34. Semantic interpretation of multispectral maps for precision agriculture: a machine learning approach

The effectiveness of precision agriculture relies on accurate, proper and extensive crop knowledge. For this task, remote sensing plays a crucial role, with a large amount of data acquired by sensors mounted on aerial unmanned vehicles (UAVs) that allow crop status maps to be obtained. In this work, a method to semantically interpret multispectral imagery of an apple orchard is presented, which is based on the U-net machine learning approach. The algorithm is able to automatically detect pixels of the map which represent the crop canopy by analysing multispectral images. In addition, an application of this method in a precision agriculture task was studied: the generation of refined vigour maps of apple orchards. The proposed method was validated using data acquired during an experimental campaign. Once trained with a small section of the multispectral map, the semantic interpretation algorithm was used to process the entire orchard map, obtaining a classification error lower than 5%. Processing orchards map with dense inter-row grassing, the U-Net based method showed higher accuracy with respect to the standard Otsu method. In addition, the vigour map, refined by the proposed semantic interpretation method, was found to be less biased by the interrow status, enhancing the reliability of the NDVI evaluation.