for Open-set Land Cover Classification of Satellite Imagery Razieh - - PowerPoint PPT Presentation

▶

Oct 20, 2023 171 likes •378 views

Representative-Discriminative Learning for Open-set Land Cover Classification of Satellite Imagery Razieh Kaviani Baghbaderani 1 , Ying Qu 1 , Hairong Qi 1 , Craig Stutts 2 1 University of Tennessee 2 Applied Research Associates Land Cover

SLIDE 1

Representative-Discriminative Learning for Open-set Land Cover Classification

f Satellite Imagery

Razieh Kaviani Baghbaderani1, Ying Qu1, Hairong Qi1, Craig Stutts2

1 University of Tennessee 2 Applied Research Associates

SLIDE 2

Land Cover Classification

http://lesun.weebly.com/hyperspectral-data-set.html [Christophe et al. 2018]

SLIDE 3

Challenge

Vast area of versatile cover materials
The training data is non-representative

https://en.wikipedia.org/wiki/Land_cover

SLIDE 4

Challenge

Vast area of versatile cover materials
The training data is non-representative

https://en.wikipedia.org/wiki/Land_cover

SLIDE 5

❑Existing methods

“Closed-set assumption”

❑This paper

“Open-set assumption”

SLIDE 6

Image Space (𝑌) Embedding Space (𝑎𝐺) Abundance Space (𝑇)

Representative-Discriminative Learning

SLIDE 7

Image Space (𝑌) Embedding Space (𝑎𝐺)

Representative-Discriminative Learning

Closed-set Embedding Learning Representative-discriminative Feature Learning

Abundance Space (𝑇) Image Space (𝑌) Embedding Space (𝑎𝐺) Abundance Space (𝑇)

SLIDE 8

Closed-set Embedding Learning

Classification loss: Sparsity loss: Image space Embedding space

SLIDE 9

Representative-Discriminative Feature Learning

Instead of an Auto-encoder structure

→ A multi-task representative-discriminative feature learning structure

SLIDE 10

Representative-Discriminative Feature Learning

❑Inspired by Spectral Unmixing concept,

Physical constraints:
Non-negativity
Sum-to-one

Bases Abundance

𝑎𝐺

▪ Known: ▪ Unknown:

𝑎𝐺 = 𝑡𝐶 𝑎𝐺 = 𝑡𝐶 𝑡 ∉ 𝑡

= 𝑡𝐶

SLIDE 11

Representative-Discriminative Feature Learning

𝑎𝐺 = 𝑡𝐶 ❑B → Decoder D

s ∈ Dirichlet distribution

Using stick-breaking structure:

𝑨𝐺 ෞ 𝑨𝐺

❑s → Encoder E

SLIDE 12

Representative-Discriminative Feature Learning

❑To increase the discriminative capacity,

SLIDE 13

Hyperspectral data

Pavia University (PU) Pavia Center (PC) Indian Pines (IN)

SLIDE 14

Experiments on Hyperspectral data

Pavia University (PU) Pavia Center (PC)

SLIDE 15

Experiments on Hyperspectral data

Pavia University (PU) Pavia Center (PC)

SLIDE 16

Experiments on Hyperspectral data

[26] Bendale & Boult, 2016. [30] Oza & Patel, 2019.

SLIDE 17

Experiments on RGB Images

… 𝑌

Classifier (F) Encoder (E)

Reconstruction loss

Dirichlet-Net

𝑇

ෞ 𝑨𝐺

𝑍

Decoder (D) Classifier (C)

𝛾 𝑣 𝑊 𝑨𝐺 𝑍

Substituted with a CNN-based structure

SLIDE 18

Experiments on RGB Images

[26] Bendale & Boult, 2016. [33] Neal et al., 2018. [34] Oza & Patel, 2019. [31] Perera et al., 2020.

SLIDE 19

Ablation Study

SLIDE 20

Contributions

For more details:

https://github.com/raziehkaviani/rdosr

✓Exploited both representative and discriminative power of feature learning. ✓Being the first to address Open-set land cover classification in satellite imagery. ✓Showed the generalization capacity to RGB images.