A Weed Detection in Teff Crops for Drone-Based Selective Herbicide Spraying Using Fine-Tuned Deep Learning Models

Abstract

Teff is a well-adapted and widely grown staple food in Ethiopia. However, the productivity of Teff is affected by different factors including weeds. Currently, farmers control weeds using manual hand weeding and chemical herbicides. The application of chemical herbicide to the farm field can be done through backpack sprayers, robots, helicopters, and drones. These techniques spray chemical herbicides uniformly on a farm field despite variations in the severity of weeds over a large farm. Therefore, this paper develops a machine learning model that can detect weeds from a Teff farm with their exact location for drone based selective herbicide spraying. For this purpose, 1308 images were captured using a high-resolution drone-mounted digital camera at the range of 1-2 meter heights, over different plant growth stages, and at different weather conditions (clear, cloudy, and rainy) from self-planted University of Gondar Agricultural Research Farm and two Azezeo Kebele farmer’s farm field. Data preprocessing activities such as resizing, enhancement, and unsharp mask filter were applied to improve quality of the dataset. Experimentation was conducted with four different deep-learning models namely MobileNetv2, InceptionResNetV2, DenseNet201, and VGG16 with and without fine-tuning. From the available dataset 941 images were used for training, 105 for validation, and 262 for testing. Based on the experimental result, fine-tuned MobileNetv2 model achieved an accuracy of 92.38% for class labels, a COCO mAP of 42.84% indicating high detection performance, and the lowest inference time of 0.067567 seconds. Future researchers can focus on integrating this fine-tuned model with a target device (drone) and collecting overlapping images of Teff crop and weed leaves to train the model for application in a real commercial environment.