Abstract

Changes in crops’ health may be gradual in nature, whereas traditional inspections on fields provide rare opportunities to observe it. The paper describes an approach for continuous monitoring and estimation of the probabilityof diseases using environmental sensors and crop images. The IoT (Internet of Things) sensors provide temperature, humidity, and soil moisture readings, whereas the computer-vision models process visible patterns in the crop/leaf images. Both data sets are analyzed independently, and their combination results into a disease-risk estimation based on all inputs. The approach contains explainability mechanisms in order to ensure that the reason of raising the alarm consists of the key observations that affect the prediction result. Recent literature indicates interest in IoT monitoring, lightweight machine learning (deep learning), multi-modal analytics, and edge computing. Explainable AI is increasingly considered a vital part of any AI system applied to agriculture. Nevertheless, issues of dissimilarity of lab-generated data and field environment, lack of diversity of input data, accuracy of sensors, computation costs, and potential for false alarms have to be taken into account. This paper focuses on development and evaluation of field-oriented research framework, in which prediction quality, time of issuing warning, consumption of resources, and usefulness of generated explanation are all evaluated jointly. The goal of the research is a simple decisionsupport tool which allows detecting crop health problems early, but still leaves the final decision up to the farmer or agricultural expert.

Keywords

How to Cite This Article

Rights