Deep Learning for Remote Sensing Images with Open Source Software

18

 

Inference

 

We will apply the previously exported SavedModel on the Sentinel-1 and Sentinel-2 images.

 

18.1  Inputs and outputs

At inference time, only the inputs for S1_t, S2_t−1 and S2_t+1 are fed to the model. The S2t^ input is not used anymore: we only care about retrieving the S2_t image at the output of the model.

Another important point is the output used to retrieve S2_t. As explained in section 12.2.2.1, computing the exact output of a fully convolutional model is one strategy to avoid blocking artifacts. For this purpose, we compute the exact output of the model in extracting the center of the final convolution output. We then remove 256 pixels from each side of the output tensor named prediction in the spatial dimensions. The ...