Instance Segmentation and Yield Estimation for Orchards Based on Enhanced Deep Learning and Multi-Scale Feature Fusion

Abstract

Quickly and accurately calculating orchard yield to address long-standing issues of fruit density, frequent occlusion, and various lighting conditions. This paper will construct an instance segmentation framework to study various orchard environments. This new technology can be used to improve segmentation accuracy and prediction results. It can be used for context-aware data augmentation, soft label smoothing, adaptive loss reweighting, and multi-scale feature fusion. In the experiment, various orchard environments and five different fruit varieties were selected. The evaluation used image-based annotations and field-recorded harvest data. The optimized framework achieved an average Intersection over Union (IoU) of 0.86 and an F1-score of 0.89 on the test set. Under commercial conditions, the average absolute error of yield estimation is 7.4–8.8 kg per block. Surpassed the historical peak and reduced the relative yield estimation error by 32%. Improving computational efficiency will aid in the real-time deployment of mobile and fixed agricultural platforms. The system proposed here enhances the technology supporting orchard automation management and provides significant benefits for large-scale precision agriculture. It also provides new directions for future crop monitoring and autonomous yield prediction.