Land cover classification using UNet

Problem statement

Model architecture

Model architecture is similar to the one presented in original paper U-Net: Convolutional Networks for Biomedical Image Segmentation, mainly in case of depth and number of features at each level. Although there is some differences:

• 3 input channels (RGB) - 1 channel orignal,
• input images resized to 224 x 224 - that results in 28 x 28 feature maps in bottleneck (572 x 572 original),
• 24 classes prediction - 2 classes in original,
• no information loss due to small padding - no padding originally

Things worth considering

• more channels:

• near infrared (NIR), 750–900 nm, used primarily for imaging vegetation

• thermal infrared, 10400-12500 nm, uses emitted instead of reflected radiation

• night images