SEMANTIC SEGMENTATION OF OIL-CONTAMINATED LAND FROM PLANETSCOPE IMAGERY USING NEURAL NETWORK MODELS

Abstract

Monitoring of oil-contaminated land is a pressing challenge for Kazakhstan. During the Soviet era, oil production at a number of fields relied on the so-called pit method, whereby drilling waste, process fluids, and associated wastewater were collected in purpose-dug earthen pits adjacent to extraction sites, leaving a substantial legacy of oil-contaminated areas requiring remediation and systematic monitoring. Remote sensing combined with deep learning-based image processing enables more accurate segmentation of such territories and facilitates their ongoing monitoring. This paper presents ROSID-HR, a dataset constructed from PlanetScope multispectral imagery designed for the segmentation of small-area oil contamination. In developing ROSID-HR, oil-contaminated sites from our previously published ROSID dataset were used as reference annotations, ensuring consistency between the two datasets. Through systematic experimentation, the optimal PlanetScope channel configuration for separating oil-contaminated areas from shadows and anthropogenic objects was identified as Blue-Red-NIR (bands 2-6-8). To validate the dataset, semantic segmentation was performed using the convolutional DeepLabv3+ model and the transformer-based Mask2Former architecture. Mask2Former achieved an IoU of 69.73 % for the oil class on the test set, exceeding the DeepLabv3+ result of 21.63 % by a factor of more than three. The results confirm the effectiveness and advantages of using PlanetScope data and the Mask2Former architecture together as part of the considered approaches to detecting small areas contaminated with oil.