Note_automatic water-body segmentation from high-resolution satellite Images via deep Networks

Basic information

Section Two, Water division
Automatic water-body segmentation from high-resolution satellite Images via deep Networks

Notes starting point

1. Water Division is the basic task of remote sensing.
2. The traditional method relies on spectra and can only handle low resolution images. And the resolution of the picture, contains more details.
3. The robustness of the method is tested by the data obtained from different data sensors.

The main innovation point

1. Proposed a new segmented network RRF deconvnet Network (Restricted receptive field deconvolution networks)
2. It is considered that the loss, or L2 distance, or Euler distance, can not well highlight the role of the boundary in the segmentation, so the use of Ewloss is the Gaussian weighted Euler distance.

Detailed description

Network structure information and network configuration.

It can be seen that the main changes are in the network structure, and a large number of empty convolution to replace the ordinary convolution to obtain a greater sense of the wild.

Using the Gaussian weighted distance as the cost function, the main characteristic idea is to highlight the marginal, the closer the segmentation boundary, the greater the weighted.

Experiment

Data is extracted from Google Earth above, 0.5M, size is 512*512, mainly in Sichuan, Wuhan region, a total of 9,000 pictures, 7:2:1=train:validation:test

Mainly six experiments, (two class loss functions, three networks)

Evaluation indicators are also two, one is overrall pixels, the other is edge pixels. The former is conventional.

\[op= \frac{tp+tn}{tp+tn+fp+fn}\]

The next one is not very clear, and the approximate description is this (pixels with a maximum of 5 checkerboard distances from the boundary are considered to be edge pixels):

 the pixels who have a maximum 5 chessboard distance to boundaries are considered as edge pixels

Looking at the result, it seems that the evaluation function has little effect on the result, or, basically, no effect. OP is the mainstream of the evaluation indicators, why the promotion is not obvious, it is obvious, because the image of the pixel too much, the boundary part of the pixel too little, a little sample imbalance reasons, so even if there is a rise, there will be no obvious change. Therefore, the introduction of new evaluation variables, the purpose of evaluation loss is to achieve, indeed, to make the edge more accurate and clear.

Summarize

1. Change the network structure
2. Change loss
3. Design favorable evaluation Index
4. It is worth mentioning that the framework adopted is mxnet

Note_automatic water-body segmentation from high-resolution satellite Images via deep Networks