Pairwise Comparison Network for Remote-Sensing Scene Classification

Abstract

Remote sensing scene classification aims to assign a specific semantic label to a remote sensing image. Recently, convolutional neural networks have greatly improved the performance of remote sensing scene classification. However, some confused images may be easily recognized as the incorrect category, which generally degrade the performance. The differences between image pairs can be used to distinguish image categories. This paper proposed a pairwise comparison network, which contains two main steps: pairwise selection and pairwise representation. The proposed network first selects similar image pairs, and then represents the image pairs with pairwise representations. The self-representation is introduced to highlight the informative parts of each image itself, while the mutualrepresentation is proposed to capture the subtle differences between image pairs. Comprehensive experimental results on two challenging datasets (AID, NWPU-RESISC45) demonstrate the effectiveness of the proposed network. The codes are provided in https://github.com/spectralpublic/PCNet.git.

FCIHMRT: Feature Cross-Layer Interaction Hybrid Method Based on Res2Net and Transformer for Remote Sensing Scene Classification

“…The experiment was split into two parts, using 50% of the training samples or 20% of the test samples. Table 3 [40] 89.64 ± 0.36 86.59 ± 0.29 CRAN [42] 96.65 ± 0.20 95.24 ± 0.16 MobileNet V2 [43] 95.96 ± 0.27 94.13 ± 0.28 SE-MDPMNet [44] 97.14 ± 0.15 94.68 ± 0.07 Two-Stream Fusion [45] 94.58 ± 0.25 92.32 ± 0.41 ViT [4] 96.88 ± 0.19 95.58 ± 0.18 CFDNN [46] 96.56 ± 0.24 94.56 ± 0.24 Inception-v3-CapsNet [18] 96.32 ± 0.12 93.79 ± 0.13 GSSF [47] 97.65 ± 0.80 95.71 ± 0.22 PCNet [48] 96.76 ± 0.25 95.53 ± 0.16 GAN [26] 96. 45 Table 3 shows that FCIHMRT produced the greatest outcomes at both 50% and 20% training rates.…”

“…Meanwhile, it was demonstrated that the OA of palace scenes was reduced to 80%, in which some palace scenes were categorized into the church intersection and island classes, indicating that the classification capacity of FCIHMRT still requires improvement for similar scenes but in general, it can distinguish different scenes with rich spatial information. [40] 79.79 ± 0.15 76.47 ± 0.18 CRAN [42] 94.07 ± 0.08 91.28 ± 0.19 MobileNet V2 [43] 83.26 ± 0.17 80.32 ± 0.16 SE-MDPMNet [44] 94.11 ± 0.03 91.80 ± 0.07 Two-Stream Fusion [45] 83.16 ± 0.18 80.22 ± 0.22 ViT [4] 94.50 ± 0.18 91.17 ± 0.13 CFDNN [46] 93.83 ± 0.09 91.17 ± 0.13 Inception-v3-CapsNet [18] 92.6 ± 0.11 89.03 ± 0.21 GSSF [47] 94.48 ± 0.26 91.98 ± 0.19 PCNet [48] 94.59 ± 0.07 92.64 ± 0.13 GAN [26] 93.63 ± 0.12 91.06 ± 0. classification accuracy of the golf course and mobile home park classes reached 99%, which indicates that FCIHMRT has a good classification performance for scenes with a small feature complexity. Meanwhile, it was demonstrated that the OA of palace scenes was reduced to 80%, in which some palace scenes were categorized into the church intersection and island classes, indicating that the classification capacity of FCIHMRT still requires improvement for similar scenes but in general, it can distinguish different scenes with rich spatial information.…”

“…GoogLeNet [40] 94.31 ± 0.89 92.70 ± 0.60 VGG-16 [40] 95.21 ± 1.20 94.14 ± 0.69 CRAN [42] 95.75 ± 0.80 94.21 ± 0.75 MobileNet V2 [43] 99.01 ± 0.21 97.88 ± 0.31 SE-MDPMNet [44] 98.95 ± 0.12 98.36 ± 0.14 Two-Stream Fusion [45] 98.02 ± 1.03 96.97 ± 0.75 ViT [4] 99.29 ± 0.34 98.75 ± 0.21 CFDNN [46] 98.62 ± 0.27 97.65 ± 0.18 Inception-v3-CapsNet [18] 99.05 ± 0.24 97.59 ± 0.16 GSSF [47] 99.24 ± 0.47 97.86 ± 0.56 PCNet [48] 99.25 ± 0.37 98.71 ± 0.22 GAN [26] 98.58 ± 0.33 97.54 ± 0. As shown in Figure 7, the confusion matrix with all 21 classes was also created to further examine the performance of FCIHMRT with an 80% training ratio.…”

Do We Still Need Imagenet Pre-Training in Remote Sensing Scene Classification?

“…Target dataset RESISC45 AID ResNet50+EAN (Zhao et al, 2020) 93.51 93.64 GLDBS (Xu et al, 2021) 94.46 95.45 PCNet (Zhang et al, 2021) 94.59 95.53 Million-AID (Long et al, 2022) 94.26 95.40 Domain-adaptive pre-training (ours)…”

“…In Table 10 the results on RESISC45 and AID obtained using DA pre-training are compared to recent state-of-the-art HRRS scene classification methods. Two of the methods, ResNet50+EAN (Zhao et al, 2020) and PCNet (Zhang et al, 2021) use the same ResNet-50 backbone as in our experiments, while GLDBS (Xu et al, 2021) uses ResNet-34. The best results obtained using pre-training on Million-AID used DenseNet-169 and ResNet-101 for classification of RESISC45 and AID, respectively.…”