - stages run.

7. Methods and equations

The registry core/analytics/__init__.py holds every method with its LaTeX and number; the HTML report shows the same equations with the same numbers. All vectors are unit vectors after equation (1).

(1) De-quantisation

xc=sign⁡(qc)(qc127.5)2,qc∈[−127,127],x^=x∥x∥(1)x_c = \operatorname{sign}(q_c)\left(\frac{q_c}{127.5}\right)^{2},\quad q_c \in [-127, 127],\qquad \hat{\mathbf{x}} = \frac{\mathbf{x}}{\lVert \mathbf{x} \rVert} \tag{1}

AlphaEarth stores each of the 64 channels as a signed byte qcq_c. The value −128-128 is nodata; a pixel with any channel at −128-128 is invalid. After de-quantisation the vector is normalised, so every method works on unit vectors x^\hat{\mathbf x}. The inverse used by the demo and the tests is qc=round⁡(sign⁡(xc)∣xc∣⋅127.5)q_c = \operatorname{round}(\operatorname{sign}(x_c)\sqrt{\lvert x_c \rvert}\cdot 127.5). Real AlphaEarth vectors have ∥x∥≈1\lVert \mathbf x \rVert \approx 1 before normalisation; the Data page shows the mean norm as a health check. Never average the raw integers (for example with gdaladdo -r average): the coding is not linear. Reference: Google 2026; Brown et al. 2025. Output: unit vectors.

(2) PCA colour composite

X−xˉ=UΣV⊤,ρc=σc2∑lσl2(2)\mathbf{X} - \bar{\mathbf{x}} = \mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^{\top},\qquad\rho_c = \frac{\sigma_c^{2}}{\sum_{l} \sigma_l^{2}} \tag{2}

X\mathbf X is a random sample of valid unit vectors (default 50 000), centred on its mean; ρc\rho_c is the share of variance of component cc. Reference: Jolliffe 2002. Output: PCA RGB raster.

(3) Similarity, mean prototype

s(x^)=x^⊤p^,p^=∑ix^i∥∑ix^i∥(3)s(\hat{\mathbf{x}}) = \hat{\mathbf{x}}^{\top}\hat{\mathbf{p}},\qquad\hat{\mathbf{p}} = \frac{\sum_i \hat{\mathbf{x}}_i}{\lVert \sum_i \hat{\mathbf{x}}_i \rVert} \tag{3}

p^\hat{\mathbf p} is the normalised sum of the positive examples' unit vectors. Reference: Brown et al. 2025. Output: score raster.

(4) Similarity, nearest example

s(x^)=max⁡ix^⊤x^i(4)s(\hat{\mathbf{x}}) = \max_i \hat{\mathbf{x}}^{\top}\hat{\mathbf{x}}_i \tag{4}

The similarity to the nearest single example; useful when the examples are varied. Reference: Brown et al. 2025. Output: score raster.

(5) Similarity, contrast

s(x^)=x^⊤p^+−x^⊤p^−(5)s(\hat{\mathbf{x}}) = \hat{\mathbf{x}}^{\top}\hat{\mathbf{p}}_{+} - \hat{\mathbf{x}}^{\top}\hat{\mathbf{p}}_{-} \tag{5}

p^+\hat{\mathbf p}_{+} and p^−\hat{\mathbf p}_{-} are the prototypes of the positive and negative examples. With the Otsu rule the threshold is taken on the pixels with s>0s > 0. Reference: Brown et al. 2025. Output: score raster.

(6) Otsu threshold

t∗=arg⁡max⁡t ω0(t) ω1(t) (μ0(t)−μ1(t))2(6)t^{*} = \arg\max_{t}\ \omega_0(t)\,\omega_1(t)\,\bigl(\mu_0(t) - \mu_1(t)\bigr)^{2} \tag{6}

ω0,ω1\omega_0, \omega_1 are the shares and μ0,μ1\mu_0, \mu_1 the means of the values below and above tt on a 256-bin histogram. When several bins share the maximum (an empty gap between two modes) the middle of that plateau is used. Reference: Otsu 1979. Output: threshold.

(7) MAD threshold

t=d~+k⋅1.4826⋅median⁡i∣di−d~∣(7)t = \tilde{d} + k \cdot 1.4826 \cdot \operatorname{median}_i \lvert d_i - \tilde{d} \rvert \tag{7}

d~\tilde d is the median; 1.4826 makes the MAD a consistent estimate of the standard deviation for normal data; default k=3k = 3. Reference: Leys et al. 2013. Output: threshold.

(8) Spherical k-means

k(i)=arg⁡max⁡kx^i⊤ck,ck=∑i:k(i)=kx^i∥∑i:k(i)=kx^i∥,J=∑i(1−x^i⊤ck(i))(8)k(i) = \arg\max_{k} \hat{\mathbf{x}}_i^{\top}\mathbf{c}_k,\qquad\mathbf{c}_k = \frac{\sum_{i: k(i) = k} \hat{\mathbf{x}}_i}{\lVert \sum_{i: k(i) = k} \hat{\mathbf{x}}_i \rVert},\qquad J = \sum_i \bigl(1 - \hat{\mathbf{x}}_i^{\top}\mathbf{c}_{k(i)}\bigr) \tag{8}

Assignment by the largest cosine, centroids as normalised member sums, k-means++ seeding, several starts (the lowest objective JJ wins). Reference: Dhillon & Modha 2001; Arthur & Vassilvitskii 2007. Output: cluster raster.

(9) Silhouette (cosine)

si=bi−aimax⁡(ai,bi),d(x^,y^)=1−x^⊤y^(9)s_i = \frac{b_i - a_i}{\max(a_i, b_i)},\qquad d(\hat{\mathbf{x}}, \hat{\mathbf{y}}) = 1 - \hat{\mathbf{x}}^{\top}\hat{\mathbf{y}} \tag{9}

aia_i is the mean cosine distance of pixel ii to its own cluster and bib_i to the nearest other cluster; computed on a 4 000-pixel subsample. Reference: Rousseeuw 1987. Output: k diagnostics.

(10) Calinski-Harabasz and Davies-Bouldin

CH=B/(k−1)W/(n−k),DB=1k∑cmax⁡c′≠cSc+Sc′∥μc−μc′∥,∥x^−y^∥2=2−2 x^⊤y^(10)\mathrm{CH} = \frac{B/(k-1)}{W/(n-k)},\qquad \mathrm{DB} = \frac{1}{k}\sum_{c}\max_{c' \ne c}\frac{S_c + S_{c'}}{\lVert \boldsymbol{\mu}_c - \boldsymbol{\mu}_{c'} \rVert},\qquad\lVert \hat{\mathbf{x}} - \hat{\mathbf{y}} \rVert^{2} = 2 - 2\,\hat{\mathbf{x}}^{\top}\hat{\mathbf{y}} \tag{10}

On chordal distance; BB and WW are the between- and within-cluster sums of squares, ScS_c the mean distance of members to their centroid μc\boldsymbol\mu_c. Higher CH and lower DB are better. Reference: Calinski & Harabasz 1974; Davies & Bouldin 1979. Output: k diagnostics.

(11) k nearest neighbours (cosine)

P(c∣x^)=∑i∈Nk(x^), yi=cwi∑i∈Nk(x^)wi,wi=max⁡(x^⊤x^i, 10−6)(11)P(c \mid \hat{\mathbf{x}}) = \frac{\sum_{i \in N_k(\hat{\mathbf{x}}),\, y_i = c} w_i}{\sum_{i \in N_k(\hat{\mathbf{x}})} w_i},\qquad w_i = \max(\hat{\mathbf{x}}^{\top}\hat{\mathbf{x}}_i,\ 10^{-6}) \tag{11}

Nk(x^)N_k(\hat{\mathbf x}) are the kk most similar training vectors. Reference: Cover & Hart 1967. Output: class, probability, margin.

(12) Nearest prototype

P(c∣x^)=exp⁡(x^⊤p^c/τ)∑lexp⁡(x^⊤p^l/τ),τ=0.05(12)P(c \mid \hat{\mathbf{x}}) = \frac{\exp(\hat{\mathbf{x}}^{\top}\hat{\mathbf{p}}_c/\tau)}{\sum_{l}\exp(\hat{\mathbf{x}}^{\top}\hat{\mathbf{p}}_l/\tau)},\qquad \tau = 0.05 \tag{12}

p^c\hat{\mathbf p}_c is the prototype of class cc; the temperature τ=0.05\tau = 0.05 turns cosine similarities into probabilities. Reference: Brown et al. 2025. Output: class, probability, margin.

(13) Multinomial logistic regression

min⁡W −1n∑ilog⁡softmax⁡(W⊤x~i)yi+λ2∥W∥2,L=12λmax⁡(X~⊤X~)/n+λ(13)\min_{\mathbf{W}}\ -\frac{1}{n}\sum_{i}\log \operatorname{softmax}(\mathbf{W}^{\top}\tilde{\mathbf{x}}_i)_{y_i} + \frac{\lambda}{2}\lVert \mathbf{W} \rVert^{2},\qquad L = \tfrac{1}{2}\lambda_{\max}(\tilde{\mathbf{X}}^{\top}\tilde{\mathbf{X}})/n + \lambda \tag{13}

x~\tilde{\mathbf x} is the unit vector with a constant 1 appended; λ=10−3\lambda = 10^{-3}; zero start, full-batch Nesterov gradient descent with step 1/L1/L, at most 500 iterations or until the change is below 10−610^{-6}. Reference: Bishop 2006. Output: class, probability, margin.

(14) Random forest (scikit-learn)

P(c∣x)=1B∑b=1BPb(c∣x)(14)P(c \mid \mathbf{x}) = \frac{1}{B}\sum_{b=1}^{B} P_b(c \mid \mathbf{x}) \tag{14}

A bridge to scikit-learn's random forest; the option is greyed out with an install hint when the package is missing. Reference: Breiman 2001. Output: class, probability, margin.

(15) Spatial block cross-validation

Δopt=OArandom−OAblocks(15)\Delta_{\mathrm{opt}} = \mathrm{OA}_{\mathrm{random}} - \mathrm{OA}_{\mathrm{blocks}} \tag{15}

Both cross-validations use the same classifier and samples; a large gap means random folds put neighbouring, near-identical pixels in training and test. Reference: Roberts et al. 2017; Valavi et al. 2019. Output: OA (blocks), OA (random), optimism gap.

(16) Quantity and allocation disagreement

Q=12∑g∣p+g−pg+∣,A=12∑g2min⁡(pg+−pgg, p+g−pgg)(16)Q = \tfrac{1}{2}\sum_{g}\lvert p_{+g} - p_{g+} \rvert,\qquad A = \tfrac{1}{2}\sum_{g} 2\min(p_{g+} - p_{gg},\ p_{+g} - p_{gg}) \tag{16}

pggp_{gg}, pg+p_{g+} and p+gp_{+g} are the diagonal, row and column proportions of the confusion matrix; Q+A=1−OAQ + A = 1 - \mathrm{OA}. Reference: Pontius & Millones 2011. Output: Q, A.

(17) Area of applicability

DI(x^)=min⁡i∥x^−x^i∥dˉ,AOA={DI≤min⁡(Q3+1.5 IQR, max⁡iDIi)}(17)\mathrm{DI}(\hat{\mathbf{x}}) = \frac{\min_i \lVert \hat{\mathbf{x}} - \hat{\mathbf{x}}_i \rVert}{\bar{d}},\qquad \mathrm{AOA} = \{\mathrm{DI} \le \min(Q_3 + 1.5\,\mathrm{IQR},\ \max_i \mathrm{DI}_i)\} \tag{17}

Unweighted, because all embedding axes share one scale; dˉ\bar d is the mean pairwise chordal distance of the training vectors, and each training sample's DIi\mathrm{DI}_i is computed against the samples of other folds; Q1Q_1, Q3Q_3 are type-7 quartiles of the training DIi\mathrm{DI}_i, as in CAST. Reference: Meyer & Pebesma 2021. Output: DI raster, AOA mask.

(18) Design-based accuracy and area

p^ij=Winijni⋅,A^j=A∑ip^ij,S(p^⋅j)=∑iWi2nijni⋅(1−nijni⋅)ni⋅−1,A^j±1.96 A S(p^⋅j)(18)\hat{p}_{ij} = W_i\frac{n_{ij}}{n_{i\cdot}},\quad \hat{A}_j = A\sum_i \hat{p}_{ij},\quad S(\hat{p}_{\cdot j}) = \sqrt{\sum_i W_i^{2}\frac{\frac{n_{ij}}{n_{i\cdot}}\bigl(1 - \frac{n_{ij}}{n_{i\cdot}}\bigr)}{n_{i\cdot} - 1}},\quad \hat{A}_j \pm 1.96\,A\,S(\hat{p}_{\cdot j}) \tag{18}

WiW_i is the mapped area share of class ii, nijn_{ij} the sample count with map class ii and reference class jj, AA the total mapped area. OA, UA and PA and their intervals follow Olofsson et al. (2014), equations 1 to 7 and 10 to 11. Reference: Olofsson et al. 2014. Output: area +/- CI, OA/UA/PA +/- CI.

(19) Change distance

dt=1−x^t⊤x^t+1∈[0,2],t∗=arg⁡max⁡tdt,ychange=yt∗+1(19)d_t = 1 - \hat{\mathbf{x}}_t^{\top}\hat{\mathbf{x}}_{t+1} \in [0, 2],\qquad t^{*} = \arg\max_t d_t,\qquad y_{\mathrm{change}} = y_{t^{*}+1} \tag{19}

dtd_t for each consecutive pair (yt,yt+1)(y_t, y_{t+1}); with more than two years the total ∑tdt\sum_t d_t and the endpoint distance d(y1,yT)d(y_1, y_T) are written too. Reference: Brown et al. 2025. Output: distance, change year, change regions.

(20) Polygon embedding

mˉ=1n∑i=1nx^i,Rˉ=∥mˉ∥∈[0,1],e^=mˉ/Rˉ(20)\bar{\mathbf{m}} = \frac{1}{n}\sum_{i=1}^{n}\hat{\mathbf{x}}_i,\qquad \bar{R} = \lVert \bar{\mathbf{m}} \rVert \in [0, 1],\qquad \hat{\mathbf{e}} = \bar{\mathbf{m}}/\bar{R} \tag{20}

nn is the number of valid pixels inside the polygon (Mardia & Jupp 2000, section 2.3.1). Reference: Mardia & Jupp 2000. Output: e00..e63, coherence.

(21) Sample size for a target precision

n=(∑iWiSiS(O^))2,Si=Ui (1−Ui)(21)n = \left(\frac{\sum_i W_i S_i}{S(\hat{O})}\right)^{2},\qquad S_i = \sqrt{U_i\,(1 - U_i)} \tag{21}

WiW_i is the mapped area share of class ii, UiU_i its expected user's accuracy (a guess made before labelling) and S(O^)S(\hat{O}) the standard error wanted for the overall accuracy; nn is rounded up. This is equation 13 of Olofsson et al. (2014); their worked example, W=(0.02,0.015,0.32,0.645)W = (0.02, 0.015, 0.32, 0.645), U=(0.7,0.6,0.9,0.95)U = (0.7, 0.6, 0.9, 0.95) and S(O^)=0.01S(\hat{O}) = 0.01, gives n=641n = 641. Reference: Olofsson et al. 2014. Output: total sample size n.

8. Outputs and nodata

OutputTypeNodataValues
Cube year rasterInt8, 64 bands−128-128quantised qc∈[−127,127]q_c \in [-127, 127]
PCA compositeByte, 3 bands01 to 255
Similarity score, probability, margin, DI, change distanceFloat32NaNs∈[−1,1]s \in [-1, 1] (contrast [−2,2][-2, 2]), P∈(0,1]P \in (0, 1], DI≥0\mathrm{DI} \ge 0, d∈[0,2]d \in [0, 2]
Change yearFloat32NaN (also outside the change mask)year yt∗+1y_{t^{*}+1}
Masks (match, change, AOA)Byte2550 / 1
Clusters, classesByte2550…k−10 \dots k-1, class index (see the class table)
Zone fieldsdouble / integerNULLe00..e63, Rˉ\bar R, valid share, sim_to_ref

Results are deterministic for a fixed seed. Rasters are processed 512 x 512 pixels at a time, never as a whole floating-point cube.

9. Limits and caveats

10. References

  1. Arthur, D., Vassilvitskii, S. (2007). k-means++: the advantages of careful seeding. Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms, 1027-1035.
  2. Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer, New York.
  3. Breiman, L. (2001). Random forests. Machine Learning 45(1), 5-32.
  4. Brown, C. F. et al. (2025). AlphaEarth Foundations: an embedding field model for accurate and efficient global mapping from sparse label data. arXiv:2507.22291.
  5. Calinski, T., Harabasz, J. (1974). A dendrite method for cluster analysis. Communications in Statistics 3(1), 1-27.
  6. Cover, T., Hart, P. (1967). Nearest neighbor pattern classification. IEEE Transactions on Information Theory 13(1), 21-27.
  7. Davies, D. L., Bouldin, D. W. (1979). A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence 1(2), 224-227.
  8. Dhillon, I. S., Modha, D. S. (2001). Concept decompositions for large sparse text data using clustering. Machine Learning 42(1-2), 143-175.
  9. GDAL/OGR contributors (2025). GDAL/OGR Geospatial Data Abstraction software Library. Open Source Geospatial Foundation. https://gdal.org
  10. Google (2026). AlphaEarth Foundations GCS data, README of gs://alphaearth_foundations/satellite_embedding/v1/annual (de-quantisation, overviews, index). Satellite Embedding V1 annual, CC BY 4.0.
  11. Jolliffe, I. T. (2002). Principal Component Analysis, 2nd ed. Springer, New York.
  12. Leys, C., Ley, C., Klein, O., Bernard, P., Licata, L. (2013). Detecting outliers: do not use standard deviation around the mean, use absolute deviation around the median. Journal of Experimental Social Psychology 49(4), 764-766.
  13. Mardia, K. V., Jupp, P. E. (2000). Directional Statistics. Wiley, Chichester.
  14. Meyer, H., Pebesma, E. (2021). Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods in Ecology and Evolution 12(9), 1620-1633.
  15. Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment 148, 42-57.
  16. Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics 9(1), 62-66.
  17. Pontius, R. G., Millones, M. (2011). Death to Kappa: birth of quantity disagreement and allocation disagreement for accuracy assessment. International Journal of Remote Sensing 32(15), 4407-4429.
  18. Roberts, D. R. et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8), 913-929.
  19. Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics 20, 53-65.
  20. Valavi, R., Elith, J., Lahoz-Monfort, J. J., Guillera-Arroita, G. (2019). blockCV: an R package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods in Ecology and Evolution 10(2), 225-232.

11. Licence, attribution and citation

Plugin: GPL-2.0-or-later. Data: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind.

Cite the plugin as: Eminoğlu, Y. (2026). 02EarthEmbeddings - Earth Embeddings Lab (version 0.1.1) [QGIS plugin]. https://gitlab.com/geophilo1/zero2earthembeddings. https://doi.org/10.5281/zenodo.24263654. The DOI 10.5281/zenodo.24263654 always resolves to the latest version; version 0.1.1 has DOI https://doi.org/10.5281/zenodo.24263655. The repository holds CITATION.cff. Cite the data as Brown et al. (2025) with the attribution sentence above.

Repository: gitlab.com/geophilo1/zero2earthembeddings. Issues: work items. Homepage: geophilo.com/zero2earthembeddings.