Crop classification for large-scale irrigable areas using interoperable satellite-based and UAV-based remote sensing information

Precision agriculture aims to boost the efficiency of the agriculture monitoring systems both at regional and national scale by providing well-timed information of the phenological state of crops, water stress, and yield estimation among many other important applications. Crop classification maps are essential for irrigation management at basin scale and currently there are multitude of remote sensing tools that offers this information, however, remote sensing data must be processed first to extract the necessary information in order to identify homogeneous groups of pixels that represent thematic classes. The success of crop classification process relies on many factors such as the choice of the spectral, spatial and temporal resolution of the sensors, and of course the choice of a suitable classification procedure. Therefore, in this study, the first main objective is to perform a classification of crops in the Duero basin with the Normalized Difference Vegetation Index NDVI generated by Landsat 8 (L8) and Sentinel 2 (S2) using the TICPA-Classifier program, developed by the research group in which this work has been carried out, which allows interoperability between these two products. Concerning the second objective is about performing a comparison between geomatic products obtained with unmanned aerial vehicles UAV and Sentinel 2 data. The aim of this comparison is trying to find explanation for possible errors that can be generated during the classification process using only satellite-based information, also studying the potential combined use of UAV-based and S2-based information for potential upscaling. The innovation that the TICPA-Classifier program has been able to bring to the classification theme is allowing the crops classification in a large basin that has an area equal to 98,073 Km2. This was possible thanks to the introduction of all the NDVI scenes obtained by L8 and S2 (1270 scenes), that cover all the surface of the Duero basin, during the period March-July of the year 2017 and the proper massive data arrangement with an agronomic point of view combined with the application of an efficient and robust classification algorithm. Three supervised classification method: Ensembled Bagged Trees EBT, Weighted Nearest Neighbor WNN and Decision Trees DT are implemented in the TICPA-Classifier program were tested twice: once the classification was performed with separated crop classes and the other one with aggregated crop classes. Classification with separated crop classes shows that the method with the highest Overall Accuracy OA is Ensembled Bagged Trees EBT 87 per cent, and the two other classifiers their OA were very close 81 per cent and 80 per cent respectively for Decision Tree DT and Weighted Nearest Neighbor WNN. Regarding the classification with aggregated crop classes, the results reveals that the classifier with the highest Overall accuracy is the Ensembled Bagged Trees 92 per cent, followed by Weighted Nearest Neighbor 89 per cent and in the third place Decision Trees 88 per cent. The TICPA-Classifier program allowed us to have very promising results thanks to: -The combined use of Landsat 8 and Sentinel 2 data in order to benefit of their distinct strengths. This improvement was possible by using the concept of generation of Tuplekeys special files in which the NDVI raster scenes of L8 and S2 are located in the Local Nested Grid (database in which data with different pixels size are organized at all the zoom level). Local Nested Grid (estructura espacial en la que se organizan datos con diferentes tamaños de píxeles en todo el nivel de zoom). - The introduction of important agronomic criteria allowing the massive information obtained to be filtered with Landsat 8 and Sentinel 2 during the period March-July 2017 with the purpose to increase the accuracy and efficiency of the classification. Correlation analysis showed strong agreement between NDVI obtained with Sentinel 2 and UAV sensors, that means the potential possibilities of upscaling of UAV data to S2 data level and the satellite-based information can be calibrated using accurate UAV based-information instead of using atmospheric models