Data-driven robotic systems are imperative in precision agriculture. Currently, Agri-Robot precision sprayers lack automated methods to assess the efficacy of their spraying. In this paper, images were collected from an RGB camera mounted to an Agri-robot system to locate spray deposits on target weeds or non-target lettuces. We propose an explainable deep learning pipeline to classify and localise spray deposits without using existing manual agricultural methods. We implement a novel stratification and sampling methodology to improve classification results. Spray deposits are identified with over 90% Area Under the Receiver Operating Characteristic and over 50% Intersection over Union for a Weakly Supervised Object Localisation task. This approach utilises near real-time architectures and methods to achieve inference for both classification and localisation in 0.062 s on average.
|Title of host publication||Towards Autonomous Robotic Systems - 24th Annual Conference, TAROS 2023, Proceedings|
|Editors||Fumiya Iida, Perla Maiolino, Arsen Abdulali, Mingfeng Wang|
|Number of pages||12|
|Publication status||Published - 8 Sep 2023|
- Computer Vision