1 of 12

Using Semantic segmentation on hyperspectral images to predict the landcover

Group 7 黃世宇 雷聿文 

2 of 12

outline

  • Brief intro.
  • Dataset
  • Model intro.
  • 損失函數與評估標準
  • 實驗結果
  • 結論與未來可改進的部分

3 of 12

Brief intro.

  • Model : U-Net backbone, with resnet18 as encoder, and bilinear interpolation as decoder. 4 skip connection is used. For encoder part, we add SE module to see if it will perform better or not.

ref. [Deep Residual Learning for Image Recognition Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun]

此圖第三個column是resnet18的架構,我們的模型的卷積核大小以及數量就是遵循這個resnet18的架構(conv2_x~conv5_x)。(當然output size不同)

4 of 12

Dataset info.

  • Training set : Validation set : testing set = 6:2:2

  • 裁切圖像塊數之shape : 96 x 96 x 200 (height x width x channel)

  • mask pixel value range : 0,1,2,...,29 (total 30 classes)

5 of 12

  • ref.HSI-TransUNet: A transformer based semantic segmentation model for crop mapping from UAV hyperspectral imagery Bowen Niu , Quanlong Feng , Boan Chen , Cong Ou , Yiming Liu , Jianyu Yang ​

6 of 12

Model intro.

  • spatial,channel,layer attention modules的抉擇
  • 在許多channel attention module中,選SE module的原因
  • 選resnet18,不是34、50、101的原因
  • 解碼端選bilinear interpolation,不是deconvolution的原因

7 of 12

上圖是U-Net原始論文的模型架構圖,我們跟他的模型區別只在於我們模型

  • 做捲積時有做zero padding,使得捲積前後的特徵圖長寬不變

  • 編碼端是用residual block + SE module

  • 每一次的捲積的#kernel是遵循resnet18的#kernel

  • 反捲積部分改成雙線性內插

  • 輸入的影像大小CxHxW = 200x96x96

  • 輸出的特徵圖大小CxHxW = 30x96x96,30是類別數

8 of 12

損失函數與評估標準

  • 損失函數:
  • cross entropy,focal loss and weighted

  • 評估標準(evaluation metric):
  • MIOU
  • Overall accuracy

9 of 12

實驗結果

ref.HSI-TransUNet: A transformer based semantic segmentation model for crop mapping from UAV hyperspectral imagery Bowen Niu , Quanlong Feng , Boan Chen , Cong Ou , Yiming Liu , Jianyu Yang 

10 of 12

11 of 12

結論與未來可改進的部分

  • 解決樣本數少的類別的像素分類很差的問題 
  • 可以使用一些更複雜的模組,或是更深更廣的網路架構 �

12 of 12

  • END