Class-Aligned and Class-Balancing Generative Domain Adaptation for Hyperspectral Image Classification. [IEEE Trans. Geosci. Remote. Sens. 62: 1-17 (2024)]

This is our official implementation of CCGDA!

by Jie Feng, Ziyu Zhou, Ronghua Shang, Jinjian Wu, Tianshu Zhang, Xiangrong Zhang, Licheng Jiao

Introduction

Abstract

The task of hyperspectral image (HSI) classification is fundamental and crucial in HSI processing. Currently, domain adaptive methods have become a research hotspot in HSI classification. However, most domain adaptive methods ignore the class alignment in different domains. Additionally, HSIs have the characteristics of category imbalance and complex spatial–spectral distribution, which restricts the adaptation performance in HSIs. To address these problems, a class-aligned and class-balancing generative domain adaptation (CCGDA) method is proposed for HSI classification. The architecture of CCGDA is designed by using the classifier, domain discriminator, sampler, and two weight-sharing generators. In the classifier, split-level capsule network (CapsNet) is constructed by extracting rich spatial information of shallow layer and spectral features of deep layer with equivariant characteristic. Then, the classifier provides the pseudo-label of samples in the target domain. To prevent the generators from mode collapse caused by category imbalance, the sampler is designed. It samples and resamples the samples of the target domain in an adaptive proportion according to the statistical calculation through confidence and distribution of pseudo-labels. Finally, a novel class-aligned domain adversarial loss is defined to jointly optimize the generators and discriminator. It incorporates the class shift adjusting and adaptive sampling for the samples of the target domain to better adapt the discriminant boundary of the classifier to the target domain. Experiments on benchmark HSI datasets verify the superiority of the proposed method for domain adaptive classification.

Fig.1 Flowchart of the proposed method. The framework consists of two generators, a classifier, a discriminator, and a data sampler. CCE refers to the class correlation evaluation, which works in the training process of the second generator.

Fig.2 Structure of classifier with split-level CapsNet.

Fig.3 Structure of domain discriminator and generator.

For further details, please check out our paper.

HSI域自适应

数据集描述

Houston数据集

类别	名称	Houston13	Houston18
1	Grass healthy	345	1353
2	Grass stressed	365	4888
3	Trees	365	2766
4	Water	285	22
5	Residential buildings	319	5347
6	Non-residential buildings	408	32459
7	Road	443	6365
total	total	2530	53200
shape	N * H * C	210 * 954 * 48	210 * 954 * 48

HyRANK数据集

类别	名称	Dioni	Loukia
1	Dense urban fabric	1262	288
2	Mineral extraction sites	204	67
3	Non irrigated land	614	542
4	Fruit trees	150	79
5	Olive Groves	1768	1401
6	Coniferous Forest	361	900
7	Dense Vegetation	5035	3793
8	Sparce Vegetation	6374	2803
9	Sparce Areas	1754	404
10	Rocks and Sand	492	487
11	Water	1612	1393
12	Coastal Water	398	451
total	total	20024	12208
shape	N * H * C	250 * 1376 * 176	249 * 945 * 176

ShanghaiHangzhou数据集

类别	名称	Shanghai	Hangzhou
1	Water	18043	123123
2	Land/Building	77450	161689
3	Plant	40207	83188
total	total	135700	368000
shape	N * H * C	1600 * 260 * 198	590 * 230 * 198

数据预处理

包括Z-Score归一化、图像裁剪、筛选类别和调整标签等

Houston数据集

python preprocess/preprocess.py configs/preprocess/houston.yaml ^
      --path E:/zts/dataset/houston_preprocessed

HyRANK数据集

python preprocess/preprocess.py configs/preprocess/hyrank.yaml ^
      --path E:/zts/dataset/hyrank_preprocessed

ShanghaiHangzhou数据集

python preprocess/preprocess.py configs/preprocess/shanghang.yaml ^
      --path E:/zts/dataset/shanghaihangzhou_preprocessed

支持的模型

用法

训练

运行 train/[model]/[dataset].bat文件
或者运行如下命令

python train/ddc/train.py configs/houston/ddc.yaml ^
       --path ./runs/houston/ddc-train ^
       --nodes 1 ^
       --gpus 1 ^
       --rank-node 0 ^
       --backend gloo ^
       --master-ip localhost ^
       --master-port 8886 ^
       --seed 30 ^
       --opt-level O2

测试

验证集等于测试集，无需再另行测试

试验记录

Dataset	Model	loss	loss-ratio	kernel	batch-size	OA-best	OA-worst
Houston	MCD	softmax+ce, discrepancy	-	l	64	0.633	0.608
Houston	DANN	softmax+ce	-	l	64	0.633	0.608
Houston	PixelDA	softmax+ce	-	l	64	0.633	0.608
HyRANK	MCD	softmax+ce, discrepancy	-	l	64	0.633	0.608
HyRANK	DANN	softmax+ce	-	l	64	0.633	0.608
HyRANK	PixelDA	softmax+ce	-	l	64	0.633	0.608

Cite

@article{jiefeng0109,
    title={Class-Aligned and Class-Balancing Generative Domain Adaptation for Hyperspectral Image Classification},
    author={Jie Feng, Ziyu Zhou, Ronghua Shang, Jinjian Wu, Tianshu Zhang, Xiangrong Zhang, Licheng Jiao},
    journal={TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING},
    volume={62},
    pages={1-17},
    year={2024},
    publisher={IEEE},
    doi={10.1109/TGRS.2024.3367765}
}

许可证

This project is released under the MIT(LICENSE) license.

breakmyhabits/CCGDA