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

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

Challenge • Vast area of versatile cover materials • The training data is non-representative https://en.wikipedia.org/wiki/Land_cover

Challenge • Vast area of versatile cover materials • The training data is non-representative https://en.wikipedia.org/wiki/Land_cover

❑ Existing methods “Closed - set assumption” ❑ This paper “ Open-set assumption ”

Representative-Discriminative Learning Image Space Embedding Space Abundance Space (𝑌) (𝑎 𝐺 ) (𝑇)

Representative-Discriminative Learning Image Space Image Space Embedding Space Embedding Space Abundance Space Abundance Space (𝑌) (𝑌) (𝑎 𝐺 ) (𝑎 𝐺 ) (𝑇) (𝑇) Closed-set Embedding Learning Representative-discriminative Feature Learning

Closed-set Embedding Learning Image space Embedding space Classification loss: Sparsity loss:

Representative-Discriminative Feature Learning • Instead of an Auto-encoder structure → A multi-task representative-discriminative feature learning structure

Representative-Discriminative Feature Learning ❑ Inspired by Spectral Unmixing concept, 𝑎 𝐺 = 𝑡 𝐶 ▪ Known: 𝑎 𝐺 = 𝑡𝐶 Abundance Bases ▪ Unknown: 𝑎 𝐺 = 𝑡𝐶 o Physical constraints: 𝑡 ∉ 𝑡 • Non-negativity • Sum-to-one

Representative-Discriminative Feature Learning 𝑎 𝐺 = 𝑡𝐶 ❑ B → Decoder D 𝑨 𝐺 ෞ ❑ s → Encoder E 𝑨 𝐺 • s ∈ Dirichlet distribution Using stick-breaking structure:

Representative-Discriminative Feature Learning ❑ To increase the discriminative capacity,

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

Experiments on Hyperspectral data Pavia University (PU) Pavia Center (PC)

Experiments on Hyperspectral data Pavia University (PU) Pavia Center (PC)

Experiments on Hyperspectral data [26] Bendale & Boult, 2016. [30] Oza & Patel, 2019.

Experiments on RGB Images Classifier (F) Reconstruction loss Encoder (E) Decoder (D) Dirichlet-Net 𝑌 𝛾 𝑨 𝐺 ෞ 𝑨 𝐺 𝑇 … 𝑊 𝑣 Classifier (C) 𝑍 𝑍 Substituted with a CNN-based structure

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

Ablation Study

Contributions ✓ 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. For more details: https://github.com/raziehkaviani/rdosr