Floating litter detection using remote sensing images

Abstract

After a thorough Literature Review was conducted on the ”Marida” dataset, the UNet method was found to be the most accurate and effective model. For the tar- get class, Marine Debris, an IoU (Intersection over Union) value of 0.78 was specifi- cally reached. Renowned as a benchmark dataset, the Marida dataset’s wide range of classes—observable via satellite observations over the ocean—make it useful for training models.This paper presents a series of experiments conducted on the MARIDA dataset using various image segmentation algorithms. We explored the performance of the UNet algorithm with different input configurations, including all bands with augmented data and a selective combination of indexes and bands like FDI Contrast Stretching, NDVI Contrast Stretching, and B0 Contrast Stretching. In data augmentation, we didn’t follow the traditional augmentation techniques. We did pixel wise augmentation. Both config- urations resulted in an Intersection over Union (IoU) score of 0, indicating no effective segmentation. Similarly, a Hybrid Approach which is a combination of autoencoder and random forest, incorporating all bands also failed to produce meaningful results, as evi- denced by an IoU score of 0. In contrast, a statistical Approach using the same subset of bands (FDI contrast stretching, NDVI contrast stretching , and B0 contrast stretching ) yielded promising results, particularly on images without land features. These findings highlight the challenges of segmenting the MARIDA dataset and suggest that specialized statistical methods may offer more robust performance in certain scenarios.