12-16, 15:45–16:00 (Asia/Bangkok), Auditorium Hall 2
Clouds, atmospheric disturbances, and sensor failures influence the quality of Earth observation (EO) data and satellite images, in particular. Many modeling techniques and statistical analysis, to be applied to EO data, require the detection and removal of such aberrations. However, the data gap created after removing the involved pixels needs to be imputed with numerical values that resemble the expected noncorrupted ones. Several imputation, or gap-filling, methods available in literature are based on time-series reconstruction, working only on the temporal dimension of each pixel to input the missing values. Such methods, compared to alternative ones that also consider spatial neighbor pixels or data fusion with other sensors, have the advantage of maintaining the same spatial resolution and spectral consistency in the imputed data.
In contrast with methods that only work with a local temporal window, some of these methods take advantage of the whole time series of each pixel to reconstruct each missing value, allowing the full reconstruction of each gappy time-series. Nevertheless, such methods, like most recent image propagation or linear interpolation, often require an iterative search of available values along the time-series. When the time-series is composed of several samples and/or the involved number of pixels is large, the application of such methods leads to prohibitive computational costs.
We present in this work a computational framework based on discrete convolution that numerically approximates such methods and does not require iterating over the time-series to be applied [1]. In addition, the framework flexibility allows the application of different time-series reconstruction methods by only adapting the convolution kernel. The framework has been used to reconstruct the PetaByte scale Landsat Analysis Ready Data (ARD) collection provided by the Global Land Analysis and Discovery team (GLAD) [2]. New research fronts include the extension of the method to data-fusion approaches that combine time-series of multiple sensors to maintain the highest spatial resolution while also using the temporal information provided by all the sensors. The code, developed in Python with a C++ backend to guarantee usability and high computational efficiency, is openly available at https://github.com/openlandmap/scikit-map.
[1] Consoli, Davide & Parente, Leandro & Simoes, Rolf & Murat, & Tian, Xuemeng & Witjes, Martijn & Sloat, Lindsey & Hengl, Tomislav. (2024). A computational framework for processing time-series of Earth Observation data based on discrete convolution: global-scale historical Landsat cloud-free aggregates at 30 m spatial resolution. 10.21203/rs.3.rs-4465582/v1.
[2] Potapov, Peter & Hansen, Matthew & Kommareddy, Anil & Kommareddy, Anil & Turubanova, Svetlana & Pickens, Amy & Adusei, Bernard & Tyukavina, Alexandra & Ying, Qing. (2020). Landsat Analysis Ready Data for Global Land Cover and Land Cover Change Mapping. Remote Sensing. 12. 426. 10.3390/rs12030426.
Clouds, atmospheric disturbances, and sensor failures affect the quality of Earth observation (EO) data, necessitating the detection and removal of these aberrations, followed by imputation of missing values. Traditional gap-filling methods based on time-series reconstruction can be computationally expensive. We present a computational framework that using discrete convolution offers an efficient solution for reconstructing large-scale EO datasets and supports various time-series reconstruction methods. The code is openly available on GitHub under the scikit-map repository.
