Per‐tree citrus count regression from single‐view images via lightweight deep learning

Fruit yield is a critical parameter in orchard management and market forecasting. However, achieving efficient, low‐cost, and low‐equipment fruit quantity estimation remains a significant challenge in real‐world agricultural production. Existing methods typically rely on multi‐view images or high‐performance computing resources, making them complex to deploy and expensive. To address these issues, this article proposes a lightweight approach to estimating the quantity of mandarin orange ( Citrus reticulata Blanco) based on single‐view images, designed for typical orchard environments under natural conditions. The method employs the EfficientNet‐B0 network as the backbone, incorporating label transformation and sample weighting strategies to improve regression accuracy and stability across different yield levels. Field‐collected image data are used in the experiments, and model performance is assessed through cross‐validation. Gradient‐weighted class activation mapping visualization and uncertainty analysis are conducted to examine the effects of environmental factors such as occlusion, lighting, and viewing angle on prediction performance. Experimental results demonstrate that the proposed method maintains robust predictive capability and deployment potential even in resource‐constrained and complex orchard settings, offering an efficient, cost‐effective, and scalable solution for fruit yield estimation.