Wine Disease Detection
Winegrowing in Baden and the Palatinate is an integral part of the collective identity of these regions. In the coming years and decades, the economic fate of the viniculture will be shaped by climate change and its consequences like that of hardly any other economic sector. This complex problem includes the spread of certain bacterial and fungal diseases in the vineyard as well as research into resistance to those diseases. Thus, we decided on a use case in the field of disease control in vineyards. We set ourselves the task of using computer vision to enable reliable detection of important vine diseases and implementing this on an edge computer, the NVIDIA Jetson Nano. In viticulture, there are a number of different diseases that affect the vine and are partly responsible for quality or quantity losses in the harvest. The amount of recognizable, relevant diseases can be limited to two essential ones: Peronospora (Pero) and Esca disease.
Our goal was to use object detection of diseases on grape leaves to better study disease outbreaks. This involves using cameras on drones to enable cost-effective large-scale surveys of grapevines at regular frequencies. The image data and the diseases detected from them enable the creation of a database and the mapping of the spread of diseases in different vineyards in the growing regions as seen in the prototype picture below.
We chose the Yolo (You only look once) approach, which is one of the most popular and promising approaches. The new versions regularly set new benchmarks in terms of object detection performance and inference time. We opted for the latest version of Yolo at the time, Yolov5 from ultralytics.
The training of the model was done in several iterations, in which we used the Mean Average Precision as well as Precision and Recall for individual classes. First, we enriched the training data set with our own images from the vineyard and images of typical backgrounds. We found that object detection on a video posed new challenges. Again, we added our own images to the training data to increase the robustness of the model, for example with regard to leaves of different ages. This was followed by an optimization of about 30 different hyperparameters and the use of test time augmentation. The resulting model had a mAP@0.5 of 0.765. The following pictures illustrate the result.
Overall, we can draw a positive conclusion regarding the development of the model. Strategies for improving performance were applied systematically. In doing so, we repeatedly thought about the meaningfulness of certain measures and thus made well-founded decisions. It was crucial that we acquired some understanding of viniculture in addition to expertise in object detection and Yolo. The composition of the training data was a constant part of our exchange with each other. Regular communication with the inclusion of various influencing factors proved to be useful and necessary in order to make the resulting model as real-world-suitable and robust as possible.
The application in the vineyard is already theoretically possible with our model. The Jetson can be mounted on a tractor, on which it can easily fit thanks to its size. As a result, infestations of Esca and Peronospora can be "casually" detected each time it is driven through a vineyard. Frequent use is therefore not associated with increased time expenditure. On the contrary, the development of the infestation can be precisely documented in very small-time intervals and conventional detection "on foot" can be made superfluous. However, due to its compact design and small size, the application of the Jetson is not limited to tractors. In many wine-growing regions, spraying agents are already applied with drones. Since drones are usually already equipped with cameras, it makes sense here to think about integrating the function of the Jetson with our model in such a drone. This would allow our model to be used in almost all wine-growing areas, regardless of accessibility by tractors.
The training of the model took place in early summer and summer. Certain diseases and older leaves could not yet be observed in the vineyard at this time, which influences the robustness of the model. The generation of more heterogeneous training data, especially videos, is therefore an opportunity for future improvements of the model.
In a next step, the construction of a database can be considered. In this database the results of many different companies of a region should be collected. The central calculation of the results is also suitable for this purpose. The detected disease outbreaks are linked to their coordinates, thus enabling the construction of a detailed, regularly updated map or heat map. This allows the progression over time and location to be tracked in detail and examined in conjunction with additional data, such as weather data.
In principle, the idea of optical disease detection can of course also be transferred to other agriculturally interesting crops. In addition to viniculture, many other crops are sprayed against diseases in Germany, and to a large extent. In this work, it was shown that this type of detection is possible in principle. Depending on the type of symptomatology, however, the detection can probably be transferred with it not only to other plants, but also to other diseases.