Optimization of irrigation timing for sprinkler irrigation system using convolutional neural network-based mobile application for sustainable agriculture

Abstract

The conventional sprinkler irrigation time calculation does not consider current soil moisture, leading to inadequate or excess irrigation due to daily weather changes. In present study, a new approach for calculating sprinkler irrigation time was proposed, which used in-field soil moisture values obtained by a convolutional neural network (CNN). To enhance accuracy and limit its size, the CNN architecture incorporated depth-wise separable convolution and residual connections. A deep-learning-based mobile application was developed, which predicted soil moisture class via in-field soil images, crop factors, and sprinkler system details. The system was assessed based on soil moisture class prediction accuracy, energy usage, water savings, and water productivity. The CNN model had an average classification accuracy of 97.10%, precision of 85.50%, recall of 86.80%, F1-score of 85.80%, and true prediction score of 75.30%. The developed mobile application saved 27.59% water and 27.42% energy by accurately estimating irrigation time using the predicted soil moisture class. The developed system corrected conventional irrigation depth in some instances, thereby reducing crop yield losses due to inadequate water supply. Compared to conventional sprinkler irrigation, the developed system increased water productivity by 32.75%. Thus, the system has the potential to conserve water and energy while improving crop yield.