Comparison of SIFT Encoded and Deep Learning Features for the Classification and Detection of Esca Disease in Bordeaux Vineyards

Abstract

Grapevine wood fungal diseases such as esca are among the biggest threats in vineyards nowadays. The lack of very efficient preventive (best results using commercial products report 20% efficiency) and curative means induces huge economic losses. The study presented in this paper is centered around the in-field detection of foliar esca symptoms during summer, exhibiting a typical "striped" pattern. Indeed, in-field disease detection has shown great potential for commercial applications and has been successfully used for other agricultural needs such as yield estimation. Differentiation with foliar symptoms caused by other diseases or abiotic stresses was also considered. Two vineyards from the Bordeaux region (France, Aquitaine) were chosen as the basis for the experiment. Pictures of diseased and healthy vine plants were acquired during summer 2017 and labeled at the leaf scale, resulting in a patch database of around 6000 images (224 × 224 pixels) divided into red cultivar and white cultivar samples. Then, we tackled the classification part of the problem comparing state-of-the-art SIFT encoding and pre-trained deep learning feature extractors for the classification of database patches. In the best case, 91% overall accuracy was obtained using deep features extracted from MobileNet network trained on ImageNet database, demonstrating the efficiency of simple transfer learning approaches without the need to design an ad-hoc specific feature extractor. The third part aimed at disease detection (using bounding boxes) within full plant images. For this purpose, we integrated the deep learning base network within a "one-step" detection network (RetinaNet), allowing us to perform detection queries in real time (approximately six frames per second on GPU). Recall/Precision (RP) and Average Precision (AP) metrics then allowed us to evaluate the performance of the network on a 91-image (plants) validation database. Overall, 90% precision for a 40% recall was obtained while best esca AP was about 70%. Good correlation between annotated and detected symptomatic surface per plant was also obtained, meaning slightly symptomatic plants can be efficiently separated from severely attacked plants.

Evaluating the Suitability of Hyper- and Multispectral Imaging to Detect Foliar Symptoms of the Grapevine Trunk Disease Esca in Vineyards

“…This model was trained on heterogeneous field data and can therefore be considered more robust as well as better adjusted to mixed leaf samples than the model trained on optimized annotated data. A comparable approach using RGB-images was performed by Rançon et al [30] who, at first, tested the feasibility of Esca symptom detection on selected leaves, thereby, achieving satisfying overall accuracies of 88 and 91 % for white and red cultivars, respectively. In a next step, they evaluated disease detection on plant-scale and found it to be more complex due to varying symptom intensity and overlapping effects of leaves.…”

“…While Junges et al [28] measured and calculated the chlorophyll index per leaf to discriminate different symptom intensities; Al-Saddik et al [27] combined the spectral data with textural data gained from RGB images. Further analyses were performed by Gallo et al [29] and Rançon et al [30] both using platform prototypes for the in-field detection of Esca leaf symptoms. Two platform prototypes developed by Gallo et al [29] were each equipped with multispectral cameras calculating the Normalized Difference Vegetation Index (NDVI) for disease detection and one vehicle was additionally equipped with LiDAR sensors to gain information on canopy thickness.…”

“…Two platform prototypes developed by Gallo et al [29] were each equipped with multispectral cameras calculating the Normalized Difference Vegetation Index (NDVI) for disease detection and one vehicle was additionally equipped with LiDAR sensors to gain information on canopy thickness. Rançon et al [30] compared different algorithms and methods for a classification both on leaf and plant scale on the basis of digital images. In an airborne approach Kerckech et al [31] also worked with RGB images combining deep learning with different color spaces and vegetation indices for vine disease detection in manually annotated data.…”

An Enhanced Plant Disease Classifier Model Based on Deep Learning Techniques

“…In addition, the precision is 99.11% whereas the recall and F1-score are 99.49% and 99.29%, respectively. Because of efficient parameter selection during transfer learning, relying on new data augmentation methods, and using fixed number of images in each category, the proposed model [8,9] use a small number images, reducing variability in the dataset. As observed in Table 4, over the last year years, DL techniques have shown a remarkable improvement for plant disease detection as compared to traditional approaches [8,9] such as SIFT, HoG, and SURF because such methods lack the ability of transfer learning, which is used by DL models.…”

In Field Detection of Downy Mildew Symptoms with Proximal Colour Imaging

“…Recent studies proposed more operable strategies for the automatic detection of grapevine diseases. The authors of [17] proposed comparing methods based on Scale-Invariant Feature Transform (SIFT) encoding and deep learning strategies for the real-time detection of Esca and Flavescence dorée diseases. The authors of [18] proposed a deep learning training strategy for the detection of downy mildew in real conditions and show the difficulties of building a robust and replicable model and the requirements in accurate annotations of the database.…”