Scalable pixel-based crop classification combining Sentinel-2 and Landsat-8 data time series: Case study of the Duero river basin

Abstract Satellite imagery is the foremost source of information to analyze and monitor land covers in several time ranges, especially over large areas. However, it is not always either freely available or easily compatible for the final users due to the different resolutions offered by sensors onboard the satellite platforms. Crop classification is an important task to control and make decisions related to the agricultural practice and its regulation. However, it is not trivial, especially for extensive areas. Thus, this paper proposes a new approach for crop classification in large areas by a combined use of multi-temporal open-source remote sensing data from Sentinel-2 (S2) and Landsat-8 (L8) satellite platforms. Having to deal with different spatial and temporal resolutions, special spatial regions (called Tuplekeys) were created within a local nested grid to allow a proper integration between the data of both sensors. Temporal variation of the Normalized Difference Vegetation Index (NDVI) was the chosen input to classify crops. Moreover, due to the massive quantity of data collected, filters considering some agronomic and edaphic criteria were applied with the dual goal of decreasing redundancies and increasing the process efficiency. Out of three different machine learning classifiers analyzed, a plot-based approach was considered for the algorithms calibration while a pixel-based approach was used for the final classification process. The methodology was both tested and validated in the Duero river basin (Spain), 78,859 km 2 , for the 2017 spring and summer seasons. Finally, classification outputs were analyzed throughout their overall accuracy (OA), not only for the whole basin but also for each of the Tuplekeys so that the OA spatial distribution was evaluated as well. The Ensemble Bagged Trees (EBT) algorithm showed the maximum OA, 87% and 92%, when classifying crops individually (15 classes) and grouped (7 classes), respectively, proving both the accuracy and efficiency of the developed approach.