EVENT DOSSIER
The researchers proposed the UnorthoDOS dataset and method, aimed at training machine learning models to directly process uncropped hyperspectral images, replacing traditional ground-based processing procedures. This study utilized EMIT sensor data and trained the methane detection model using the U-Net model on uncropped images, achieving an average intersection over union (IoU) of 16.91%, which is close to 18.47% for cropped data, and significantly better than the mag1c matching filter baseline (IoU 4.76%). Additionally, FP16 compression reduced the model size by half, with an output deviation of less than 0.3%, demonstrating its feasibility for satellite onboard deployment. The trained machine learning models, two sets of ML-ready datasets, and code have been made available on Hugging Face and GitHub.
研究人员提出 UnorthoDOS 数据集与方法，旨在训练机器学习模型直接处理未正射校正的高光谱图像，以替代传统的地面处理流程。该研究利用 EMIT 传感器数据，通过 U-Net 模型在未经正射校正的图像上训练甲烷检测，其平均交并比（IoU）达 16.91%，接近正射校正数据的 18.47% 表现，且均显著优于 mag1c 匹配滤波器基线（IoU 4.76%）。此外，FP16 压缩将模型体积减半，输出偏差小于 0.3%，证明了其在卫星机载部署的可行性。相关训练好的机器学习模型、两套 ML 就绪数据集及代码已公开于 Hugging Face 和 GitHub。