Practical Weed Detection Based On Data Fusion Techniques In Precision Agriculture

Abstract

Artificial intelligence (AI) technologies have seen interesting applications in the agricultural sectors over the last few decades. Researchers have demonstrated that DL-based solutions are more accurate than previous classical techniques. Currently, weed control has received increased attention due to its potential to increase agricultural productivity. The weeds occlusion, variation in weather conditions, complex scenes background, and failures to transfer learning opened up a new world for researchers in this domain. This work aims to develop a fusion technique based on transfer learning to identify and localize more than 20 types of weeds in Morocco. Our strategy uses a data merging technique to address the needs of the actual agricultural field. In this regard, the you only look once (YOLOv5) framework was trained on a dynamic dataset with interventions, which results in a reasonable mean average precision (mAP) exceeding 96.38% and a precision of 97.95% with a recall value of 93.25% on the validation set. Therefore, the selection of the dataset demonstrates the robustness of this research as it contains images collected in real agricultural environments with a specific range of standards, which robots can use to effectively control weeds in real environments.