Hi man , here is a link to a writeup + code you could look at to understand how to go about this image regression challenge :https://zindi.africa/competitions/lacuna-solar-survey-challenge/discussions/25675 .

Here's a quick summary of a transfer learning approach you could use: Use a pretrained model from the timm library as your backbone, then add regression heads.

Models to try EfficientNet: tf_efficientnet_b5 or efficientnet_b3 — good balance of speed and accuracy Swin Transformer: swin_base_patch4_window7_224 — good for dense counting ResNet: resnet50 simple and reliable

Architecture pattern : Load a pretrained backbone (freeze initially or fine-tune)

-Add regression heads

-Train end-to-end