The great contribution of artificial intelligence to science (and agriculture)

In 2013, faced with the advance of a strange form of rapid desiccation of the olive tree in Puglia, the late prof. John Martelli first he asked himself a fateful question, demonstrating his intuition and his ability as a scientist to connect disparate information that he had learned during his long career: what if it was Xylella? Together with other experts, such as Maria Saponari and Donato Boscia, he had to suffer attacks and denigrations of all sorts, up to threats and investigations; but they were rightand the imbeciles who attacked and in some cases still attack or ride absurd conspiracy theories, even without ever apologizing and demonstrating their human and cognitive level in this too, have nevertheless ended up in oblivion, where they hide to avoid reminding all of having brought with their memes the main aid to the genes of the bacterium, which has transformed large areas of the magnificent olive groves of Salento into a cemetery.

Now, a new work, as well as reminding me of the story, raises an interesting question: what would happen if the first, fundamental early diagnosis of a phytopathology could be depersonalised, i.e. it could be attributed not to a person, but to a machine which, with an artificial intelligence system, could provide a first probabilistic indication to be further investigated later? In addition to the fact that it would be more difficult for researchers to suspect a conspiracy, there would be several advantages in combining this tool with human experts: the possibility of more widespread surveillance, to begin with, given that the sensors can be multiplied more easily than the heads of the Martelli, the Boscia and the Saponari, and secondly also the possibility of revealing signals not immediately perceptible to the human senses, for example by examining chemical signals and the appearance of plants at frequencies not visible. The work I mentioned, which has just been published, is based on the use of a new type of neural network applied to the extraction of data from images of controlled plants. In particular, theand neural networks, able to associate particular “patterns” in the data to specific diseases, have produced promising results in the classification of plant diseases. However, the conventional methods used so far require prior training for better predictions, and thus are particularly vulnerable to lack of adequate training data. Furthermore, good data collection on which to train neural networks is possible in controlled environments, but not trivial in the real world. In the field, diseases may be rare or not readily observed, or, as in the case of Xylella, may be relatively poorly documented on a new host plant. For these reasons, up to now the automatic classifiers of plant diseases based on neural networks have proved to be limited in their usefulness when used on real data, not included in the set used for training.

Now, a group of researchers, authors of the cited work, has developed a new relatively simple neural network called “Multi-Representation Subdomain Adaptation Network with Uncertainty Regularization for Cross-Species Plant Disease Classification” (Msun), which apparently can accurately classify plant diseases in natural environments. To do this, they applied a technique called unsupervised domain adaptation (UDA), which allows models learned by the AI ​​during lab training to be adapted to the situation found in the field, without the need for human supervision. This technique made it possible to overcome the difficulties related to the complexity of the images collected in the field, such as the presence of many leaves, unusual camera angles, blurring and other confounders related to photographic acquisition. Furthermore, the technique proved to be robust to the simultaneous presence of multiple pathologies on the same plant, documented in a single image, and also to the relative similarity of symptoms produced by different pathogens. In the end, using huge datasets of phytopathological images collected in the field all over the world, the research group authoring the work in question was able to demonstrate how the accuracy and precision of diagnosis obtained was comparable to that achieved during training on laboratory images, thus surpassing any other automatic diagnosis method available up to now. If the work of these scientists were to prove valid and confirmed by the analysis of other independent groups, very interesting scenarios would open up, some of which immediately come to mind.

First, the same type of technique could be used for the analysis of images and other types other than simple optical images, increasing the amount of information available for more accurate discrimination. Furthermore, instead of limiting itself to the analysis of the leaves of single plants, one could try to use the same method by applying it to the processing of multispectral images obtained from drones or airplanes, at least for certain particular conditions of stress and pathology detectable on vast cultivated areas, such as orchards, olive groves, vineyards and other types of cultivated areas. Collectively, help from AI could provide plant pathologists with the data-driven force that, when under human oversight, has been in other fields that can vastly improve our ability to respond promptly, accurately, and effectively.; and perhaps, beyond the fact that the hot-headed conspiracy theorists will in any case imagine new deceptions against them, it may be that the next environmental catastrophe linked to some pathogen or simple climatic variables can be better tackled, thanks to precision agriculture methods supported by modern research in artificial intelligence.