This paper proposes an artificial intelligence-based yield prediction system (AIYPS) using environmental change detection for smart farms to compose agricultural environments, to predict the yield of crops, to detect environmental changes of farms, to facilitate the management of smart farms, and to reduce the causes of rapid changes in the production rate of the crops being grown, which consists of four modules as follows.
First, the Farm Environment Management Module optimizes the farm environment by collecting weather and soil data so that crops can grow well. It uses environmental data as input and utilizes farm management data such as nutrient supply and moisture replenishment as output. Second, the Farm Disease Diagnosis Module takes pictures of leaves of crops being cultivated by using a drone and diagnoses which diseases the crops have based on CNN. Third, the Farm Yield Prediction Module based on environments predicts the annual yield by using the result data of the Farm Environment Management Module and the Farm Disease Diagnosis Module, the current farm environment, and crop diseases as inputs to the ANN. Fourth, the Harvest Anomaly Detection Module analyzes the production process of crops by applying the condition of crops to Nelson rules after the actual harvest is over. It determines the overall condition and uniformity of the crops and manages the farm afterwards by using only some of the Nelson rules that can be applied to the farm.
As a result of the performance evaluation of the proposed AIYPS, the average accuracy of the Farm Environment Management Module was about 15% higher than that of RF and about 8% higher than that of GBT, and even if the data was increased, the accuracy was reduced less than that of RF or GBT. The Farm Disease Diagnosis Module is about 37 seconds faster than R-CNN and about 72 seconds faster than YOLO in runtime. According to the linear regression, the operation time slope of the Farm Disease Diagnosis Module was 8.1032E-3, R-CNN was 1.5642E0, and YOLO was 1.0934E0. And when the Farm Yield Prediction Module used ReLU as the activation function, the average accuracy was 97.8332%, when Sigmoid was used, the accuracy was 96.2454%, and when the step function was used, the accuracy was 92.759%. Finally, the predicted yield of the farm to which the Harvest Anomaly Detection Module was applied was 14.16% for apples, 14.58% for onions, 26.46% for tomatoes, 28.09% for potatoes, and 28.52% for cabbages. In particular, the production of potatoes and onions, which are root plants, was predicted to increase by about 30% when the Harvest Anomaly Detection Module was applied. Therefore, the AIYPS can manage more efficiently than existing farms from the start to the end of agriculture.