Comparison of techniques for missing lines reconstruction of RapidEye imagery

Abstract

Due to sensor or ground station failure and poor atmospheric conditions, missing information is a common problem that reduces the usage of optical remote sensing data. A large number of algorithms have been developed to reconstruct missing information. Simple methods are compared for filling missing data in the case that one spectral band presents missing information but the other bands are complete. The methods are (1) spatial convolution filtering, (2) unsupervised classification using the complete bands and assignment of the cluster’s average value to missing stripe pixels, (3) global regression models, (4) geographically weighted regression (GWR) models, and (5) a combination of classification and linear models. To evaluate the performance of the different methods, missing line stripes of different sizes are simulated and reconstructed using the five methods. Then, root mean square error (RMSE), correlation, maximum deviance, and bias are computed. To identify the conditions related to the performance of each method, some characteristics of the missing lines, as correlation between spectral bands and spatial autocorrelation, are also calculated. The unsupervised classification based on the larger number of clusters, in particular when combined with linear regression, presents a low RMSE due to its ability to identify the spectral signature of the objects present in the scene. The GWR model performs better than global regression model because it is able to fit the relationship between the missing band and the complete bands locally, which is an important advantage in heterogeneous landscape. Spatial filtering is the most inaccurate method except for one pixel-width missing line.