Deepness Model ZOO

The Model ZOO is a collection of pre-trained, deep learning models in the ONNX format. It allows for an easy-to-use start with the plugin.

NOTE: the provided models are not universal tools and will perform well only on similar data as in the training datasets. If you notice the model is not perfroming well on your data, consider re-training (or fine-tuning) it on your data.

If you do not have machine learning expertise, feel free to contact the plugin authors for help or advice.

Segmentation models

Model

Input size

CM/PX

Description

Example image

Corn Field Damage Segmentation

512

3

PUT Vision model for Corn Field Damage Segmentation created on own dataset labeled by experts. We used the classical UNet++ model. It generates 3 outputs: healthy crop, damaged crop, and out-of-field area.

Image

Land Cover Segmentation

512

40

The model is trained on the LandCover.ai dataset. It provides satellite images with 25 cm/px and 50 cm/px resolution. Annotation masks for the following classes are provided for the images: building (1), woodland (2), water(3), road(4). We use DeepLabV3+ model with tu-semnasnet_100 backend and FocalDice as a loss function. NOTE: the dataset covers only the area of Poland, therefore the performance may be inferior in other parts of the world.

Image

Buildings Segmentation

256

40

Trained on the RampDataset dataset. Annotation masks for buildings and background. Xunet network. Val F1-score 81.0

Image

Land Cover Segmentation Sentinel-2

64

1000

Trained on the Eurosat dataset. Uses 13 spectral bands from Sentinel-2, with 10 classes. Model ConvNeXt.

Image

Agriculture segmentation RGB+NIR

256

30

Trained on the Agriculture Vision 2021 dataset. 4 channels input (RGB + NIR). 9 output classes within agricultural field (weed_cluster, waterway, …). Uses X-UNet.

Image

Fire risk assesment

384

100

Trained on the FireRisk dataset (RGB data). Classifies risk of fires (ver_high, high, low, …). Uses ConvNeXt XXL. Val F1-score 65.5.

Image

Roads Segmentation

512

21

The model segments the Google Earth satellite images into ‘road’ and ‘not-road’ classes. Model works best on wide car roads, crossroads and roundabouts.

Image

Solar PV Segmentation

512

3

Model trained by M Kleebauer et al. in “Multi-resolution segmentation of solar photovoltaic systems using deep learning on a diverse range of image data, spanning UAV, aerial, and satellite imagery at both native and aggregated resolutions of 0.1 m, 0.2 m, 0.3 m, 0.8 m, 1.6 m, and 3.2 m.

Image

Regression models

Recognition models

Model

Input size

CM/PX

Description

Example image

NAIP Place recognition

224

100

ConvNeXt nano trained using SimSiam onn NAIP imagery. Rank1-accuracy 75.0.

Image

Object detection models

Model

Input size

CM/PX

Description

Example image

Airbus Planes Detection

256

70

YOLOv7 tiny model for object detection on satellite images. Based on the Airbus Aircraft Detection dataset.

Image

Airbus Oil Storage Detection

512

150

YOLOv5-m model for object detection on satellite images. Based on the Airbus Oil Storage Detection dataset.

Image

Aerial Cars Detection

640

10

YOLOv7-m model for cars detection on aerial images. Based on the ITCVD.

Image

UAVVaste Instance Segmentation

640

0.5

YOLOv8-L Instance Segmentation model for litter detection on high-quality UAV images. Based on the UAVVaste dataset.

Image

Super Resolution Models

Model

Input size

CM/PX

Scale Factor

Description

Example image

Residual Dense Network (RDN X2)

64

Trained on 10 cm/px images set it same as input data

X2

Model originally trained by H Zhang et. al. in “A Comparative Study on CNN-Based Single-Image Super-Resolution Techniques for Satellite Images“ converted to onnx format

Image from Massachusetts Roads Dataset Dataset in kaggle

Residual Dense Network (RDN X4)

64

Trained on 10 cm/px images set it same as input data

X4

Model originally trained by H Zhang et. al. in “A Comparative Study on CNN-Based Single-Image Super-Resolution Techniques for Satellite Images“ converted to onnx format

Image from Massachusetts Roads Dataset Dataset in kaggle

Contributing

PRs with models are welcome!