Remote sensing image recognition misclassification under multi class recognition

Antecedents feed

Following the previous article ：
Remote sensing image recognition - Train with larger datasets
Train through the data set made last time , Finally, we get the fitting depth learning model , Next, analyze the problems and solutions .

1. Preparation before training ( Add )

Because the previous article did not point out the use of data sets, the data sets were mistaken for small , So prepare to draw pictures and instructions to fill the hole left in front .

1.1 Data preprocessing

Because the resolution of the original data set is relatively large , And the resolution of each image is different , Therefore, it cannot be directly put into the model for training , There are two reasons ： First, the resolution is too large , Video memory will explode ; Second, the data size required for training needs to be unified . To do this, do the following operations on the data .

This is a local process , In fact, it is similar to convolution kernel sliding window in deep learning , It is also a sliding cut image . The following figure shows the overall situation , Looking at it may be more direct .

Data set size change results ：
$$30\to 3740\to 7480$$
there $3740\to 7480$ It is the data that doubles the cutting image through data enhancement .

The image resolution can be expressed as ：
$$\mathrm{W}\times \mathrm{H}\sim 2560\times 1440$$
Resolution size change results ：
$$\begin{gathered} \mathrm{W}=\{1004,\dots ,15088\}\to \mathrm{W}=\{512\} \\ \mathrm{H}=\{751,\dots ,8017\}\to \mathrm{H}=\{512\} \end{gathered}$$

The above is the work of data preparation .

2. The effect after training

The last one is to use the current data to run , The previous one used five large pictures to run .

The evaluation index ：

3. What happened

As can be seen from the above figure , The model predicts the green part of the map as water .

3.1 reason

1. The color of water and green space is relatively similar , In terms of the shape of the distribution, it is also relatively similar , And the data are collected from urban areas , Therefore, the waters are Turquoise , Close to green space
2. Because there is almost no annotation of green space in the electronic map , Therefore, the category of green space was not added to the model for recognition in the initial training , It also leads to the model's confusion between green space and water

3.2 reflection

For the above reasons , There are mainly several views

• At the beginning, the data set was made around the theme of electronic map , Then train prediction , But this method is more limited , The quality of the annotation depends entirely on the electronic map , And the category is also determined by it , Therefore, the effect of the model is not ideal .
• If you're just looking for results , Then there may be another way , It is to train different models by using existing data sets of different categories , Finally, integrate the results .

4. Separate identification

Now try to waters and Green space Use a single model to train recognition , Finally, integrate the results with another model , At present, the data set has been made , Then go into training .