Laser & Optoelectronics Progress, Volume. 58, Issue 2, 0228002(2021)
Classification of Individual Tree Species in High-Resolution Remote Sensing Imagery Based on Convolution Neural Network
Figures & Tables(14)
Fig. 1. Location of the research area. (a) Huangshan City, Anhui Province; (b) true color schematic of WorldView3, the box indicates the location of Huangshan Mountain
Fig. 2. Construction steps of sample set of remote sensing imagery of individual tree species. (a) Remote sensing imagery of research area; (b) distribution diagram of tree species; (c) delineation diagram of tree crown; (d) labeling diagram of tree crown category; (e) remote sensing imagery of individual tree species; (f) sample set of remote sensing imagery of individual tree species
Fig. 3. Classification labeling result of sample set of remote sensing imagery of individual tree species
Fig. 4. Histogram of training accuracy, validation accuracy, and network layers when CNN model converges
Table 1. Results of sample set division before and after data augmentation
View tableTable 1. Results of sample set division before and after data augmentation
Tree species Training sample set Validation sample set Test sample set Before After Before After Ph.h 66 396 23 138 23 E.a 266 1596 89 534 89 C.l 83 498 28 168 28 Pi.h 1199 7194 401 2406 401 D.a 369 2214 124 744 124 Total 1983 11898 665 3990 665
Table 2. LeNet5_relu model parameter
View tableTable 2. LeNet5_relu model parameter
Layer Output size Parameter Input 32×32×8 - Convolutional C1 28×28×6 Kernel 5×5, filter 6, stride 1, ReLU Pooling S1 14×14×6 Average_pooling 2×2, stride 2 Convolutional C2 10×10×16 Kernel 5×5, filter 16, stride 1, ReLU Pooling S2 5×5×16 Average_pooling 2×2, stride 2 Convolutional C3 1×1×120 Kernel 5×5, filter 120, stride 1, ReLU Fully-connected F1 84 Node 84, FC, ReLU Classification 5 Node 5, FC, Softmax
Table 3. AlexNet_mini model parameter
View tableTable 3. AlexNet_mini model parameter
Layer Output size Parameter Input 32×32×8 - Convolutional C1 32×32×12 Kernel 7×7, filter 12, stride 1, ReLU Pooling S1 15×15×12 Average_pooling 3×3, stride 2 Convolutional C2 15×15×36 Kernel 5×5, filter 36, stride 1, ReLU Pooling S2 7×7×36 Average_pooling 3×3, stride 2 Convolutional C3 7×7×54 Kernel 3×3, filter 54, stride 1, ReLU Convolutional C4 7×7×54 Kernel 3×3, filter 54, stride 1, ReLU Convolutional C5 3×3×36 Kernel 3×3, filter 36, stride 1, ReLU Pooling S3 3×3×36 Average_pooling 3×3, stride 2 Fully-connected F1 320 Node 320, FC, ReLU Fully-connected F2 100 Node 100, FC, ReLU Classification 5 Node 5, FC, Softmax
Table 4. GoogLeNet_mini56 model parameter
View tableTable 4. GoogLeNet_mini56 model parameter
Layer Output size Parameter Input 32×32×8 -- Convolutional C1 32×32×12 Kernel 7×7, filter 12, stride 1 Inception V1 block (1a) 32×32×32 -- Inception V1 block (1b) 32×32×60 -- Pooling S1 15×15×60 Max_pooling 3×3, stride 2 Inception V1 block (2a) 15×15×64 -- Inception V1 block (2b) 15×15×64 -- Inception V1 block (2c) 15×15×64 -- Inception V1 block (2d) 15×15×66 -- Inception V1 block (2e) 15×15×104 -- Pooling S2 7×7×104 Max_pooling 3×3, stride 2 Inception V1 block (3a) 7×7×104 -- Inception V1 block (3b) 7×7×128 -- Pooling S3 1×1×128 Average_pooling 7×7, stride 1 Classification 5 Node 5, FC, Softmax
Table 5. Inception V1 block parameter
View tableTable 5. Inception V1 block parameter
Layer Inception V1 block Ⅰ Ⅱ Ⅲ Ⅳ Convolutional kernel 1×1 Bottleneck kernel 1×1 Convolutional kernel 3×3 Bottleneck kernel 1×1 Convolutional kernel 5×5 Pooling 3×3 Bottleneck kernel 1×1 1a Filter 8 Filter 12 Filter 16 Filter 2 Filter 4 Stride 1 Filter 4 1b Filter 16 Filter 16 Filter 24 Filter 4 Filter 12 Stride 1 Filter 8 2a Filter 24 Filter 12 Filter 26 Filter 2 Filter 6 Stride 1 Filter 8 2b Filter 20 Filter 14 Filter 28 Filter 3 Filter 8 Stride 1 Filter 8 2c Filter 16 Filter 16 Filter 32 Filter 3 Filter 8 Stride 1 Filter 8 2d Filter 14 Filter 18 Filter 36 Filter 4 Filter 8 Stride 1 Filter 8 2e Filter 32 Filter 20 Filter 40 Filter 4 Filter 16 Stride 1 Filter 16 3a Filter 32 Filter 20 Filter 40 Filter 4 Filter 16 Stride 1 Filter 16 3b Filter 48 Filter 24 Filter 48 Filter 6 Filter 16 Stride 1 Filter 16
Table 6. ResNet_mini56 model parameter
View tableTable 6. ResNet_mini56 model parameter
Layer Output size Parameter Residual blockfilter Ⅰ Ⅱ Ⅲ Bottleneck kernel 1×1 Convolutional kernel 3×3 Bottleneck kernel 1×1 Input 32×32×8 -- -- Convolutional C1 16×16×12 Kernel 7×7, filter 12, stride 2 -- Residual block (1) Pooling S1 8×8×12 Max_pooling 3×3, stride 2 -- Bottleneck 8×8×12 × 3 12 -- -- Convolutional 8×8×12 -- 12 -- Bottleneck 8×8×48 -- -- 48 Residual block (2) Bottleneck 4×4×24 × 4 24 -- -- Convolutional 4×4×24 -- 24 -- Bottleneck 4×4×96 -- -- 96 Residual block (3) Bottleneck 2×2×48 ×8 48 -- -- Convolutional 2×2×48 -- 48 -- Bottleneck 2×2×192 -- -- 192 Residual block (4) Bottleneck 1×1×96 × 3 96 -- -- Convolutional 1×1×96 -- 96 -- Bottleneck 1×1×384 -- -- 384 Classification 384 Global_average_pooling -- 5 Node 5, FC,Softmax
Table 7. DenseNet_BC_mini56 model parameter
View tableTable 7. DenseNet_BC_mini56 model parameter
Layer Output size Parameter Dense block filter Ⅰ Ⅱ Bottleneck kernel 1×1 Convolutional kernel 1×1 Input 32×32×8 -- -- Convolutional C1 32×32×12 Kernel 3×3, filter 12, stride 2 -- Dense block (1) Bottleneck 32×32×42 ×5 24 -- Convolutional -- 6 Compression (1) 32×32×21 Kernel 1×1 -- 16×16×21 Average_pooling 2×2, stride 2 Dense block (2) Bottleneck 16×16×51 × 5 24 -- Convolutional -- 6 Compression (2) 16×16×36 Kernel 1×1 -- 8×8×36 Average_pooling 2×2, stride 2 Dense block (3) Bottleneck 8×8×66 ×5 24 -- Convolutional -- 6 Compression (3) 8×8×51 Kernel 1×1 -- 4×4×51 Average_pooling 2×2, stride 2 Dense block (4) Bottleneck 4×4×81 × 5 24 -- Convolutional -- 6 Compression (4) 4×4×66 Kernel 1×1 -- 2×2×66 Average_pooling 2×2, stride 2 Dense block (5) Bottleneck 2×2×99 ×5 24 -- Convolutional -- 6 Classification 99 Global_average_pooling -- 5 Node 5, FC, Softmax
Table 8. CNN model parameter
View tableTable 8. CNN model parameter
Model name Total parameter Trainable parameter Non-trainable parameter Network layer LeNet5_relu 62331 62331 0 5 AlexNet_mini 213537 213537 0 8 GoogLeNet_mini56 97251 97227 24 56 ResNet_mini56 934025 924401 9624 56 DenseNet_BC_mini56 82979 78839 4140 56
Table 9. Classification accuracy evaluation index of CNN model
View tableTable 9. Classification accuracy evaluation index of CNN model
Model name Evaluation index Tree species Ph.h E.a C.l Pi.h D.a LeNet5_relu Producer accuracy /% 86.96 67.42 75.00 98.00 87.90 User accuracy /% 90.91 76.92 87.50 93.57 90.08 Overall accuracy /% 90.68 Kappa coefficient 0.84 AlexNet_mini Producer accuracy /% 86.96 71.91 64.29 97.76 92.74 User accuracy /% 90.91 79.01 78.26 94.00 94.26 Overall accuracy /% 91.58 Kappa coefficient 0.85 GoogLeNet_mini56 Producer accuracy /% 95.65 74.16 75.00 97.76 97.58 User accuracy /% 100.00 84.62 95.45 94.92 93.08 Overall accuracy /% 93.53 Kappa coefficient 0.89 ResNet_mini56 Producer accuracy /% 95.65 76.40 78.57 97.51 92.74 User accuracy /% 100.00 80.95 100.00 94.90 92.00 Overall accuracy /% 92.93 Kappa coefficient 0.88 DenseNet_BC_mini56 Producer accuracy /% 95.65 75.28 85.71 98.25 95.97 User accuracy /% 100.00 87.01 100.00 94.71 94.44 Overall accuracy /% 94.14 Kappa coefficient 0.90
Tools
Get Citation
Copy Citation Text
Guang Ouyang, Linhai Jing, Shijie Yan, Hui Li, Yunwei Tang, Bingxiang Tan. Classification of Individual Tree Species in High-Resolution Remote Sensing Imagery Based on Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0228002
Paper Information
Category: Remote Sensing and Sensors
Received: Jun. 12, 2020
Accepted: Jul. 3, 2020
Published Online: Jan. 11, 2021
The Author Email: Jing Linhai (jinglh@radi.ac.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20