基于像元级不变特征的遥感器响应衰变智能自定标研究

Aiming at the problem of the lack of calibration information of historical remote sensing data and the bad data consistency, this study proposes a set of intelligent self-calibration method based on the analysis of sequence images by using artificial intelligence and big data analysis technology to overcome the high input of traditional site calibraton and the large amount of manual processing of the pseudo-invariant targets calibration. In this method, the large uniform target required by traditional vicarious calibration is reduced to the pixel level. It adopts deep learning of CNN and multivariate change detection (MAD) technology to extract pixel invariant features, and uses two-stage Trimap algorithm to detect the automatic filtering of cloud, fog, haze, dust and other significant change targets in the image before invariant pixel extraction, and realizes automatic quality control technology such as data defect and geolocation error of image. From invariant pixels based on long sequence of satellite image pairs, a weighted 1-norm polynomial fit model is adopted to realize that degradation modeling of instument, and finally, the accuracy of self-calibration method is validated based on various traditional calibration methods. This study relies solely on satellite imagery to implement the intelligent self-calibration of the remote sensing instruments, which reduces the manual intervention for the mass data reprocessing. It will provide a novel technological approach to the historical data recalibration of the multi-seriel remote sensing satellites in our country.

针对历史遥感资料定标信息缺失和数据一致性不足的问题，克服传统场地替代定标试验高投入和地球稳定场景定标的大量人工处理，本研究借助于人工智能和大数据分析技术，提出了一套只依赖于序列图像分析的智能自定标方法。该方法将传统替代定标所需的较大范围均匀目标区域缩减到像元级别，它采用深度学习CNN和多元变化检测（MAD）技术手段智能提取像元级不变特征，且在不变像元提取前，采用两阶段抠图（Trimap）算法检测图像中云、雾、霾、沙尘等显著变化目标的自动过滤，并实现图象的数据缺陷、定位偏差等自动质量控制技术。基于上述智能提取的长序列卫星图像对的不变像元，采用加权的1-范数多项式拟合模型实现遥感器衰变建模，最后基于多种传统定标方法开展自定标方法精度验证。本研究仅依赖卫星图像实现遥感器响应衰减的智能自定标，能减少大规模资料再定标处理的人工干预，将为我国多序列遥感卫星历史数据再定标提供全新的技术手段。

本研究针对历史遥感资料定标信息缺失和数据一致性不足的问题，克服传统场地替代定标试验高投入和地球稳定场景定标的大量人工处理，借助于人工智能和大数据分析手段，提出了一套只依赖于序列图像分析的智能自定标方法。该方法首先采用人工智能抠图、语义分割等方式检测出卫星图像中的云、雾、霾等显著变化目标并对其过滤，然后利用图像对进行智能学习自动甄别出其中的准不变像元，最后基于这些准不变像元拟合辐射校正系数和响应衰减曲线，实现仅基于图像的遥感仪器自定标。该方法克服了传统场地定标方式人工、时间以及经济成本过高的缺点，同时解决了场地定标方法较难对长序列历史数据进行再定标的困窘，该技术只依赖卫星图像实现遥感器响应衰减的智能自定标，能减少大规模资料再定标处理的人工干预，为我国多序列遥感卫星历史数据再定标、全球准不变定标场智能选取和遥感仪器间辐射基准传递提供全新的技术手段。