Supervised Classification

Create a thematic class raster from labelled training samples. Spectral classifiers learn pixel-feature relationships; CNN/U-Net use spatial neighborhoods or semantic segmentation context.

Remote Sensing CoreEnd-User DocumentationTheory + Formula + Parameters

1. What This Feature Does

Create a thematic class raster from labelled training samples. Spectral classifiers learn pixel-feature relationships; CNN/U-Net use spatial neighborhoods or semantic segmentation context.

Cloud workflow: choose data → configure the scientific method → run → review the resulting layer/report. No programming is required.

2. Recommended Workflow

Prepare analysis-ready predictor raster. ↓ Create representative training samples/areas with a class field. ↓ Select predictor bands. ↓ Choose a classifier and tune its parameters. ↓ Run classification. ↓ Assess the result with independent reference samples using Classification Accuracy Assessment.

3. Theory, Methods & Equations

Maximum Likelihood

Assumes each class follows a multivariate Gaussian distribution and assigns the class with the highest discriminant probability.

g_i(x)=−½ln|Σ_i|−½(x−μ_i)^TΣ_i⁻¹(x−μ_i)+lnP(ω_i)

Best used when: classes are reasonably Gaussian and training covariance can be estimated reliably.

Nearest Neighbour

Assigns a class from nearby labelled samples in feature space.

d(x,x_i)=√Σ(x_j−x_ij)²

Best used when: local sample similarity is meaningful and training density is adequate.

Mahalanobis Distance

Measures distance from a class mean while accounting for covariance.

d_M²=(x−μ)^TΣ⁻¹(x−μ)

Best used when: band covariance differs from simple Euclidean geometry.

Random Forest

Combines many decision trees; each tree votes for a class.

ĉ(x)=mode{f₁(x),…,f_T(x)}

Gini = 1−Σp_k²

Best used when: nonlinear class boundaries and mixed predictor types are expected.

CNN

Learns spectral-spatial filters from image patches around labelled pixels.

Y=φ(W * X + b)* denotes convolution.

Best used when: local texture/shape around a pixel helps classification.

U-Net

Encoder–decoder semantic segmentation network with skip connections that restores pixel-level class boundaries.

CE = −Σ y_c log p_c

Dice = 2|P∩G|/(|P|+|G|)

Best used when: training polygons and spatially coherent semantic classes are available.

KAN

Nonlinear classifier using learnable univariate functions on connections.

f(x) ≈ Σ_q Φ_q(Σ_p φ_qp(x_p))

Best used when: advanced nonlinear spectral classification is being explored.

Transformer

Uses self-attention to learn relationships among predictor bands/features.

Attention(Q,K,V)=softmax(QK^T/√d_k)V

Best used when: cross-band dependencies are important and enough training data exist.

4. Input Data

Input	Type	Requirement	Description
Predictor Raster	raster	Required	Multi-band raster used as classification predictors.
Training Samples	vector	Required for selected method	Point or polygon training features containing the selected class field.
Training Areas	vector	Required for selected method	Polygon training areas are recommended for CNN and U-Net. Point labels are accepted but provide sparse supervision.
Validation Samples (Optional)	vector	Conditional / Optional	Independent Point/MultiPoint or Polygon/MultiPolygon reference samples. When supplied, the same run computes confusion matrix, Overall Accuracy, Kappa, Producer/User Accuracy, Precision, Recall and F1.

5. Parameters Available in the Application

Parameter	Default	Choices / Range	How to Use It
Classification Algorithm	—	Maximum Likelihood, Nearest Neighbour, Mahalanobis Distance, Random Forest, CNN, U-Net, KAN, Transformer
Class Field	—	—	Attribute containing the training class label. Shown for: maximum_likelihood, nearest_neighbour, mahalanobis, random_forest, kan, transformer, cnn, unet
Validation Class Field	—	—	Reference class field in Validation Samples. Leave empty to use the same Class Field.
Predictor Bands	—	—	Raster band numbers, for example [1,2,3,4]. Leave empty to use all bands. Shown for: maximum_likelihood, nearest_neighbour, mahalanobis, random_forest, kan, transformer, cnn, unet
All Touched Pixels	Off	—	When enabled, every pixel touched by a training geometry is labelled. Shown for: maximum_likelihood, nearest_neighbour, mahalanobis, random_forest, kan, transformer, cnn, unet
Random State	42	—	Shown for: nearest_neighbour, random_forest, cnn, unet, kan, transformer
Maximum Training Pixels	200000	100 to None	Balanced maximum number of labelled pixels used by spectral classifiers. Shown for: maximum_likelihood, nearest_neighbour, mahalanobis, random_forest, kan, transformer
Prediction Batch Size	200000	1000 to 2000000	Shown for: maximum_likelihood, nearest_neighbour, mahalanobis, random_forest, kan, transformer
Covariance Regularization	1e-06	1e-12 to —	Shown for: maximum_likelihood
Class Priors	empirical	empirical, equal	Shown for: maximum_likelihood
Number of Neighbours	5	1 to 100	Shown for: nearest_neighbour
Neighbour Weights	distance	uniform, distance	Shown for: nearest_neighbour
Distance Metric	euclidean	euclidean, manhattan	Shown for: nearest_neighbour
Standardize Predictors	On	—	Shown for: nearest_neighbour
Covariance Regularization	1e-06	1e-12 to —	Shown for: mahalanobis
Number of Trees	300	10 to 3000	Shown for: random_forest
Maximum Tree Depth (0 = unlimited)	0	0 to 500	Shown for: random_forest
Minimum Samples Split	2	2 to 100	Shown for: random_forest
Maximum Features	sqrt	sqrt, log2, all	Shown for: random_forest
CNN Patch Size	15	5 to 63	Shown for: cnn
CNN Channels	[32, 64]	—	Shown for: cnn
CNN Epochs	80	1 to 2000	Shown for: cnn
CNN Learning Rate	0.001	1e-07 to —	Shown for: cnn
CNN Batch Size	16	1 to 512	Shown for: cnn
CNN Dropout	0.1	0.0 to 0.9	Shown for: cnn
Maximum CNN Training Patches	5000	10 to None	Shown for: cnn
U-Net Patch Size	128	32 to 1024	Shown for: unet
U-Net Base Channels	24	4 to 256	Shown for: unet
U-Net Depth	3	2 to 5	Shown for: unet
U-Net Epochs	60	1 to 2000	Shown for: unet
U-Net Learning Rate	0.001	1e-07 to —	Shown for: unet
U-Net Batch Size	4	1 to 128	Shown for: unet
U-Net Dropout	0.1	0.0 to 0.9	Shown for: unet
Maximum U-Net Training Patches	2500	10 to None	Shown for: unet
Compute Device	auto	auto, cpu, cuda	Shown for: cnn, unet, kan, transformer
KAN Hidden Layers	[32, 16]	—	Shown for: kan
KAN Grid Size	5	2 to 32	Shown for: kan
KAN Spline Order	3	1 to 5	Shown for: kan
KAN Epochs	150	1 to 2000	Shown for: kan
KAN Learning Rate	0.01	1e-07 to —	Shown for: kan
KAN Batch Size	512	2 to 65536	Shown for: kan
KAN Weight Decay	0.0	0.0 to —	Shown for: kan
Transformer Embedding Size	64	4 to 1024	Shown for: transformer
Transformer Attention Heads	4	1 to 32	Shown for: transformer
Transformer Encoder Layers	2	1 to 12	Shown for: transformer
Transformer Feed-forward Size	128	4 to 8192	Shown for: transformer
Transformer Dropout	0.1	0.0 to 0.9	Shown for: transformer
Transformer Epochs	150	1 to 2000	Shown for: transformer
Transformer Learning Rate	0.001	1e-07 to —	Shown for: transformer
Transformer Batch Size	512	2 to 65536	Shown for: transformer
Transformer Weight Decay	1e-05	0.0 to —	Shown for: transformer

6. Output & Interpretation

The result should be interpreted according to the selected method and the physical meaning of the input data. Preserve source units, coordinate reference information, NoData meaning, acquisition date, and preprocessing level when comparing results.

Scientific interpretation: an algorithm can produce a numerically valid result even when the input data are unsuitable. Always check masks, units, sensor characteristics, spatial resolution and reference data.

Important: Training data quality is more important than simply increasing model complexity. Separate training and independent validation samples whenever possible.

7. Best Practices

Use analysis-ready inputs and remove invalid/cloud/noise artifacts that are not part of the target phenomenon.
Choose parameters from the physical scale of the data, not only from visual appearance.
Keep categorical and continuous rasters conceptually separate when selecting interpolation/resampling methods.
Compare the result with the source image and independent reference information.
Document the settings used when results will be compared across dates, sensors or study areas.

← Raster Modeller Unsupervised Classification →