Pixel-level crop segmentation is essential for precision agriculture yet remains severely constrained by the prohibitive cost and labor intensity of generating large-scale annotated datasets, a critical bottleneck limiting autonomous crop monitoring adoption across agricultural systems. This study presents a paradigm-shifting, zero-shot segmentation framework that eliminates pixel-level annotation requirements entirely by leveraging monocular depth estimation models (Depth Anything, Depth Pro, DPT Large) combined with unsupervised clustering algorithms to segment crops based on 3D structural geometry. Evaluated on a comprehensive agricultural dataset of 19,940 images spanning potatoes, kidney beans, and soybeans across distinct agro-climatic regions of Prince Edward Island and Ontario, Canada, the proposed depth-aware pipeline substantially outperformed classical segmentation baselines and competed favorably with emerging foundation model approaches. The method achieved a mean F1-score of 0.875 and Intersection over Union (IoU) of 0.803, representing 93% to 164% improvement over traditional techniques (Otsu thresholding, Watershed, K-means), while maintaining superior consistency across crop types (standard deviation 0.057 versus 0.127 for vision-language models). Notably, Depth Anything combined with unsupervised Mini-Batch K-means clustering achieved this performance with an inference latency of 400 ms per image, making it practical for real-time field deployment. The framework's robustness stems from its utilization of geometric priors and inherent depth discontinuities, which differentiate crop vegetation from soil backgrounds under complex conditions where traditional color-based and even zero-shot vision-language models struggle. By circumventing the data scarcity trap that has plagued supervised deep learning in agricultural applications, this work establishes 3D structural reasoning as a scalable, highly accurate, and annotation-free solution for autonomous crop monitoring, directly advancing the accessibility of AI-driven precision agriculture for small and large-scale farming operations worldwide. • Introduces a novel zero-shot depth vision framework for crop segmentation. • RGB to 3D depth estimation using depth vision foundation models • Provides a scalable, annotation-free solution for crop segmentation • Replaces the labor-intensive annotated supervised datasets • Achieves high accuracy across multiple crops and regions.
Afzaal et al. (Fri,) studied this question.