MACHINE LEARNING BASED REAL-TIME DETECTION OF RIPE AND UNRIPE SORGHUM
The physiological maturity, which is a conventional way of detecting the ripeness of Sorghum involves moisture detection and physical sampling of the crop. It involves plucking out a few grains from the panicle and check the point of attachment, if the tip is brownish-black, it signifies that the grain is mature. However, this method of ripeness detection is destructive, time-consuming, labor intensive particularly in large farms and inaccurate. Another important issue that causes a reduction in yields and economic losses is delayed harvesting, which can result in grain loss due to pest infestation, shatter, and sprouting. Sprouting becomes an issue when sorghum grain has reached maturity and is exposed to long periods of wet, warm weather. This work proposes to develop a machine learning (ML) based system to detect the ripeness of Sorghum and alerts the farmer to harvest at the appropriate time. It encompasses the training of a machine learning model to effectively learn patterns within a given dataset. The dataset was annotated using Roboflow and divided into training, validation and test set. The trained model is deployed onto a Raspberry pi using Python code, allowing the model to operate directly on the device to execute the model's predictions locally enabling it to make accurate predictions of the state of the crop. Finally, it Utilizes a camera sensor on the Raspberry device, to perform real-time inference with the trained model. YOLOv5 object detector model is used. Precision, Recall, F1 score and mean average precision (mAP) are the evaluation metrics used to gauge the performance of the model. The trained model is successfully deployed on the raspberry pi to achieve a real-time detection on the field. The real-time system exhibits an overall Precision of 79.5%, a Recall of 87.3%, an F1 score of 83%, and an mAP of 87% using the YOLOv5 object detector model. For future work, expanding the size of the dataset and using a more robust model would improve the accuracy and efficiency of detecting the ripeness of the crop. The model could be further integrated into robotic systems and drones for automatic harvest and to compute the yield percentage.
Authors : udu, A.I. and Ndirmbula, S.M.
Category : Open Access Volume (Issue) : 10(2) Date Uploaded : 29th October 2024