UAV-enabled semantic segmentation for precision farming using deep learning

Abstract

In this work, we study the semantic segmentation of captured UAV (Unmanned Aerial Vehicle) images for enhanced crop and weed segmentation in precision agriculture. In the existing literature, researchers studied segmentation techniques; however, there is a need for deep feature extraction to capture the spatial and contextual information, especially for the complex agriculture domain, which involves the overlapping of crop, weed, and background pixels. To address this, we present VResUNet++ architecture that combines VGG16 and ResNet50 in the backbone of UNet for deep semantic feature extraction and helps in improving the segmentation accuracy and performance. This improved segmentation method helps in weed detection, crop health monitoring, and early disease detection. Our hybrid model has outperformed the state-of-the-art models like UNet, UNetResNet50, and UNetVGG16. The outcome of the extensive experiment shows a significant improvement in the precision of 99.83%, recall of 98.65%, and accuracy of 98.69% on the Weedmap dataset.