Chinese Journal of Lasers, Volume. 48, Issue 16, 1610003(2021)
Multi-Dimensional CNN Fused Algorithm for Hyperspectral Remote Sensing Image Classification
Figures & Tables(11)
Fig. 2. Process diagrams of decomposed 3D CNN and 3D-2D-1D CNN. (a) Decomposed 3D CNN; (b) 3D-2D-1D CNN
Fig. 3. Hyperspectral data used in experiment. (a) Indian Pines; (b) Pavia University; (c) Salinas Scene; (d) WHU-Hi-Han Chuan
Fig. 4. Correlation coefficient graphs of spectral and spatial features of Indian Pines dataset. (a) Correlation coefficient of spectral features; (b) correlation coefficient of spatial features
Fig. 5. Test results of each model classification in Indian Pines dataset. (a) Ground truth; (b) SVM; (c) 2D CNN; (d) 3D CNN; (e) 3D-2D CNN; (f) 3D-2D-1D CNN
Fig. 6. Overall classification accuracy and loss of proposed model in 100 epochs. (a) Overall classification accuracy; (b) loss
Table 1. Datasets of Indian Pines, Pavia University, Salinas Scene, and WHU-Hi-Han Chuan
View tableTable 1. Datasets of Indian Pines, Pavia University, Salinas Scene, and WHU-Hi-Han Chuan
Indian Pines Salinas Scene WHU-Hi-Han Chuan Pavia University Class Number of samples Class Number of samples Class Number of samples Class Number of samples Alfalfa 46 Brocoli green weeds 1 2009 Strawberry 44735 Asphalt 6631 Corn notill 1428 Brocoli green weeds 2 3726 Cowpea 22753 Meadows 18649 Corn mintill 830 Fallow 1976 Soybean 10287 Gravel 2099 Corn 237 Fallow rough plow 1394 Sorghum 5353 Trees 3064 Grass pasture 483 Fallow smooth 2678 Water spinach 1200 Shadows 947 Grass trees 730 Stubble 3959 Watermelon 4533 Bare soil 5029 Grass pasture mowed 28 Celery 3579 Greens 5903 Self blocking bricks 3682 Hay windrowed 478 Grapes untrained 11271 Trees 17978 Oats 20 Soil vinyard develop 6203 Grass 9469 Bitumen 1330 Soybean notill 972 Vinyard untrained 7268 Red roof 10516 Painted metal sheets 1345 Soybean mintill 2455 Lettuce romaine 4wk 1068 Gray roof 16911 Soybean clean 593 Lettuce romaine 5wk 1927 Plastic 3679 Wheat 205 Lettuce romaine 6wk 916 Bare soil 9116 Woods 1265 Lettuce romaine 7wk 1070 Road 18560 Buildings grass trees drives 386 Corn-senesced green weeds 3278 Bright object 1136 Stone steel towers 93 Vinyard vertical trellis 1807 Water 75401
Table 2. Convolution training model of Indian Pines dataset
View tableTable 2. Convolution training model of Indian Pines dataset
Layer (type) Output shape Number of parameters Input_1 (Input layer) (None, 25, 25, 30, 1) 0 Conv3d (Conv3D) (None, 23, 23, 24, 8) 512 Conv3d_1 (Conv3D) (None, 21, 21, 20, 16) 5776 Reshape (Reshape) (None, 21, 21, 320) 0 Conv2d (Conv2D) (None, 19, 19, 32) 92192 Reshape_1 (Reshape) (None, 19, 608) 0 Conv1d (Conv1D) (None, 17, 64) 116800 Flatten (Flatten) (None, 1088) 0 Dense (Dense) (None, 256) 278784 Dropout (Dropout) (None, 256) 0 Dense_1 (Dense) (None, 128) 32896 Dropout_1 (Dropout) (None, 128) 0 Dense_2 (Dense) (None, 16) 2064 Total number of parameters: 5361913
Table 3. Classification accuracies of each model in Indian Pines, Pavia University, Salinas Scene and WHU-Hi-Han Chuan datasets unit: %
View tableTable 3. Classification accuracies of each model in Indian Pines, Pavia University, Salinas Scene and WHU-Hi-Han Chuan datasets unit: %
Dataset Accuracy of classification SVM 2D CNN 3D CNN 3D-2D CNN 3D-2D CNN (new) 3D-2D-1D CNN Indian Pines OA 69.675 89.560 96.962 99.331 94.328 99.652 AA 51.652 94.443 97.638 98.141 93.483 98.974 KAPPA 64.498 84.364 96.526 99.238 93.512 99.603 Pavia University OA 71.690 97.262 98.834 99.930 99.953 99.947 AA 55.553 98.655 98.456 99.882 99.909 99.883 KAPPA 57.041 91.957 98.456 99.907 99.938 99.929 Salinas Scene OA 93.418 95.238 99.937 100.000 100.000 100.000 AA 96.784 99.993 99.895 100.000 100.000 100.000 KAPPA 92.663 93.103 99.929 100.000 100.000 100.000 WHU-Hi-Han Chuan OA 81.575 99.271 99.917 99.956 99.953 99.849 AA 62.336 98.205 99.800 99.802 99.872 99.816 KAPPA 78.136 99.146 99.903 99.948 99.945 99.823
Table 4. Classification accuracy of each model for each ground object sample in Indian Pines dataset unit:%
View tableTable 4. Classification accuracy of each model for each ground object sample in Indian Pines dataset unit:%
No. Class SVM 2D CNN 3D CNN 3D-2D CNN 3D-2D CNN(new) 3D-2D-1D CNN 1 Alfalfa 52.174 15.217 100.000 87.500 87.500 100.000 2 Corn notill 84.804 71.919 96.700 98.500 90.600 99.600 3 Corn mintill 73.855 50.723 81.928 100.000 87.608 100.000 4 Corn 64.135 25.316 100.000 100.000 83.133 98.795 5 Grass pasture 89.648 65.010 98.225 100.000 97.633 99.408 6 Grass trees 96.027 87.945 99.804 100.000 99.609 99.609 7 Grass pasture mowed 71.429 21.429 100.000 95.000 100.000 90.000 8 Hay windrowed 89.331 61.088 100.000 100.000 100.000 100.000 9 Oats 45.000 5.000 100.000 92.857 92.857 100.000 10 Soybean notill 82.510 57.819 95.000 100.000 88.235 100.000 11 Soybean mintill 89.939 84.236 99.767 99.011 99.651 99.767 12 Soybean clean 77.572 45.363 97.831 98.313 91.807 99.277 13 Wheat 97.561 52.195 100.000 99.301 98.601 98.601 14 Woods 94.704 91.700 98.871 99.774 94.018 100.000 15 Buildings grass trees drives 69.430 40.933 94.074 100.000 93.704 98.519 16 Stone steel towers 79.570 50.538 100.000 100.000 90.769 100.000
Table 5. Training time and test time of each model in Indian Pines, Pavia University, Salinas Scene, and WHU-Hi-Han Chuan datasets
View tableTable 5. Training time and test time of each model in Indian Pines, Pavia University, Salinas Scene, and WHU-Hi-Han Chuan datasets
Dataset Classification performance SVM 2D CNN 3D CNN 3D-2D CNN 3D-2D CNN(new) 3D-2D-1D CNN Indian Pines Training time /s 1.18 901.20 1477.93 968.71 642.97 600.77 Test time /s 0.98 6.51 18.04 20.04 14.28 13.37 Pavia University Train time /s 2.23 180.28 953.88 725.11 564.86 564.32 Test time /s 4.38 4.37 21.20 23.29 20.17 19.81 Salinas Scene Train time /s 3.41 572.69 1155.50 920.70 706.34 664.93 Test time /s 8.34 5.5 25.93 29.42 25.43 22.83 WHU-Hi-Han Chuan Training time /s 678.335 1396.175 7935.224 5372.430 3410.025 3228.423 Test time /s 862.016 27.298 148.706 149.471 129.047 114.472
Tools
Get Citation
Copy Citation Text
Jinxiang Liu, Wei Ban, Yu Chen, Yaqin Sun, Huifu Zhuang, Erjiang Fu, Kefei Zhang. Multi-Dimensional CNN Fused Algorithm for Hyperspectral Remote Sensing Image Classification[J]. Chinese Journal of Lasers, 2021, 48(16): 1610003
Paper Information
Category: remote sensing and sensor
Received: Dec. 23, 2020
Accepted: Feb. 25, 2021
Published Online: Jul. 30, 2021
The Author Email: Kefei Zhang (profkzhang@gmail.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20