multispectral images of plants

The automatic analysis of the images with algorithms of the ML/DL type (machine learning/deep learning) allows a good ability to recognize plants with some type of suffering. As described in deliverable D421 the accuracy of the various models is around 80%, and similarly the F1-score for diseased plants. In particular this means that the models have few false negatives (i.e. plants with some problems indicate as healthy). False negatives are evidently the worst cases from a farm management point of view, while false positives (plants without problems indicate as unhealthy) are more, but also less serious from a farm management point of view. It should also be noted that the algorithms work on sub-images of plants (for example the image of a plant is divided into 9 smaller squares), and signal that a plant has a problem if at least one sub-image signals a problem. This allows for a more precise analysis, and corresponds to the way of working of people who monitor plants. In most cases, a plant has a problem in only one part, and the plant manager intervenes as soon as possible to prevent the problem from spreading. Image analysis outputs a binary value (healthy / unhealthy). It was not possible to obtain (particularly for the value 'unhealthy') a better analysis capability. In fact, the 'unhealthy' value corresponds to tens, if not hundreds of possible cases (attacks by parasites, insects, viruses, pollutants, water shortages). The detailed level of analysis, capable of identifying the specific type of unhealthy case, requires the availability of an extremely higher number of images and corresponding pathophysiological analyses.