Objective

2

Quick view data generated by high-resolution satellites provide real-time reception and full resolution for quick view imaging. Such imaging offers a timely source of data for practical applications

such as fire detection

moving window display

disaster observation

and military information acquisition. Road extraction from remote sensing images has been a popular research topic in the field of remote sensing image analysis. Traditional object-oriented methods are not highly automated

and road features require prior knowledge for manual selection and design. These conditions lead to problems in real-time road information acquisition. The popular deep learning road extraction method mainly focuses on the improvement of precision and lacks research on the timeliness of road information extraction. Transfer learning can rapidly complete the task in the target area through weight sharing among different fields and make the model algorithm highly personalized. A transfer learning deep network for rapidly extracting roads is constructed to utilize quick view data from high-resolution satellites.

Method

2

First

we propose a least-square fitting method of devignetting to solve the most serious radiation problem of TDICCD (time delay and integration charge coupled devices) vignetting phenomenon appearing in raw quick view data. The results of the preprocessing of the quick view data serve as our training dataset. Then

we choose LinkNet as the target network after comparing the performance among different real-time semantic segmentation networks

such as ENet

U-Net

LinkNet

and D-LinkNet. LinkNet is efficient in computation memory

can learn from a relatively small training set

and allows residual unit ease training of deep networks. The rich bypass links each encoder with decoder. Thus

the networks can be designed with few parameters. The encoder starts with a kernel of size 7×7. In the next encoder block

its contracting path to capture context uses 3×3 full convolution. We use batch normalization in each convolutional layer

followed by ReLU nonlinearity. Reflection padding is used to extrapolate the missing context in the training data for predicting the pixels in the border region of the input image. The input of each encoder layer of LinkNet is bypassed to the output of its corresponding decoder. Lost spatial information about the max pooling can then be recovered by the decoder and its upsampling operations. Finally

we modify LinkNet to keep it consistent with ResNet34 network layer features

the so-called fine tuning

for accelerating LinkNet network training process. Fine tuning is a useful efficient method of transfer learning. The use of ResNet34 weight parameter pretrained on ImageNet initializing LinkNet34 can accelerate the network convergence and lead to improved performance with almost no additional cost.

Result

2

In the process of devignetting quick view data

the least-square linear fitting method proposed in this study can efficiently remove the vignetting strip of the original image

which meets practical applications. In our road extraction experiment

LinkNet34 using the pretrained ResNet34 as encoder has a 6% improvement in Dice accuracy compared with that when using ResNet34 not pretrained on the valid dataset. The time consumption of a single test feature map is reduced by 39 ms

and the test Dice accuracy can reach 88.3%. Pretrained networks substantially reduce training time that also helps prevent overfitting. Consequently

we achieve over 88 % test accuracy and 40 ms test time on the quick view dataset. With an input feature map size of 3×256×256 pixels

the data of Tianjin Binhai with a size of 7 304×6 980 pixels take 54 s. The original LinkNet using ResNet18 as its encoder only has a Dice coefficient of 85.7%. We evaluate ResNet50 and ResNet101 as pretrained encoders. The Dice accuracy of the former is not improved

whereas the latter takes too much test time. We compare the performance of LinkNet34 with those of three other popular deep transfer models for classification

namely

U-Net; two modifications of TernausNet and AlubNet using VGG11 (visual geometry group) and ResNet34 as encoders separately; and a modification of D-LinkNet. The two U-Net modifications are likely to incorrectly recognize roads as background or recognize something nonroad

such as tree

as road. D-LinkNet has higher Dice than LinkNet34 on the validation set

but the testing time takes 59 ms more than that of LinkNet34. LinkNet34 avoids the weaknesses of TernuasNet and AlubNet and makes better predictions than them. The small nonroad gap between two roads can also be avoided. Many methods mix the two roads into one. The method proposed in this study generally achieves good connectivity

accurate edge

and clear outline in the case of complete extraction of the entire road and fine location. It is especially suitable for rural linear roads and the extraction of area roads in towns. However

the extraction effect for complex road networks in urban areas is incomplete.

Conclusion

2

In this study

we build a deep transfer learning neural network

LinkNet34

which uses a pretrained network

ResNet34

as an encoder. ResNet34 allows LinkNet34 to learn without any significant increase in the number of parameters

solves the problem that the bottom layer features randomly initialized with weights of neural networks are inadequately rich

and accelerates network convergence. Our approach demonstrates the improvement in LinkNet34 by the use of the pretrained encoder and the better performance of LinkNet34 than other real-time segmentation architecture. The experimental results show that LinkNet34 can handle road properties

such as narrowness

connectivity

complexity

and long span

to some extent. This architecture proves useful for binary classification with limited data and realizes fast and accurate acquisition of road information. Future research should consider increasing the quick view database. The pretrained network LinkNet34 trains on the expanded quick view database and then transfers. The "semantic gap" between the source and target networks is reduced

and the data distribute similarly. These features are conducive to model initialization.