Export our model to ONNX as Deepness Segmentor

[rojberr]

User Story:

As a machine learning engineer,
I want to export our trained model in Deepness Segmentor format compatible with QGIS,
So that users can perform semantic segmentation on selected map polygons with optimized inference.

---

Acceptance Criteria:

• Model converted to ONNX format with metadata requirements
• Deepness-compatible segmentation masks validated (correct class mapping)
• Output produces single-class masks per input patch (entire patch = one class)
• Documentation added for export process and requirements

---

Definition of Done:

• All acceptance criteria met
• Code reviewed and approved
• Unit tests for export script
• Documentation updated in /docs/deployment.md
• Sample exported model added to model registry

---

Description:

Labels for Polish trees in QGIS: https://essd.copernicus.org/articles/16/2877/2024/#section4

Export model to Deepness Segmentor format (reference: Fire Risk Model example) with these specifications:

• Input: 4-channel patches (Near-Infrared, Red, Green, Blue) from QGIS
• Output: Single-class segmentation mask per patch (all pixels = same class)
• Focus: Forest areas only (no "other" classes in initial version)
• Normalization: Handled automatically by Deepness
• Polygon support: Inference on user-selected map regions
• Channel selection: QGIS users can select input channels

Metadata Requirements (source):

metadata = {
    'model_type': 'segmentor',
    'task_type': 'segmentation',
    'classes': '["forest"]',  # Single-class initially
    'input_channels': 4,      # NIR-R-G-B
    'normalization_mean': "[0.485, 0.456, 0.406, 0.440]",
    'normalization_std': "[0.229, 0.224, 0.225, 0.242]"
}

Resources:

Konwertowanie modelu do ONNX: https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_example_onnx_model.html,
Deepness model zoo: https://qgis-plugin-deepness.readthedocs.io/en/latest/main/main_model_zoo.html#deepness-model-zoo
(Szczególnie Fire Risk): https://qgis-plugin-deepness.readthedocs.io/en/latest/example/example_segmentation_landcover.html

Metadata: https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_add_metadata_to_model.html
Deepness does normalization by default.

[rojberr]

Our model would segmentation,

return one value for all output mask

take tile 64x64, we do classification for the tile (return one class out of list of classes)

jako output robie powtorzenie tej wartosci jako maske wyjsciowa

to by trzeba bylo zaimplementowac w sieci, bo Deepness tego nierobi, wiec trzeba moze funkcje repeat w dpowiednich osiach,

ta klasa dla wszystkich pixeli z danego obszaru

wiec dla 64x64, na wyjsiu mamy maske 64x64 i przypisujemy wartosc klasyfikacji

zrob na patchach z ta sama rozdzielczoscia przestrzenna co trenowalismy, 256x250 0,2m

UWAGA bo ktos mial problem model zwracal maske wyjsiowa zmiejszona o 60 pixeli, pokrycie bylo wieksze niz 10% miedzy tile i zhardkodowal tileing, zawsze krycie bylo wieksze niz 10% miedzy kolejnymi tileami.

[rojberr]

UWAGA
Resolution cm/px rozdzielczosci przestrzennej jest wyjatkowo wazne bo okresla wielkosc obiektow.

Modele sa na nie bardzo czule. Jesli trenujemy 10cm/px to powinnismy ten model tez tak stosowac.

WIec uwaga na augumentacje skalowaniem.

[rojberr]

Sources:

• https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_example_onnx_model.html
• https://docs.pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html