Cross-Feature Granularity Fusion Network for Land Cover Classification of Hyperspectral Remote Sensing Images and LiDAR

Dan Fan; Zhengwei Yang; Xia Li; Chao Feng; Chuangjiang Rao

doi:10.3788/LOP242189

Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1028002(2025)

Cross-Feature Granularity Fusion Network for Land Cover Classification of Hyperspectral Remote Sensing Images and LiDAR

Dan Fan¹, Zhengwei Yang^1、*, Xia Li², Chao Feng¹, and Chuangjiang Rao²

Author Affiliations

¹Yunnan Water Resources and Hydropower Survey and Design Institute Co., Ltd., Kunming 650032, Yunnan , China

²Yunnan Institute of Water & Hydropower Engineering Investigation, Design and Research, Kunming 650032, Yunnan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Network framework

Download full size

Fig. 2. Structure of MCFF

Download full size

Fig. 3. Structure of CGFI

Download full size

Fig. 4. MUUFL dataset. (a) Hyperspectral image; (b) LiDAR image; (c) ground-truth map

Download full size

Fig. 5. Houston 2018 dataset. (a) Hyperspectral image; (b) LiDAR image; (c) ground-truth map

Download full size

Fig. 6. Trento dataset. (a) Hyperspectral image; (b) LiDAR image; (c) ground-truth map

Download full size

Fig. 7. Classification results of different models on the MUUFL dataset. (a) Context CNN; (b) CRNN; (c) ViT; (d) SpectralFormer; (e) Two-Branch CNN; (f) Coupled CNN; (g) FusAtNet; (h) CFCGNet

Download full size

Fig. 8. Classification results of different models on the Houston 2018 dataset. (a) Context CNN; (b) CRNN; (c) ViT; (d) SpectralFormer; (e) Two-Branch CNN; (f) Coupled CNN; (g) FusAtNet; (h) CFCGNet

Download full size

Fig. 9. Classification results of different models on the Trento dataset. (a) Context CNN; (b) CRNN; (c) ViT; (d) SpectralFormer; (e) Two-Branch CNN; (f) Coupled CNN; (g) FusAtNet; (h) CFCGNet

Download full size

Table 1. Classification evaluation metrics of different models on the MUUFL dataset

View table

Table 1. Classification evaluation metrics of different models on the MUUFL dataset

Class	Two-Branch CNN	Coupled CNN	Context CNN	CRNN	FusAtNet	ViT	SpectralFormer	CFCGNet
trees	92.78	98.50	91.43	87.25	93.62	85.69	90.26	94.62
mostly grass	59.74	78.66	63.76	85.36	84.64	81.56	75.46	88.54
mixed ground surface	94.15	90.29	81.62	90.21	87.14	73.87	78.84	92.84
dirt and sand	93.12	90.05	93.24	94.67	92.78	86.03	86.87	96.08
road	92.46	96.83	89.06	84.51	86.13	86.31	89.13	92.33
water	98.02	75.09	92.69	81.48	83.54	93.07	98.45	99.04
building shadow	95.28	73.78	84.31	93.59	89.91	83.82	88.19	92.21
building	94.42	96.81	81.72	90.22	95.28	83.74	77.82	92.47
sidewalk	86.53	64.51	81.53	87.23	85.33	71.13	74.33	84.13
yellow curb	99.37	19.75	99.03	93.28	91.80	100.00	94.08	100.00
cloth panels	96.81	62.36	98.67	99.36	98.26	99.63	97.60	99.23
OA	90.52	92.62	86.13	88.03	90.87	83.44	85.54	93.13
AA	91.15	76.97	87.01	89.74	89.86	85.90	86.46	93.71
Kappa	84.72	88.39	82.15	83.59	86.08	76.24	78.32	89.34

Table 2. Classification evaluation metrics of different models on the Houston 2018 dataset

View table

Table 2. Classification evaluation metrics of different models on the Houston 2018 dataset

Class	Two-Branch CNN	Coupled CNN	Context CNN	CRNN	FusAtNet	ViT	SpectralFormer	CFCGNet
healthy grass	93.41	91.05	94.72	79.02	77.98	75.54	85.64	76.04
stressed grass	90.96	91.12	90.20	93.61	93.95	92.75	86.12	96.75
artificial turf	99.15	99.68	99.23	98.49	99.54	76.02	99.31	98.86
evergreen trees	94.53	95.41	93.38	96.54	93.85	97.13	96.89	97.37
deciduous trees	95.47	95.97	85.96	86.59	88.21	81.76	85.83	94.02
bare earth	99.54	99.08	99.23	99.63	100.00	97.88	89.57	99.91
water	100.00	99.10	98.35	67.05	75.85	100.00	92.28	98.95
residential buildings	83.86	88.44	85.81	93.96	94.81	88.79	88.73	98.31
non-residential buildings	92.77	93.07	95.64	98.78	98.55	98.63	97.45	98.21
roads	72.60	85.21	61.61	78.74	80.79	83.05	70.87	86.23
sidewalks	63.39	58.48	61.40	73.04	77.38	74.84	73.20	80.95
crosswalks	67.92	65.84	70.08	31.85	21.68	27.90	18.28	59.71
major thoroughfares	70.19	65.36	69.39	88.30	86.84	83.65	82.99	97.50
highways	91.85	93.26	85.11	90.02	89.31	84.74	82.27	89.27
railways	96.47	99.58	96.48	100.00	99.15	99.31	99.03	98.60
paved parking lots	93.17	89.45	88.30	95.96	94.99	93.46	88.54	97.50
unpaved parking lots	98.83	99.91	99.08	55.63	46.52	23.60	64.09	72.27
cars	94.52	95.53	93.52	81.95	85.29	86.83	85.82	93.97
trains	96.58	97.76	96.41	95.15	95.17	95.74	95.16	98.83
stadium seats	99.80	98.08	97.36	98.97	99.63	98.61	97.80	99.48
OA	86.34	87.18	86.14	92.29	92.53	91.60	89.34	95.73
AA	89.75	90.07	88.06	85.16	84.97	83.01	84.00	91.64
Kappa	87.31	89.61	86.89	90.42	89.90	87.95	85.57	92.68

Table 3. Classification evaluation metrics of different models on the Trento dataset

View table

Table 3. Classification evaluation metrics of different models on the Trento dataset

Class	Two-Branch CNN	Coupled CNN	Context CNN	CRNN	FusAtNet	ViT	SpectralFormer	CFCGNet
apple trees	99.21	99.05	96.45	94.30	99.67	93.91	94.86	99.10
buildings	97.12	97.77	94.63	93.34	96.83	94.81	93.44	96.77
ground	83.54	82.84	80.29	79.53	79.01	82.43	80.03	88.27
wood	98.99	99.27	96.93	95.73	99.74	97.35	96.05	99.80
vineyard	93.96	93.00	93.61	93.81	98.92	99.00	95.86	99.77
roads	88.52	88.62	85.71	87.87	89.75	76.41	80.74	90.65
OA	95.74	95.54	94.05	93.58	97.81	94.74	93.70	98.26
AA	93.56	93.43	91.27	90.76	93.99	90.65	90.16	95.73
Kappa	93.35	93.22	93.54	92.45	95.57	92.20	89.46	97.41

Table 4. The computational cost of different models on various datasets

View table

Table 4. The computational cost of different models on various datasets

Dataset	Complexity	Two-Branch CNN	Coupled CNN	Context CNN	CRNN	FusAtNet	ViT	SpectralFormer	CFCGNet
Houston	Training time	727.39	184.43	2198.58	289.62	1329.44	309.49	399.01	1162.37
Houston	Testing time	15.24	1.37	9.02	1.41	5.15	1.76	2.01	4.45
MUUFL	Training time	338.32	88.67	972.82	135.97	773.52	155.44	175.74	589.76
MUUFL	Testing time	7.04	0.76	3.97	0.70	3.60	4.94	6.31	2.94
Trento	Training time	549.03	117.29	1947.73	217.76	394.22	76.2	95.81	312.40
Trento	Testing time	17.21	1.65	8.70	1.06	12.44	4.85	3.74	9.47

Table 5. Experimental results of ablation of model structure

View table

Table 5. Experimental results of ablation of model structure

Model			MUUFL			Honston 2018			Trento
MCFF_base	MCFF	CGFI	OA	AA	Kappa	OA	AA	Kappa	OA	AA	Kappa
×	×	×	86.51	87.29	83.01	88.65	86.73	87.53	95.74	91.18	93.04
×	√	×	90.44	91.18	86.85	92.57	88.09	89.97	96.93	93.71	95.20
×	×	√	89.15	90.68	86.21	91.84	88.81	90.16	96.42	94.86	96.07
√	×	√	90.59	92.36	87.73	92.40	89.92	90.70	97.39	95.35	96.92
×	√	√	92.13	93.77	89.04	94.77	91.14	91.68	98.26	95.73	97.41

Table 6. Experimental results of ablation of loss function
View table
Table 6. Experimental results of ablation of loss function
Model MUUFL Honston 2018 Trento
$L_{C E}$ $L_{W C E}$ OA AA Kappa OA AA Kappa OA AA Kappa
√ × 89.04 90.18 86.74 91.53 87.73 88.21 96.44 94.06 95.10
× √ 92.13 93.77 89.04 94.77 91.14 91.68 98.26 95.73 97.41

Tools

Get Citation

Copy Citation Text

Dan Fan, Zhengwei Yang, Xia Li, Chao Feng, Chuangjiang Rao. Cross-Feature Granularity Fusion Network for Land Cover Classification of Hyperspectral Remote Sensing Images and LiDAR[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1028002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Remote Sensing and Sensors

Received: Oct. 28, 2024

Accepted: Feb. 10, 2025

Published Online: Apr. 23, 2025

The Author Email: Zhengwei Yang (3274458043@qq.com)

DOI:10.3788/LOP242189

CSTR:32186.14.LOP242189

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Classification evaluation metrics of different models on the MUUFL dataset

Table 1. Classification evaluation metrics of different models on the MUUFL dataset

Table 2. Classification evaluation metrics of different models on the Houston 2018 dataset

Table 2. Classification evaluation metrics of different models on the Houston 2018 dataset

Table 3. Classification evaluation metrics of different models on the Trento dataset

Table 3. Classification evaluation metrics of different models on the Trento dataset

Table 4. The computational cost of different models on various datasets

Table 4. The computational cost of different models on various datasets

Table 5. Experimental results of ablation of model structure

Table 5. Experimental results of ablation of model structure

Table 6. Experimental results of ablation of loss function

Table 6. Experimental results of ablation of loss function