Semantic segmentation of citrus-orchard using deep neural networks and multispectral UAV-based imagery

Accurately mapping farmlands is important for precision agriculture practices. Unmanned aerial vehicles (UAV) embedded with multispectral cameras are commonly used to map plants in agricultural landscapes. However, separating plantation fields from the remaining objects a scene difficult task traditional algorithms. In this connection, deep learning methods that perform semantic segmentation could help improve overall outcome. study, state-of-the-art segment citrus-trees images were evaluated. For purpose, camera operates at green (530–570 nm), red (640–680 red-edge (730–740 nm) and also near-infrared (770–810 spectral regions was used. The performance of following five pixelwise evaluated: fully convolutional network (FCN), U-Net, SegNet, dynamic dilated convolution (DDCN) DeepLabV3 + . results indicated evaluated performed similarly proposed task, returning F1-Scores between 94.00% (FCN U-Net) 94.42% (DDCN). It determined inference time needed per area and, although DDCN method slower, based on qualitative analysis, it better highly shadow-affected areas. This study demonstrated citrus orchards achievable neural networks. investigated here proved be equally suitable solve providing fast solutions varying 0.98 4.36 min hectare. approach incorporated into similar research, contribute decision-making accurate fields.