Figure 1.
Partial example plot of the Potsdam dataset.
Figure 1.
Partial example plot of the Potsdam dataset.
Figure 2.
Partial example plot of the Vaihingen dataset.
Figure 2.
Partial example plot of the Vaihingen dataset.
Figure 3.
Overall structure of the ASPP+-LANet network.
Figure 3.
Overall structure of the ASPP+-LANet network.
Figure 4.
ASPP+ module structure.
Figure 4.
ASPP+ module structure.
Figure 5.
Schematic diagram of the “grid effect” [
25]. (
a) The “grid effect” in atrous convolutions. (
b) Reasonable combination of dilation rates in atrous convolutions.
Figure 5.
Schematic diagram of the “grid effect” [
25]. (
a) The “grid effect” in atrous convolutions. (
b) Reasonable combination of dilation rates in atrous convolutions.
Figure 6.
BottleNeck structure of ResNet50.
Figure 6.
BottleNeck structure of ResNet50.
Figure 7.
Schematic diagram of FReLU.
Figure 7.
Schematic diagram of FReLU.
Figure 8.
Visual comparison of semantic segmentation for small object features on the Potsdam dataset. (
a) Image, (
b) Ground truth, (
c) UNet, (
d) SegNet, (
e) DeepLabv3+, (
f) LANet, (
g) MANet, (
h) UnetFormer, (
i) Swin-CNN, (
j) ASPP
+-LANet. The colors represent the same types of ground object as shown in
Figure 1, and the same applies to other similar images.
Figure 8.
Visual comparison of semantic segmentation for small object features on the Potsdam dataset. (
a) Image, (
b) Ground truth, (
c) UNet, (
d) SegNet, (
e) DeepLabv3+, (
f) LANet, (
g) MANet, (
h) UnetFormer, (
i) Swin-CNN, (
j) ASPP
+-LANet. The colors represent the same types of ground object as shown in
Figure 1, and the same applies to other similar images.
Figure 9.
Visual comparison of semantic segmentation for large object features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 9.
Visual comparison of semantic segmentation for large object features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 10.
Visual comparison of semantic segmentation for slender ground objects features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 10.
Visual comparison of semantic segmentation for slender ground objects features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 11.
Visual comparison of semantic segmentation for limbic ground objects features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 11.
Visual comparison of semantic segmentation for limbic ground objects features on the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 12.
Visual comparison of semantic segmentation for the missing detection of object features in the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 12.
Visual comparison of semantic segmentation for the missing detection of object features in the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 13.
Visual comparison of semantic segmentation for the false detection of object features in the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 13.
Visual comparison of semantic segmentation for the false detection of object features in the Potsdam dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 14.
Visual comparison of semantic segmentation for small object features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 14.
Visual comparison of semantic segmentation for small object features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 15.
Visual comparison of semantic segmentation for large object features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 15.
Visual comparison of semantic segmentation for large object features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 16.
Visual comparison of semantic segmentation for slender ground objects features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UNetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 16.
Visual comparison of semantic segmentation for slender ground objects features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UNetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 17.
Visual comparison of semantic segmentation for limbic ground objects features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 17.
Visual comparison of semantic segmentation for limbic ground objects features on the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 18.
Visual comparison of semantic segmentation for the missing detection of object features in the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 18.
Visual comparison of semantic segmentation for the missing detection of object features in the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 19.
Visual comparison of semantic segmentation for the false detection of object features in the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 19.
Visual comparison of semantic segmentation for the false detection of object features in the Vaihingen dataset. (a) Image, (b) Ground truth, (c) UNet, (d) SegNet, (e) DeepLabv3+, (f) LANet, (g) MANet, (h) UnetFormer, (i) Swin-CNN, (j) ASPP+-LANet.
Figure 20.
Visual comparisons of the ablation experiments conducted on the Potsdam dataset: (a) Image, (b) Grond Truth, (c) LANet, (d) LANet + ASPP, (e) LANet + ASPP+, (f) LANet + FReLU, (g) ASPP+-LANet.
Figure 20.
Visual comparisons of the ablation experiments conducted on the Potsdam dataset: (a) Image, (b) Grond Truth, (c) LANet, (d) LANet + ASPP, (e) LANet + ASPP+, (f) LANet + FReLU, (g) ASPP+-LANet.
Figure 21.
Visual comparisons of the ablation experiments conducted on the Vaihingen dataset: (a) Image, (b) Grond Truth, (c) LANet, (d) LANet + ASPP, (e) LANet + ASPP+, (f) LANet + FReLU, (g) ASPP+-LANet.
Figure 21.
Visual comparisons of the ablation experiments conducted on the Vaihingen dataset: (a) Image, (b) Grond Truth, (c) LANet, (d) LANet + ASPP, (e) LANet + ASPP+, (f) LANet + FReLU, (g) ASPP+-LANet.
Figure 22.
Effect of data augmentation on semantic segmentation results for the Potsdam dataset.
Figure 22.
Effect of data augmentation on semantic segmentation results for the Potsdam dataset.
Figure 23.
Effect of data augmentation on semantic segmentation results for the Vaihingen dataset.
Figure 23.
Effect of data augmentation on semantic segmentation results for the Vaihingen dataset.
Table 1.
Segmentation accuracy of different methods on the Potsdam dataset.
Table 1.
Segmentation accuracy of different methods on the Potsdam dataset.
Method | Parameters(M) | PA/% | F1/% | MIoU/% | Kappa |
---|
UNet | 17.27 | 92.66 | 78.08 | 71.35 | 0.9492 |
SegNet | 29.45 | 92.61 | 77.61 | 70.84 | 0.9491 |
DeepLab V3+ | 21.94 | 90.00 | 72.24 | 64.17 | 0.8913 |
LANet | 23.81 | 93.29 | 78.77 | 72.29 | 0.9496 |
MANet | 35.86 | 92.06 | 76.89 | 69.86 | 0.9256 |
UNetFormer | 11.28 | 91.23 | 75.01 | 67.51 | 0.9138 |
Swin-CNN | 66 | 94.56 | 81.68 | 76.62 | 0.9521 |
ASPP+-LANet | 27.46 | 95.53 | 82.57 | 77.81 | 0.9552 |
Table 2.
Segmentation accuracy of different methods on the Vaihingen dataset.
Table 2.
Segmentation accuracy of different methods on the Vaihingen dataset.
Method | Parameters(M) | PA/% | F1/% | MIoU/% | Kappa |
---|
UNet | 17.27 | 98.03 | 81.83 | 79.53 | 0.9637 |
SegNet | 29.45 | 96.82 | 80.21 | 76.77 | 0.9433 |
DeepLab V3+ | 21.94 | 92.77 | 73.31 | 67.33 | 0.8721 |
LANet | 23.81 | 97.55 | 80.82 | 77.77 | 0.9465 |
MANet | 35.86 | 98.08 | 81.81 | 79.55 | 0.9677 |
UNetFormer | 11.28 | 96.73 | 80.08 | 76.52 | 0.9429 |
Swin-CNN | 66 | 97.98 | 81.66 | 78.86 | 0.9625 |
ASPP+-LANet | 27.46 | 98.24 | 81.99 | 79.83 | 0.9689 |
Table 3.
Results of ablation experiments on the Potsdam dataset.
Table 3.
Results of ablation experiments on the Potsdam dataset.
Method | PA/% | F1/% | MIoU/% |
---|
LANet | 93.29 | 78.77 | 72.29 |
LANet + ASPP | 93.71 | 79.46 | 73.29 |
LANet + ASPP+ | 93.86 | 79.80 | 73.75 |
LANet + FReLU | 95.22 | 82.05 | 77.06 |
ASPP+-LANet | 95.53 | 82.57 | 77.81 |
Table 4.
Results of ablation experiments on the Vaihingen dataset.
Table 4.
Results of ablation experiments on the Vaihingen dataset.
Method | PA/% | F1/% | MIoU/% |
---|
LANet | 97.55 | 80.82 | 77.77 |
LANet + ASPP | 97.77 | 81.31 | 78.65 |
LANet + ASPP+ | 97.80 | 81.42 | 78.79 |
LANet + FReLU | 97.76 | 81.30 | 78.59 |
ASPP+-LANet | 98.24 | 81.99 | 79.83 |
Table 5.
Comparative experiments with different dilation rates of ASPP+ on the ASPP+-LANet.
Table 5.
Comparative experiments with different dilation rates of ASPP+ on the ASPP+-LANet.
Dilation Rate | PA/% | F1/% | MIoU/% |
---|
(1, 2, 4, 6, 8) | 95.50 | 82.51 | 77.74 |
(1, 2, 4, 8, 12) | 95.53 | 82.57 | 77.81 |
(1, 3, 6, 12, 18) | 95.47 | 82.39 | 77.60 |
(1, 3, 8, 16, 18) | 95.46 | 82.51 | 77.74 |
(1, 3, 8, 18, 24) | 95.51 | 82.52 | 77.73 |
Table 6.
Experimental Comparisons of Different Activation Functions on the LANet Network.
Table 6.
Experimental Comparisons of Different Activation Functions on the LANet Network.
Activation Function | PA/% | F1/% | MIoU/% |
---|
LANet + LeakyReLU [26] | 93.34 | 78.73 | 72.31 |
LANet + PReLU [28] | 94.37 | 80.65 | 74.94 |
LANet + ELU [30] | 90.23 | 72.58 | 64.83 |
LANet + Mish [31] | 89.99 | 73.50 | 65.74 |
LANet + DY-ReLU [32] | 94.10 | 80.26 | 74.40 |
LANet + FReLU | 95.22 | 82.05 | 77.06 |