Machine Learning Solutions for Precision Fruit Growth Classification and Yield Projections

Abstract

The agriculture function is critical in agrarian economies, and as population increases across the world, there is need for more innovation, efficient and accurate methods of producing foods. They subsequently reviewed that old fashioned procedures of farming are gradually being replaced with modern ways to feed the growing populations besides enhancing agricultural exports. Precision agriculture technology that employs modern technologies in the management of existing farm resources holds the central place in this transformation. This research seeks to contribute towards the development of efficient precision agriculture control by performing automatic segmentation of mature or immature citrus fruits and subsequently determining the fruit yield count. Using current detection models like CNN ResNet50, YOLOv8 classifiers or VoLt YOLOv10 coupled with DeepSort and Byte Tracker for tracking, the study aims at improving the classification of the citrus fruits. The dataset included 756 images from Changing and 694 from the National Agri Research Center (NARC) which were extended to 15, 220. To train the model, dataset was divided in 93:7 training-validation/testing ratio and tested under different conditions like different lighting, blurriness, and occlusions. Use of YOLOv10 X-Large deepsort along with byte tracker yielded better performances than models based on YOLOv8 in diverse environments with occlusions and low light conditions. In growth stage recognition, YOLOv8 models demonstrated significant improvements: YOLOv8 X-Large gained the improvement of the top-1 accuracy from 94. 37% to 97. 30% which has a significant decrease of the training loss as compared with the ResNet50 may have less effectiveness in the oranges growth classification near 32. 40 to 76. 90. Schedule 3: Comparison of various models in terms of accuracy and loss YOLOv8 Small and Nano models also received significant increase in accuracy and decline in loss; The Nano model achieved the accuracy rate of 0. 97. on social network BigML is 08% and now training loss is also minimum. In general, YOLOv8 models particularly the Nano versions outperformed ResNet50 in average inference time and similarity index, which qualified them for real-world usage across various agricultural environments. This research demonstrates that these precision agriculture technologies with remote sensing and artificial intelligence will open up other great developments in agriculture that can revolutionize food-systems around the world