A Patch-Image Based Classification Approach for Detection of Weeds in Sugar Beet Crop

Abstract

Weeds affects crops health as it shares water and nutrients from the soil, as a result it decreases crop yield. Manual weedicide spray through bag-pack is hazardous to human health. Localized autonomous weedicide spray through aerial spraying units can help save water, weedicide chemical and effect less on human health. Such systems require multi-spectral cues to classify crop, weed, and soil surface. Our focus in this paper is on the detection of weeds in the sugar beet crop, using air-borne multispectral camera sensors, which is considered as an alternative crop to sugarcane to obtain sugar in Pakistan. We developed a new framework for weed identification; a patch-based classification approach as appose to semantic segmentation that is more realistic for real-time intelligent aerial spraying systems. Our approach converts 3-class pixel classification problem into a 2-class crop-weed patch classification problem which in turns improves crop and weed classification accuracy. For classification, we developed a new VGG-Beet convolutional neural network (CNN), which is based on generic VGG16 (visual graphics group) CNN model with 11 convolutional layers. For experiments, we captured a sugar beet dataset with 3-channel multispectral sensor with a ground sampling distance (GSD) of 0.2 cm/pixel and a height of 4 meters. For better comparison, we used two publicly available sugar beet crop aerial imagery datasets, captured using a 5-channel multispectral sensor and a 4-Channel multispectral sensor with a ground sampling distance of 1cm and a height of 10 meters. We observed that patch-based method is more robust to different lighting conditions. To produce low cost weed detection system usage of Agrocam sensor is recommended, for higher accuracy Red Edge and Sequoia multispectral sensors with more channels should be deployed. We observed higher crop-weed accuracy and lower testing time for our patch-based approach as compared to U-Net and Deeplab based semantic segmentation networks.

Keywords

• Autonomous weed detection
• drone weed detection
• deep learning in agriculture
• multispectral image processing