.png){kind=icon}
.png){kind=icon}
Return
A semi-supervised deep learning framework based on multi-source data for wheat: from counting to yield estimation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.isprsjprs.2026.03.026.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Efficient quantification of wheat ears and spike grains is crucial for yield estimation but is currently hindered by the heavy reliance on labor-intensive manual annotations and complex field environments. There is also no unified approach for the simultaneous segmentation and counting of wheat ears and spike grains. To address these issues, this study proposed a semi-supervised deep learning framework for wheat. First, we constructed a Wheat Segment Anything Model (Wheat-SAM), which integrates adapters and is fine-tuned on a large-scale dataset of wheat ear and spike grain images to generate high-quality training samples. Subsequently, we developed the Wheat Segmentation and Counting U-Net (WSCU-Net), an innovative dual-task framework that combines Depthwise Separable Convolutions (DSConv) and Efficient Multi-scale Attention (EMA), enabling simultaneous segmentation and counting of wheat ears and spike grains. Specifically, we fine-tuned Wheat-SAM to generate pseudo labels, which were subsequently used to train WSCU-Net. The trained WSCU-Net was then employed to predict segmentation masks and count outputs. Experimental results demonstrate that the fine-tuned Wheat-SAM achieved the highest segmentation accuracy across multiple scales and maintained stable performance under different prompt modes, confirming its capability in generating high-quality pseudo annotations. Furthermore, the semi-supervised WSCU-Net trained on pseudo labels outperformed other dual-stage fusion methods in both segmentation and counting tasks, achieving an IoU of 0.8409 and a Coefficient of Determination (R2) of 0.766, approaching the performance of fully supervised models. Notably, WSCU-Net maintained robust performance across different wheat cultivars, highlighting its applicability in diverse field scenarios. Preliminary correlation analysis between the predicted count results and actual yield showed a strong positive statistical relationship, with an average Pearson correlation coefficient (r) of 0.6918. Our approach provides valuable insights into leveraging foundation models in wheat phenotyping tasks and holds great promises for intelligent wheat yield estimation. It offers strong support for wheat breeding, remote sensing yield prediction, and other smart agriculture applications.
Keywords:
Wheat segmentation
Wheat counting
Semi-supervised learning
Deep learning
Yield estimation
Journal
IF:
12.2
Papers:
4.3K
Citations:
3.2W
Organization
View more organizations