基于深度学习的在轨辐射定标方法研究

On-orbit Radiometric Calibration Method Research Based on Deep Learning Theory

辐射定标是将卫星传感器的计数值转化为具有物理意义的数值的关键环节。传统的在轨定标方法都是基于一天的数据,定标精度受限于当天的地面测量数据和天气情况。文章提出了一种基于深度学习的在轨定标新方法,其思想是利用定标场地的大量历史卫星影像、历史大气数据和历史光谱数据,通过对这些数据的学习和筛选,构建和真实场景最接近的定标场地模型。利用这一定标场地模型,模拟出待定标卫星成像时刻对应观测几何下的表观反射率,实现传感器的绝对辐射定标。为验证新方法的有效性,分别利用场地定标法、交叉定标法和深度学习定标法对"高分一号"卫星PMS2相机进行在轨辐射定标。结果表明,深度学习定标法的定标精度和场地定标法接近,优于交叉定标方法,且具备交叉定标方法的成本低、频率高、可实现历史数据再定标等优点,是一种比较理想的在轨定标新方法。

Radiometric calibration is the key to transform a satellite sensor dimensionless value into a physical value. Traditional on-orbit calibration methods are based on one-day data, whose calibration accuracy is limited by the in situ measurement data and the weather that day. This paper proposes a new on-orbit calibration method based on deep learning, which is to establish a model of calibration site through the deep learning of historical satellite images,atmosphere and spectral data of this calibration site. The apparent reflectance corresponding to the observation satellite geometry can be simulated through the new calibration site model, and then the absolute radiometric calibration is realized. The validation of the new method is performed with the site calibration, cross calibration and deep learning calibration of GF-1 PMS2. The results shows that the calibration precision of the new method is similar to site calibration while is better than the cross calibration method. In addition, it also has the advantages of cross calibration method, such as lower cost, and higher frequency. In brief,the new on-orbit calibration method is feasible.