This study was focused on forecasting diseases in fruit trees five days in advance using supervised machine learning. This involved photographing the undersides of leaves and sensing environmental conditions using agricultural robots. A leaf underside disease classifier that achieved 0.90 accuracy and 0.91 recall based on 330 images collected by the robot-mounted camera was developed. The classifier’s results were utilized for binary classifications to predict disease occurrences. This innovative approach aims to enhance disease management in agriculture. Using objective variables in the leaf underside disease classification and feature-increase method, we analyzed disease forecasting methods through the comparison of machine learning models, sensor types, and dataset durations required for training the models. As a result, we clarified the changes in the accuracy of the predicted number of days for each machine learning model. The recall when using the dataset collected by the robot over 16 days was 0.980. Furthermore, we confirmed that the characteristics unique to each farm appeared in the forecast for each sensor used in the observations.