Marine plastic pollution poses significant ecological, economic, and social challenges, necessitating innovative detection, management, and mitigation solutions. Spectral imaging and optical remote sensing have proven valuable tools in detecting and characterizing macroplastics in aquatic environments. Despite numerous studies focusing on bands of interest in the shortwave infrared spectrum, the high cost of sensors in this range makes it difficult to mass-produce them for long-term and large-scale applications. Therefore, we present the assessment and transfer of various machine learning models across four datasets to identify the key bands for detecting and classifying the most prevalent plastics in the marine environment within the visible and near-infrared (VNIR) range. Our study uses four different databases ranging from virgin plastics under laboratory conditions to weather plastics under field conditions. We used Sequential Feature Selection (SFS) and Random Forest (RF) models for the optimal band selection. The significance of homogeneous backgrounds for accurate detection is highlighted by a 97 % accuracy, and successful band transfers between datasets (87 %–91 %) suggest the feasibility of a sensor applicable across various scenarios. However, the model transfer requires further training for each specific dataset to achieve optimal accuracy. The results underscore the potential for broader application with continued refinement and expanded training datasets. Our findings provide valuable information for developing compelling and affordable detection sensors to address plastic pollution in coastal areas. This work paves the way towards enhancing the accuracy of marine litter detection and reduction globally, contributing to a sustainable future for our oceans.

中文翻译：

---

使用选定的光谱带在沿海地区进行有效的塑料检测


海洋塑料污染带来了重大的生态、经济和社会挑战，需要创新的检测、管理和缓解解决方案。光谱成像和光学遥感已被证明是检测和表征水生环境中大塑料的宝贵工具。尽管大量研究集中在短波红外光谱中的感兴趣波段，但该范围内的传感器成本高昂，因此很难大规模生产它们以进行长期和大规模应用。因此，我们提出了跨四个数据集的各种机器学习模型的评估和迁移，以确定在可见光和近红外（VNIR）范围内检测和分类海洋环境中最常见塑料的关键波段。我们的研究使用了四种不同的数据库，从实验室条件下的原始塑料到现场条件下的耐候塑料。我们使用顺序特征选择（SFS）和随机森林（RF）模型来进行最佳频带选择。 97% 的准确度凸显了均匀背景对于准确检测的重要性，并且数据集之间成功的谱带传输 (87%–91%) 表明传感器适用于各种场景的可行性。然而，模型迁移需要对每个特定数据集进行进一步训练才能达到最佳精度。结果强调了通过不断完善和扩展训练数据集进行更广泛应用的潜力。我们的研究结果为开发引人注目且价格实惠的检测传感器以解决沿海地区的塑料污染问题提供了宝贵的信息。这项工作为提高全球海洋垃圾检测和减少的准确性铺平了道路，为我们海洋的可持续未来做出贡献。