Résumé:
Oil spill detection has garnered increasing research interest in recent years due to the
profound impact such incidents have on marine environments, natural resources, and the
livelihoods of coastal communities. Hyperspectral remote sensing imagery offers a wealth
of spectral information, which is highly advantageous for monitoring oil spills in complex
oceanic scenarios. However, most existing methods rely on supervised or semi-supervised
frameworks, requiring substantial effort to annotate a sufficient number of high-quality
training samples. This process can be labor-intensive and time-consuming.
In this study, we use a novel approach which consists of an unsupervised oil spill detection
method based on the isolation forest algorithm tailored for hyperspectral images (HSIs).
The methodology begins with an estimation of noise variance across different spectral
bands because noise levels can vary significantly. Bands severely affected by noise are
subsequently discarded to improve data quality. Next, Principal Component Analysis
(PCA) is employed to reduce the high dimensionality inherent in HSIs, facilitating more
efficient processing.
The core of the approach involves estimating the probability that each pixel belongs to
either the seawater or oil spill class using the isolation forest. This probabilistic information
enables the automatic generation of pseudo-labeled samples through clustering algorithms,
which serve as training data for subsequent classification steps. An initial detection map is
then produced using support vector machines (SVM) on the dimension-reduced data.
To assess the effectiveness of our proposed method, we evaluated the method on dataset
termed the Hyperspectral Oil Spill Dataset (HOSD), comprising eighteen hyperspectral
images capturing oil spills over the Gulf of Mexico in 2010.
