-
Notifications
You must be signed in to change notification settings - Fork 0
Models
In this repository, we have included 3 different state-of-art models for semantic segmentation and their corresponding pre-trained weight. We believe their usage can be extended to any usecases involving Sentinel-2 as input data, being used as a first weight initialisation step.
In all our configurations, we have used an input image of 256x256 pixels, with 4 channels. With the features provided by Hugin, we tile the initial images into smaller patches and use the 10m resolution channels as inputs:
- B2 - blue
- B3 - green
- B4 - red
- B8 - nir
We have trained three networks, with two different agricultural classification use-cases. First, we have used a more coarse Ground Truth (GT) label, from the publicly available Corine Land Cover dataset, mapping 8 classes. Second, we have used a dataset with more refined labels, from the publicly available data of Land Parcel Identification System, mapping 25 classes.
| Model | Total params | Trainable params |
|---|---|---|
| U-Net | 7,772,392 | 3,053,992 |
| W-Net | 32,474,216 | 32,468,072 |
| HSN | 8,241,192 | 8,240,424 |
U-Net (Paper) is a topology designed for medical image segmentation, but has been widely used in Remote Sensing applications. As opposed to the originally proposed topology, we have also inserted Batch Normalization layer after each Convolutional Layer.
U-Net weights trained with the Corine GT may be downloaded (here)
W-Net (Paper) is a network designed for medical image segmentation, composed on two U-Net side by side, with additional skip connections transferring information from one U-Net to the other. W-Net also brings the alternation between two activations functions ELU and ReLU, to avoid the saturation.
W-Net weights trained with the Corine GT may be downloaded (here)
HSN (Paper) was designed for remote sensing data, on the ISPRS Vaihingen and Potsdam datasets. It is also based on an encoder-decoder structure, following the same principle of transferring features from the encoder to the decoder as in U-Net. However, it combines residual and inception modules to improve the performance of the model.
HSN weights trained with the Corine GT may be downloaded(here)