GeoAI Model Training

Train reusable GeoAI models from high-resolution raster imagery and labelled vector features for object detection, instance segmentation, or semantic segmentation.

Remote Sensing CoreEnd-User DocumentationTheory + Formula + Parameters

1. What This Feature Does

Train reusable GeoAI models from high-resolution raster imagery and labelled vector features for object detection, instance segmentation, or semantic segmentation.

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

2. Recommended Workflow

Prepare representative labels and class field. ↓ Choose the learning task and architecture. ↓ Choose image tile/patch size, training epochs and device. ↓ Train the model. ↓ Review training/validation metrics. ↓ Use the saved model in GeoAI extraction or semantic classification.

3. Theory, Methods & Equations

YOLO26 — Object Detection

Learns bounding boxes and object classes from labelled objects. Modern YOLO training jointly optimizes localization, object confidence and classification.

IoU=|B_p∩B_g|/|B_p∪B_g|

Best used when: individual object boxes/centroids are sufficient.

YOLO26 — Instance Segmentation

Adds an object mask branch so separate objects receive separate pixel masks.

Dice = 2|P∩G|/(|P|+|G|)

Best used when: separate object footprints are required.

U-Net

Encoder-decoder semantic segmentation with skip connections that combine coarse context and fine boundaries.

CE=−Σy_clog p_c

Dice=2|P∩G|/(|P|+|G|)

Best used when: a strong baseline for pixel-level classes is needed.

U-Net++

Nested skip pathways reduce the semantic gap between encoder and decoder features and improve multi-scale feature fusion.

Best used when: more flexible multi-scale semantic segmentation is desired.

DeepLabV3+

Uses atrous/dilated convolution and spatial pyramid context, followed by a decoder for sharper boundaries.

Best used when: objects/classes occur across multiple spatial scales.

FPN — Feature Pyramid Network

Combines features from several network resolutions into a multi-scale pyramid.

Best used when: class structures vary substantially in size.

4. Input Data

Input	Type	Requirement	Description
Training Raster	raster	Required	High-resolution satellite, aerial, or UAV raster used to generate training image tiles.
Training Labels	vector	Required	Labeled vector layer. Polygon labels are recommended. Point labels can be converted to small boxes/masks; line labels are supported for semantic segmentation.

5. Parameters Available in the Application

Parameter	Default	Choices / Range	How to Use It
Training Algorithm	yolo26_detection	YOLO26 — Object Detection, YOLO26 — Instance Segmentation, U-Net — Semantic Segmentation, U-Net++ — Semantic Segmentation, DeepLabV3+ — Semantic Segmentation, FPN — Semantic Segmentation
Class Field	—	—	Attribute containing the target class name/code, for example class_name or class_id.
RGB Bands	[1, 2, 3]	—	Three 1-based bands in Red, Green, Blue order.
Training Tile Size	640	128 to 2048
Training Tile Overlap	64	0 to 512
Validation Fraction	0.2	0.05 to 0.5
Epochs	50	1 to 1000
Batch Size	8	1 to 256
Learning Rate	0.001	1e-07 to 1.0
Device	cpu	CPU, Auto, CUDA GPU, Apple MPS
Point Label Radius (pixels)	8	1 to 128	Used when training labels contain points.
Line Label Width (pixels)	3	1 to 128	Shown for: unet, unetplusplus, deeplabv3plus, fpn
Semantic Encoder	resnet34	ResNet-18, ResNet-34, ResNet-50, EfficientNet-B0	Shown for: unet, unetplusplus, deeplabv3plus, fpn
YOLO Model Size	n	Nano, Small, Medium, Large	Shown for: yolo26_detection, yolo26_instance_segmentation
Negative Tile Ratio	0.15	0.0 to 1.0	Fraction of empty/background tiles retained relative to labeled tiles.
Random State	42	—

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 labels should cover the variability expected at inference time: season, illumination, scale, background and object appearance.

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.

← GeoAI Semantic Classification Raster Change Detection →