The inaugural UCC AI Quest 2023 challenged teams to build AI for climate action: specifically, semantic segmentation of vegetation patches from high-resolution drone images of Irish landscapes, scored on intersection-over-union (IoU). Our team placed as finalists in the vegetation patches image segmentation task.
The pipeline moves aerial image–annotation pairs through data pre-processing (augmentation and normalization), training with validation, and finally testing. On the augmentation side we leaned on random crop, JPEG compression, shift-scale-rotate, horizontal and vertical flipping, hue-saturation-value jitter, and Gaussian blur to make the model robust to the variability of drone capture.
The model is DeepLabV3 with a ResNet101 backbone. The gains that mattered most came from the add-ons: a learning-rate scheduler, in-pipeline JPEG compression, customised normalization statistics for the drone imagery, and a learning-rate / batch-size finder to pick training hyperparameters directly instead of guessing. Along the way we also explored U-Net and FCN (ResNet50 and ResNet101 backbones), batch sizes of 8, 16, 32, both Adam and AdamW optimizers, and 224 × 224 input resizing.
Code: View on GitHub.