This study addresses the issue of black-grass, a herbicide-resistant weed that threatens wheat yields in Western Europe, through the use of high- resolution Unmanned Aerial Vehicles (UAVs) and synthetic data augmentation in precision agriculture. We mitigate challenges such as the need for large labeled datasets and environmental variability by employing synthetic data augmentations in training a Mask R-CNN model. Using a minimal dataset of 43 black-grass and 12 wheat field images, we achieved a 37% increase in Area Under the Curve (AUC) over the non-augmented baseline, with scaling as the most effective augmentation. The best model attained a recall of 53% at a precision of 64%, offering a promising approach for future precision agriculture applications. © NLDL 2024. All rights reserved.

This paper presents a crucial multispectral image registration and sensor calibration method for an agricultural application. The multispectral images are obtained using a special drone equipped with multiple cameras flying at low altitudes. However, the distance between lenses, the lens distortions and the low-altitude flights lead to a lack of alignment in the built-in normalized difference vegetation index (NDVI). This lack of alignment results in a very poor performance in further analysis, especially for image segmentation and target detection to distinguish crops from invasive plants. In this work, we point out the importance of reducing this misalignment. To do so, the near-infrared and red sensors are first calibrated to remove the lens distortions. Then, the corresponding keypoints are utilized to calculate the transformation matrix and to minimize the back-projection error. The registered near-infrared and red images are then used to compute NDVI. The experimental results show higher alignment and F1-score of 0.73 which is a significant improvement in the performance of a trained deep neural network using NDVI in the detection of invasive plants. This is particularly a challenging task as the invasive plants resemble the desired crops.

This work investigates the impact of the neural networks architecture when performing fingerprint recognition. Three networks are studied; a Triplet network and two Siamese networks. They are evaluated on datasets with specified amounts of relative translation between fingerprints. The results show that the Siamese model based on contrastive loss performed best in all evaluated metrics. Moreover, the results indicate that the network with a categorical scheme performed inferior to the other models, especially in recognizing images with high confidence. The Equal Error Rate (EER) of the best model ranged between 4%−11% which was on average 6.5 percentage points lower than the categorical schemed model. When increasing the translation between images, the networks were predominantly affected once the translation reached a fourth of the image. Our work concludes that architectures designed to cluster data have an advantage when designing an authentication system based on neural networks.