Preprocessing Impact on SAR Oil Spill Image Segmentation Using YOLOv8

Abstract

Synthetic Aperature Radar (SAR) is a sensory equipment used in marine remote sensing that emits radio waves to capture a representation of the target scene. SAR images have poor quality, one of which is due to speckle noise. This research uses SAR images containing oil spills as objects that are detected using machine learning with the YOLOv8 model. The dataset was obtained from MKLab by preprocessing to improve the quality of SAR images before processing. Preprocessing involves annotating the dataset, augmenting it with flip augmentation, and filtering it using threshold and median filters in addition to a sharpen kernel that finds the optimal midway value. The default value of the YOLOv8 hyperparameter is used with addition of delta as well as subtraction of the same delta.

The implementation of preprocessing and combination of hyperparameters is examined to optimize the YOLOv8 model in detecting oil spills in SAR images. Based on 10 experimental scenarios, initial results with the original MKLab image provide an mAP50 of 49.7%. Implementing Flip augmentation alone on the data set increases the mAP50 value by 18.8%. Followed by the sharpen 1.2 kernel filter increasing the mAP50 value to 68.89%, while the median and thresholding filters tend to reduce the mAP50 value. The combination of experiments with the best results was preprocessing with flip augmentation and sharpen 1.2 kernel filter with hyperparameters: epoch 200, warmup 4.0, momentum 0.9, warmup bias lr 0.01, weight decay 0.005, and learning rate 0.000714, resulting in an mAP50 value of 68.89%.  In addition, it was found that the sharpening kernel with a real number midpoint of 1.2 and combination with flipping augmentation had the greatest impact on increasing the MAP50 value in SAR oil spill image segmentation by YOLOv8.