I want to downscale a satellite image using a deep learning approach, called extended super-resolution convolutional neural network (ESRCNN). The paper of the method is 'Deep learning-based fusion of Landsat-8 and Sentinel-2 images for a harmonized surface reflectance product'.
The authors claimed that they used some samples to train the model. My question has to do about the labelling of the samples. Assuming that I created the samples, should I label them or not? Because the task is not a classification problem I am confused on if I need to label my samples or not. As you understand I am fairly new to the area of deep learning.
My data set is two satellite images, one coarse resolution image and one fine resolution image. The samples were taken from the fine resolution image. They include various land cover types such as dense vegetation, built-up areas etc.
At p.7 in the paper, the authors mention, and I quote:
The original Landsat-8 bands 1–7 at 30 m were used for training and validating the fusion network. Similar to the self-adaptive fusion, all the input images should be of the same size (spatial extent and pixel size) as each other. In the multi-temporal fusion step, 12160 patches were randomly selected for training and 3040 patches were randomly selected for validating the fusion network.
Why they used patched for training and validation? I am asking this because according to your answer, I understand that I the inputs to the algorithm (ESRCNN) should be the whole images.
seems to me another option for you is the hierarchical or non-hierarchical clustering could be a solution. Since the fundemental diff between the clustering and classification is labelling. Unsupervised ML method
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence