Hyperspectral Image Classification Based on Three-Dimensional Dilated Convolutional Residual Neural Network

Mingjing Yan; Xiyou Su

doi:10.3788/AOS202040.1628002

Journals >Acta Optica Sinica >Volume 40 >Issue 16 >Page 1628002 > Article

Acta Optica Sinica
Vol. 40, Issue 16, 1628002 (2020)

Hyperspectral Image Classification Based on Three-Dimensional Dilated Convolutional Residual Neural Network

Mingjing Yan and Xiyou Su^*

Author Affiliations

School of Information Science and Technology, Beijing Forestry University, Beijing 100083, China

show less

DOI: 10.3788/AOS202040.1628002 Cite this Article Set citation alerts

Mingjing Yan, Xiyou Su. Hyperspectral Image Classification Based on Three-Dimensional Dilated Convolutional Residual Neural Network[J]. Acta Optica Sinica, 2020, 40(16): 1628002 Copy Citation Text

EndNote(RIS)

BibTex

Plain Text

show less

Fig. 1. Two-dimensional and three-dimensional convolution network diagrams. (a) 2D-CNN; (b) 3D-CNN

Download full size

Fig. 2. Normal and dilated convolution kernel diagrams. (a) Normal kernel; (b) dilated kernel (r=2)

Download full size

Fig. 3. Seven permutation and combination types of dilated and normal convolutional layers and two activation function distribution strategies. (a) Type 1; (b) type 2; (c) type 3; (d) type 4; (e) type 5; (f) type 6; (g) type 7; (h) distribution strategy Ⅰ; (i) distribution strategy Ⅱ

Download full size

Fig. 4. Receptive field's distributions of different convolution combinations. (a) Dilation parameter distribution is (2,2,2); (b) dilation parameter distribution is (1,2,2); (c) dilation parameter distribution is (1,1,2)

Download full size

Fig. 5. Network structure

Download full size

Fig. 6. Corresponding precision of dilation rate in two datasets. (a) Indian Pines spectral dimension; (b) Salinas spectral dimension; (c) Indian Pines spatial dimension; (d) Salinas spatial dimension

Download full size

Fig. 7. Pseudo-color composite images of two datasets. (a) Indian Pines; (b) Salinas

Download full size

Fig. 8. Classification images of different network models in Indian Pines dataset. (a) True value image; (b) SVM; (c) 2D-CNN; (d) Res-3DCNN; (e) M3D-DCNN; (f) 3D-CNN; (g) Dilated-3D-CNN

Download full size

Fig. 9. Classification images of different network models in Salinas dataset. (a) True value image; (b) SVM; (c) 2D-CNN; (d) Res-3DCNN; (e) M3D-DCNN; (f) 3D-CNN; (g) Dilated-3D-CNN

Download full size

Structure type	Indian Pines				Salinas
Structure type	Kappa	OA	AA	OA-mean	Kappa		OA		AA	OA-mean
Type 1-Ⅰ	95.852	96.976	92.041	96.942	95.797	96.986		94.524		96.823
Type 2-Ⅰ	95.676	96.851	91.499		95.534	96.794		94.748
Type 3-Ⅰ	95.885	97.000	91.383		95.388	96.690		94.694
Type 4-Ⅰ	95.155	96.480	90.501	96.505	95.862		97.028		95.027	96.943
Type 5-Ⅰ	94.811	96.232	89.872		95.808		96.991		94.772
Type 6-Ⅰ	95.609	96.801	91.559		95.554		96.810		94.666
Type 7-Ⅰ	95.680	96.855	91.335	96.855	95.320		96.644		94.192	96.644

Table 1. [in Chinese]

Structure type	Indian Pines				Salinas
Structure type	Kappa	OA	AA	OA-mean	Kappa	OA	AA	OA-mean
Type 1-Ⅱ	95.976	97.066	91.912	96.973	95.804	96.989	94.793	96.944
Type 2-Ⅱ	95.698	96.866	91.800		95.883	97.045	94.903
Type 3-Ⅱ	95.866	96.987	91.956		95.532	96.796	94.432
Type 4-Ⅱ	95.775	96.921	91.839	96.689	95.455	96.741	94.262	96.856
Type 5-Ⅱ	94.798	96.220	90.179		95.718	96.928	94.737
Type 6-Ⅱ	95.783	96.927	91.543		95.676	96.898	94.682
Type 7-Ⅱ	95.618	96.808	91.351	96.808	95.457	96.739	94.568	96.739

Table 2. [in Chinese]

Class name	Classification accuracy
Class name	SVM	2D-CNN	Res-3DCNN	M3D-DCNN	3D-CNN	Dilated-3D-CNN
Background	69.123	98.797	98.406	99.240	99.376	99.653
Alfalfa	24.348	52.174	73.587	69.130	82.065	82.065
Corn-notill	61.583	85.007	84.352	88.235	93.092	94.492
Corn-mintill	40.126	80.624	86.640	89.329	92.948	94.555
Corn	29.473	77.426	84.325	88.270	93.418	92.806
Grass-pasture	73.113	83.013	83.046	86.115	90.971	91.766
Grass-trees	75.993	89.370	89.411	92.493	94.945	96.075
Grass-pasture-mowed	23.750	58.036	70.893	70.714	78.750	81.786
Hay-windrowed	85.471	97.406	97.416	97.699	98.180	98.441
Oats	22.750	49.250	47.000	72.000	75.250	87.000
Soybean-notill	55.710	84.767	84.268	88.971	93.897	95.307
Soybean-mintill	71.536	91.440	89.668	93.733	96.449	97.132
Soybean-clean	40.780	72.686	75.941	86.169	89.771	92.119
Wheat	84.293	93.317	94.756	95.683	96.585	97.390
Woods	55.486	86.644	86.067	92.518	95.636	95.945
Buildings-Grass-Trees-Drives	28.455	58.503	63.965	77.532	87.946	90.303
Stone-Steel-Towers	33.441	77.796	81.667	86.237	90.860	89.570
Kappa	54.791	88.317	88.514	92.541	95.377	96.304
OA	64.636	91.721	91.820	94.622	96.634	97.303
AA	51.496	78.603	81.848	86.710	91.185	92.730

Table 3. [in Chinese]

Class name	Classification accuracy
Class name	SVM	2D-CNN	Res-3DCNN	M3D-DCNN	3D-CNN	Dilated-3D-CNN
Background	70.502	98.295	98.365	98.491	98.672	98.806
Brocoli-green-weeds-1	94.556	95.768	91.594	86.573	95.888	96.374
Brocoli-green-weeds-2	95.456	98.332	97.772	98.552	98.355	98.395
Fallow	44.098	86.180	73.336	86.282	89.653	90.567
Fallow-rough-plow	60.740	82.123	80.289	86.596	88.238	90.585
Fallow-smooth	47.043	87.957	87.580	89.402	91.423	91.356
Stubble	98.964	96.419	95.926	87.176	96.583	96.638
Celery	91.562	96.462	96.159	97.399	97.122	97.837
Grapes-untrained	83.636	92.165	92.618	93.300	95.896	96.643
Soil-vinyard-develop	64.362	93.482	92.085	93.865	95.142	96.079
Corn-senesced-green-weeds	79.297	92.466	90.876	92.527	95.597	95.165
Lettuce-romaine-4wk	71.196	94.180	93.153	94.094	95.781	95.058
Lettuce-romaine-5wk	31.651	95.916	90.497	94.714	97.738	98.093
Lettuce-romaine-6wk	22.858	86.096	71.331	90.291	91.745	92.651
Lettuce-romaine-7wk	52.255	86.670	83.944	88.426	89.789	90.792
Vinyard-untrained	38.987	85.982	83.519	87.330	93.218	95.129
Vinyard-vertical-trellis	97.662	95.990	96.006	96.231	96.068	96.388
Kappa	61.951	93.267	92.013	93.226	95.567	96.149
OA	70.697	95.185	94.324	95.178	96.820	97.236
AA	67.343	92.028	89.121	91.838	94.524	95.091

Table 4. [in Chinese]

Mingjing Yan, Xiyou Su. Hyperspectral Image Classification Based on Three-Dimensional Dilated Convolutional Residual Neural Network[J]. Acta Optica Sinica, 2020, 40(16): 1628002

Download Citation

EndNote(RIS)

BibTex

Plain Text

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information