bragagnololu / unet-defmapping Goto Github PK

Hi @bragagnololu !

I'm trying to run your code and I've been having some issues.
What confused me kind a bit was the fact that you proposed two different models: one for forest and one for cloud detection, and they have different dimensions for the input.
Can you give me some details on why the cloud has input_size of (512, 512, 3) while the forest uses (512, 512, 4)?

Giving more details on my current issue: I've been trying to load the forest model and test it on some images, but I'm having the following issue loading the arrays on the following code:

# loading arrays
image_array = np.load("image_array_og.npy") # array of training images
image_array[image_array > 10000] = 10000
image_array = image_array.astype(float)/10000
mask_array = np.load("mask_array_og.npy") # array of training masks

channels_imgs = 4 # number of channels of one image

bands_third = np.zeros(channels_imgs)
bands_nin = np.zeros(channels_imgs)

# getting the percentiles of the training array for normalization
for i in range(channels_imgs):
    bands_third[i] = np.percentile(image_array[:,:,:,i],3)
    bands_nin[i] = np.percentile(image_array[:,:,:,i],97)

np.save('bands_third_og.npy', bands_third)
np.save('bands_nin_og.npy', bands_nin)

PS: I did not change the gen_npy_files.py code, except by adding lines to save the numpy arrays into files.
My error message is the following:

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
/tmp/ipykernel_40875/54951181.py in <module>
     12 # getting the percentiles of the training array for normalization
     13 for i in range(channels_imgs):
---> 14     bands_third[i] = np.percentile(image_array[:,:,:,i],3)
     15     bands_nin[i] = np.percentile(image_array[:,:,:,i],97)
     16 

IndexError: index 3 is out of bounds for axis 3 with size 3

I've experiemented reducing the dimension of the input, but then I start having problems loading the weights of the model (as expected).

I think I can make it work if you give us more details on how to run the code at the /UNet folder.

Dear Lucimara,

I am trying to download the images but it seems that the rondonia_square3.geojson does not exist. Could you please inform me where can I find this file so that I can run the codes?

I am looking forward to your answer

Best regards

Hi,

I recently tried out awesome UNET-defmapping with mostly up to date packages (see below), however with no luck. I wonder maybe a little clarification on originally used package names and versions would solve all the issues I encountered tinkering around and trying to make this amazing code to work in my environment.

Python 3.7.10
? Keras 2.4.3
? Tensorflow 2.3.0
? sklearn: scikit-learn 0.24.2
rasterio 1.1.5
? rkimage ?
fiona 1.8.9.post2
? cv2: opencv-python
numpy-indexed 0.3.5
sentinelsat 1.0.0
zipfile.py
? glob2 0.7
matplotlib 3.4.2

I'd be very grateful for any kind of assistance.