Incorporation of Multiscale Hierarchical Features for Remote-Sensing Classification

Fu Lü; Yuxuan Xie; Yongan Feng

doi:10.3788/LOP241179

Journals >Laser & Optoelectronics Progress >Volume 62 >Issue 2 >Page 0228004 > Article

Laser & Optoelectronics Progress
Vol. 62, Issue 2, 0228004 (2025)

Incorporation of Multiscale Hierarchical Features for Remote-Sensing Classification

Fu Lü^1,2,*, Yuxuan Xie¹, and Yongan Feng¹

Author Affiliations

¹School of Software, Liaoning Technical University, Huludao 125105, Liaoning , China

²Department of Basic Teching, Liaoning Technical University, Huludao 125105, Liaoning , China

show less

DOI: 10.3788/LOP241179 Cite this Article Set citation alerts

Fu Lü, Yuxuan Xie, Yongan Feng. Incorporation of Multiscale Hierarchical Features for Remote-Sensing Classification[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0228004 Copy Citation Text

show less

Fig. 1. MHFNet structure diagram

Download full size

Fig. 2. Structure diagrams of two attention mechanisms. (a) MHSA; (b) MSIA

Download full size

Fig. 3. Various FFN structure diagrams. (a) FFN; (b) IRFFN; (c) GDFN; (d) MGFN

Download full size

Fig. 4. Structure diagrams of two token extractors. (a) STE; (b) CTE

Download full size

Fig. 5. Scene instances of two datasets. (a) Airplane; (b) beach; (c) ship; (d) cloud; (e) island; (f) bridge; (g) parking; (h) center; (i) storage tank; (j) playground

Download full size

Fig. 6. Confusion matrix of AID dataset under 50% training ratio

Download full size

Fig. 7. Confusion matrix of NWPU-RESISC45 dataset under 20% training ratio

Download full size

Fig. 8. Visual CAM comparison. (a) Basketball court; (b) bridge; (c) church; (d) freeway; (e) roundabout; (f) ship; (g) parking

Download full size

Parameter	Value
CPU	i7-12700
GPU	NVIDIA GeForce RTX 3090Ti 24 GB
Language	Python 3.9.12
Operating system	Windows 11
Deep learning framework	PyTorch 1.11.0+CUDA 11.3

Table 1. Experimental configuration

Parameter	Value
Epochs	300
Initial learning rate	0.0002
Final learning rate	0.00001
Optimizer	AdamW
Batch size	128
Warmup	20
Weight decay	0.003
Random seed	42

Table 2. Key parameter setting

Method	Parameters /10⁶	NWPU（1∶9）	NWPU（2∶8）	AID（2∶8）	AID（5∶5）
VGGNet16^［27］	138.36	76.47	79.79	86.59	89.64
GoogleNet^［27］	54.40	76.19	78.48	83.44	86.39
ResNet50^［29］	23.58	86.23	88.93	92.39	94.96
DenseNet121^［30］	7.98	88.31	90.47	93.76	94.73
ViT-B^［11］	85.63	87.59	90.87	91.16	94.44
PVT-M^［31］	43.32	90.51	92.66	92.84	95.93
DeiT-B^［32］	85.97	91.86	93.83	93.41	96.04
Swin-B^［33］	89.74	91.80	94.14	94.86	97.80
EMTCAL^［34］	–	91.63	93.65	94.69	96.41
GCSANet^［35］	8.11	93.39	94.65	95.96	97.53
MGSNet^［36］	–	92.40	94.57	95.46	97.18
EMSCNet（ViT-B）^［37］	173.64	93.58	95.37	96.02	97.35
MHFNet	41.85	94.09	95.73	97.12	98.63

Table 3. OA of different models on AID dataset and NWPU-RESISC45 dataset

Module	Parameters /10⁶	NWPU（1∶9）	NWPU（2∶8）	AID（2∶8）	AID（5∶5）
None	31.18	92.27	94.10	95.38	96.93
ATM	32.53	92.88	94.74	96.05	97.57
ATM+MSIT	41.85	94.09	95.73	97.12	98.63

Table 4. Impact of MSIT and ATM on model performance

Option	Parameters /10⁶	NWPU（1∶9）	NWPU（2∶8）	AID（2∶8）	AID（5∶5）
N=0	32.53	92.88	94.74	96.05	97.57
N=1	37.68	93.79	95.57	96.74	98.37
N=2	41.85	94.09	95.73	97.12	98.63
N=3	46.02	93.82	95.66	96.89	98.46
N=4	50.18	93.74	95.59	96.76	98.38

Table 5. Impact of number of MSITs in each branch on model performance

Module	Parameters /10⁶	NWPU（1∶9）	NWPU（2∶8）	AID（2∶8）	AID（5∶5）
None	32.53	92.88	94.74	96.05	97.57
MSIA	38.43	93.73	95.41	96.75	98.30
MSIA+MGFN	41.76	93.97	95.65	96.94	98.52
MSIA+MGFN+LRC	41.85	94.09	95.73	97.12	98.63

Table 6. Effectiveness of various components in MSIT

Method	Parameters /10⁶	FLOPs /G	NWPU（1∶9）	NWPU（2∶8）	AID（2∶8）	AID（5∶5）
MHFNet（ResNet50）	34.61	4.71	92.63	94.58	95.39	96.94
MHFNet（ViT-B）	96.17	56.30	93.28	94.83	95.21	96.47
MHFNet（Swin-B）	102.38	16.09	93.74	95.80	96.83	98.92
MHFNet（FasterNet-S）	41.85	5.12	94.09	95.73	97.12	98.63

Table 7. Impact of different feature extractors on model performance

Category	OA	Precision	Recall	F1 score	Category	OA	Precision	Recall	F1 score
1	99.63	99.26	100.00	99.63	16	100.00	100.00	99.55	99.77
2	99.12	100.00	100.00	100.00	17	95.04	95.93	94.19	95.05
3	100.00	100.00	100.00	100.00	18	100.00	100.00	100.00	100.00
4	98.81	99.78	99.38	99.58	19	99.83	99.04	98.59	98.81
5	99.44	99.00	99.50	99.25	20	99.45	99.39	98.22	98.80
6	99.47	99.22	90.81	94.83	21	99.61	100.00	100.00	100.00
7	97.22	97.67	100.00	98.82	22	99.80	100.00	100.00	100.00
8	97.08	97.61	98.37	97.99	23	96.33	96.35	95.60	95.97
9	97.87	97.23	100.00	98.60	24	98.09	98.78	98.11	98.44
10	97.94	98.78	100.00	99.39	25	96.91	97.72	97.83	97.77
11	98.91	98.72	100.00	99.36	26	97.32	99.59	98.65	99.12
12	99.08	100.00	100.00	100.00	27	95.15	95.25	99.60	97.38
13	98.52	99.65	99.41	99.53	28	99.44	99.16	96.19	97.65
14	97.66	98.15	100.00	99.07	29	100.00	100.00	99.01	99.50
15	99.18	99.26	100.00	99.63	30	99.28	100.00	100.00	100.00

Table 8. OA, precision, recall rate, and F1 score for each category in AID dataset

Category	OA	Precision	Recall	F1 score	Category	OA	Precision	Recall	F1 score
1	98.67	99.47	100.00	99.73	24	91.96	91.57	88.27	89.89
2	97.60	98.75	98.16	98.45	25	97.38	98.16	100.00	99.07
3	94.14	94.56	94.93	94.74	26	97.01	98.93	96.35	97.62
4	98.88	99.15	91.62	95.24	27	96.87	97.42	94.12	95.74
5	98.45	99.43	96.17	97.77	28	90.71	90.91	84.75	87.72
6	97.57	98.16	97.33	97.74	29	95.48	96.88	100.00	98.42
7	100.00	100.00	99.81	99.90	30	96.93	96.14	92.13	94.09
8	86.48	86.92	88.05	87.48	31	93.26	93.68	93.69	93.68
9	98.55	98.19	100.00	99.09	32	96.79	98.76	93.47	96.04
10	98.81	98.45	99.27	98.86	33	96.17	96.81	93.19	94.97
11	89.90	89.98	96.76	93.25	34	97.52	98.06	94.13	96.05
12	93.19	93.12	92.11	92.61	35	97.58	98.03	97.56	97.79
13	91.30	91.22	98.65	94.79	36	100.00	100.00	100.00	100.00
14	97.27	98.15	98.70	98.42	37	96.15	98.56	97.31	97.93
15	96.31	97.07	90.62	93.73	38	99.22	100.00	99.23	99.61
16	99.23	100.00	100.00	100.00	39	95.57	96.27	97.56	96.91
17	85.18	85.50	97.29	91.01	40	93.40	94.90	98.22	96.53
18	99.63	100.00	100.00	100.00	41	95.70	96.52	100.00	98.23
19	93.41	93.19	93.86	93.52	42	86.51	86.44	87.74	87.09
20	92.49	92.26	92.81	92.53	43	93.97	93.20	98.08	95.58
21	93.57	93.04	97.32	95.13	44	98.19	99.65	97.16	98.39
22	91.40	91.29	95.85	93.51	45	97.45	98.47	89.91	94.00
23	96.13	97.50	95.96	96.72

Table 9. OA, precision, recall rate, and F1 score for each category in NWPU-RESISC45 dataset

Fu Lü, Yuxuan Xie, Yongan Feng. Incorporation of Multiscale Hierarchical Features for Remote-Sensing Classification[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0228004

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享