Skip to content

DominikGithub/cloud_image_segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

37 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LWIR cloud image segmentation

Segment clouds in satellite infrared single channel images.

Download report (.pdf)

Approach

Supervised training of a model to learn the segmentation mask as a target labels (Y) from the satellite images (X).

Data set

Pairs of IR (single channel) images and binary masks for supervised learning.

  • Image/mask size: 1024x1024x1
  • Approx. 1000 samples

Data preprocessing

  • TIFF image format
  • Adaption to multi channel base model, by replicating gray scale input image to 3 channels
  • Augmentations (rotation, flip, noise, cropping)

IR color mapping

IR error sample intensity historgram

Models

Transfer learning for image semantic segmentation tasks

(CNN) VGG19

Location: /tensorflow_vgg/

(static preprocessed TFRecord dataset, no augmentation)

VGG19 setup

VGG19 results

  • Flat (one step) upsampling decoder
  • Deeper decoder designs increase training challenge significantly

(ViT) Segformer

Location: /pytorch_segformer/

segformer setup

segformer results

  • Two step upsampling decoder

Training

f-scores