As an option: use those images you already have, but! But rotate them and add some (controlled) noise. A type of noise and its intensity are to be discussed specifically, since it heavily depends on the task you want to resolve.

Michael Sadovsky its actually like I have to sample some images from the dataset where maximum images should end up having solar panels.

Is your goal detecting the actual solar panels?

Assuming it is, you could crop the images in different ways so that they contain the solar panels. You could then resize the cropped images and apply other form of augmentation such as rotation, translation etc..

One possible idea to explore would be to apply changes in the scale of the image so that the solar panels occupy more pixels and use it as training.

In addition, using a loss function that compensates for the class imbalance by assigning larger weights to true positives than true negatives might help.

Here you can find a great summary of different loss functions and their properties:

https://lars76.github.io/2018/09/27/loss-functions-for-segmentation.html

Related to your query, I suggest you follow the already existing work, e.g., Article Analysis on Solar Panel Crack Detection Using Optimization Techniques

Article Comprehensive overview of optimizing PV-DG allocation in pow...

You might try first training neural networks to detect solar panels in an image. As you have this, run prediction on the images. With transfer learning, you can train to detect solar panels with less than 300 images containing solar panels.

You could also have a look at the google DeepDream algorithm. This would be useful if your goal is to maximize the detected presence of solar panels by a neural network: https://ai.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html

That could lead to some interesting results.

Cheers, Raoul