Improving Image Alignment in vineyard environment with deep learning image matching

Abstract

Abstract. Globalisation has accelerated the spread of invasive agricultural pests, including Popillia japonica Newman, introduced to Italy in 2014. This species has caused severe damage to vineyards, highlighting the need for efficient detection methods. Manual identification, though accurate, is time-consuming and labour-intensive. This study explores a computer vision (CV)-based approach using Near-Infrared (NIR) imagery captured by Uncrewed Aerial Systems (UAS) to detect adult Popillia specimens. Conducted in two vineyards in northern Italy, the project aims to develop a standardised and replicable monitoring protocol. CV-based detections are validated by entomologists and integrated into a Geographic Information System (GIS) to generate prescription maps for targeted drone-based pesticide application. However, traditional feature extraction and matching (FEM) algorithms, such as SIFT, SURF, and ORB, struggle in vineyard environments due to repetitive structures (seriality of fixed components, such as poles, supports, etc) and limited NIR texture. These limitations hinder image alignment, especially in the absence of geodetic-grade GNSS and high-precision IMU data. To address this, the study replaces FEM methods with deep image matching (DIM) techniques like SuperPoint and DISK for feature extraction, paired with SuperGlue for graph-based matching. Applied within a visual SLAM (vSLAM) framework, these deep learning models significantly improve image connectivity and alignment. Experimental results, supported by a fine-tuned SuperPoint model trained on vineyard datasets from the DANTE2 project, demonstrate up to 90% alignment improvement over conventional methods. This work presents a robust, scalable solution for accurate pest mapping in viticulture, contributing a fine-tuned PyTorch model to the scientific community.