Remote sensing image recognition imaging synthesis

effect

Because the data set used in the previous training is sampled from cities , Therefore, the recognition effect of the model on rural areas is particularly poor , Therefore, a dozen datasets of rural areas have been made by hand , Two categories are marked , Respectively waters 、 Woodland . Then use this data set to train the model .
The following is the recognition effect of the two sampling areas on the newly trained model ：

1. City ：
   
2. rural
   

synthesis

Because recognition is for multiple categories , Therefore, it is necessary to synthesize the previously trained model and the image recognized by the current model , Final realization waters 、 building 、 road 、 Woodland The identification of .

WF : Indicates that the water area can be recognized 、 Woodland
RAW : It means that the road can be recognized 、 building 、 waters

The effect of combining the two images is as follows ：

The idea was to output the predicted probability of each pixel to the image alpha passageway , Then it is used to compare probabilities to determine which category the pixels are divided into , The following is the effect picture ：

You can see that the border color of the shaded area on the left will be darker , Because the probability represents the transparency of the image , But this is only used to store the probability of each category .
After doing the above step , It is found that if we just compare the probability , Then we can only compare the probability of water area and forest land , Because the data sets of model training are different , The number of recognition categories is also different , So from these two kinds of effects , The effect of the new training model is better than that of the previous model , Therefore, the water area is finally directly 、 The recognition results of these two types of woodlands with the newly trained model cover the original recognition results .

Ideas for the next step

Manually mark the green space in the city , Then mark the roads and buildings in rural areas , Put in training , See how it works . In addition, use richer data enhancement methods , Only rotation was used before 、 noise , But there are clouds and the like that have not been taken into account , Therefore, using the existing training model to predict the previous picture will still be wrong .

The white line in the left figure is added by yourself , It is mainly used to indicate that the influence of cloud and fog in the area surrounded by white line is relatively large .