jensleitloff / cnn-sentinel Goto Github PK

i want the classification of wheat using sentinel 2 imagery so it is able to classify images only from this data source?
What should be least number tif maps to train the model?
How to save the model in the path you mentioned in classify images?
What are the basic codes only to classify sentinel two tif?

I am using the following command to get RGB images:

gdal_translate -b 4 -b 3 -b 2 -of JPEG -scale inputimage.tif outputfile.jpeg

RGB images obtained from the above command is not same as supplied by EuroSAT.
Can you let me know what is the procedure that you have used to covert EuroSAT GEOTIFF to JPEG ?

Hi

Can you share the file "../data/karlsruhe/2018_zugeschnitten.tif".
Is the query file also need to have all the 13 bands.

But I have met other problems, ms_label only 0 and 10, not have 1 - 9 kinds.
Currently testing images in mountainous areas. (use Sentinel-2 RGB bands image)

What should I consider when test images?

Originally posted by @daniellinboy in #8 (comment)

@felixriese
I know you mean.
Thanks very much for the assistance you provide my help.

More further,I obtain result of S2 image. But I don't know value range 0.67-0.83 of ms_prob why not get an integer value. attachment figure

Sorry to bother you.

Originally posted by @daniellinboy in #7 (comment)

@jensleitloff - thanks very much for your tutorial and code which is very useful.
06_classify_image.py by default produces two outputs: a label.tif file and a prob.tif file. The label TIF file when I run the code on a VGG RGB model built using transfer learning is a grayscale image with 10 'colours' that presumably correspond to the 10 land cover classes in the eurosat dataset? Is there a legend you could share that relates the results in this TIF image to those classes?
The prob.tif output file may be the probability by pixel of the selected class? Any more information about this output is appreciated.
Thanks

I've been testing out the code and the training has been working as expected with both the pretrained and networks and custom on the eurostat dataset however, I've been having issues interpreting the results of the classifier. I also tried using the models that were provided on google drive as well as the image for prediction but I'm still facing the same issue. Any insight on this would be greatly appreciated.

The results I am getting are provided in the link below using the Models provided on google drive on other issues.
(Results for my custom own training and prediction have been giving me very similar results)
https://drive.google.com/drive/folders/1UBRjIYioRjjB8kzHoNN0bt2M85jTifrE?usp=sharing

Thanks for the code! As I just start to work with machine learning models, then I hope you can help me with your code. I tuned a RGB model and after that made a confusion matrix and classification report, for EuroSAT imagery (All classes) a got a low metrics (precision, recall and f1-score ranged within 0.09 to 0,12), however other parameters were ok (loss: 0.0208 - categorical_accuracy: 0.9924 - val_loss: 0.0805 - val_categorical_accuracy: 0.9764). For main RGB set (four main classes in landscape (deciduous forest, coniferous forest, agriculture land and lakes) the accuracy metrics ranged from 0.2 to 0.3, also low, but at the same time categorical accuracies were high. I wondering why I got such low accuracy metrics and is there any solution to increase them.

@jensleitloff I'm having trouble interpreting the classes created by a model in the label.tif file. 02_train_rgb_finetuning.py outputs the classes as:
Found 18900 images belonging to 10 classes.
{'AnnualCrop': 0, 'Forest': 1, 'HerbaceousVegetation': 2, 'Highway': 3, 'Industrial': 4, 'Pasture': 5, 'PermanentCrop': 6, 'Residential': 7, 'River': 8, 'SeaLake': 9}
Found 8100 images belonging to 10 classes.
...
When I review the grayscale TIF image created by 06_classify_image.py, the classes don't make sense. For example, the forest class (identified by overlaying the Sentinel-2 satellite image) is displayed as class #9 #090909, which according to the above legend should be the class 'SeaLake'. I must be doing something wrong!

Any advice is really appreciated.

My computer GPU memory have 16G.
GPU memory use about 16G when train_rgb_finetuping.py, but run classify image.py only use about 110 MB.
I don't know how do set memory value in 06_classify_image.py.

Thanks.

Hi all thanks for the great tutorial!
Quick question: is it possible to get a hold of the pretrained model for MS Imagery mentioned here https://github.com/jensleitloff/CNN-Sentinel/blob/master/py/06_classify_image.py#L18 I can't seem to be able to find it anywhere.
Many thanks!

Hi, Thanks for sharing these code. It would have been great to get the step how to prepare data for training as well as for query. Possible to share the same?
Thanks

Hi, Thanks a lot for the code. I trained my model and tried to classify the sentinel image with that. But all I could get is some small patches. Is there any restriction on the size of the test image? I used the same area in Germany as specified in the slide. It would be helpful if you could clarify this.

Hi, first of all, congratulations on your work. I wanna know if there is some paper about your work?