Management of Geospatial Data with an Interface for the Application of Artificial Intelligence

Abstract

The combination of digital images with artificial intelligence algorithms has several potential applications in agriculture. However, the complexity of the agricultural environment means that the information contained in such images is not enough to provide an assertive answer that actually helps in decisions. With the evolution of the use of sensors (soil, humidity, leaf wetness, etc.), as well as the increase in the reliability of historical records of meteorological and phytosanitary variables, different sources of information covering various aspects of the crop or herd can be explored together with the images for a better characterization of what happens on the property. This is particularly important for small and medium-sized producers, as it becomes possible to effectively monitor the property at relatively low cost.The construction of adequate databases is relatively simple in the case of meteorological and environmental sensors, just ensuring the continuous operation of such sensors over a representative period and adequate storage of the generated data. In the case of images, the challenge is greater, with the lack of truly representative databases being the biggest obstacle to the practical adoption of this type of technology. The literature has many works proving the predictive power of deep neural networks trained with digital images, but in all cases the image base used in the research is still limited for such tools to have practical use. As an important element in this process, machine learning can be seen as a different programming paradigm, in which computational procedures are learned directly from input data, rather than being explicitly coded by a programmer. It was only with the use of machine learning that perceptual tasks such as image recognition and image segmentation could be properly modeled to deal with the enormous variability found in the input data. Deep neural networks are examples of the success of this paradigm, not only for their good performance in perceptual tasks, but also for the possibility of end-to-end learning, in which a single model (the neural network) performs all the processing. Thus, one of the great challenges of the CPAD will be the generation of representative databases, which will demand not only a great effort in the field to capture the images, but also an effective involvement of a specialist capable of correctly labeling the large volume of data that will be collected. generated and to propose new data collection strategies using the concepts of the area. The data science specialist, foreseen for this Postdoctoral Fellowship, will develop data management strategies - involving collection, preparation and combination of variables of interest to feed machine learning algorithms. (AU)