Abstract
Precision agriculture relies heavily on accurate plant segmentation to monitor growth and health. However, traditional clustering methods like $k$-means and Density Peaks Clustering (DPC) often struggle with the non-convex shapes of plants and the "chain reaction" errors caused by misclassified boundary points. This paper proposes an improved DPC method that utilizes Gaussian kernel mapping and a novel two-stage point assignment strategy. By projecting data into a high-dimensional feature space and employing a "temporary boundary point" to halt initial assignments, the method achieves superior segmentation accuracy on complex agricultural images and standard benchmarks.
Introduction
In agricultural image processing, segmenting early-stage crops from soil is challenging due to two primary factors. First, the color profiles of young plants and surrounding soil are often highly similar, rendering simple thresholding ineffective. Second, traditional DPC algorithms suffer from a "chain reaction" error: because points are assigned to the same cluster as their nearest neighbor of higher density, a single misclassification at a cluster boundary can trigger a cascade of errors, pulling unrelated data points into the wrong group.
Methodology
The proposed method addresses these limitations through a three-phase workflow: centroid selection, high-dimensional mapping, and a two-stage allocation strategy.
1. Centroid Selection and Kernel Mapping
Following the DPC paradigm, cluster centers are identified using a decision graph that plots local density ($\rho$) against the distance from points of higher density ($\delta$). To handle non-linearly separable data, the algorithm projects the input features into a high-dimensional space using a Radial Basis Function (RBF) Gaussian kernel. This transformation allows complex, interlocking shapes to be separated more effectively than in the original Euclidean space.
2. Two-Stage Allocation Strategy
To prevent the "chain reaction" error, the authors introduce a two-stage assignment process:
- First Stage: Starting from the identified centroids, the algorithm performs a similarity search in the high-dimensional space. This process continues until it reaches a temporary boundary point, defined by a similarity threshold ($d_q$). This ensures that only points with high confidence are clustered initially.
- Second Stage: Once the core of the clusters is established, the remaining unassigned points are allocated to the cluster with which they share the highest overall similarity, rather than just following a single nearest neighbor.
Experimental Results
The algorithm was validated against $k$-means, DPC, and Fuzzy c-means across synthetic, UCI, and agricultural datasets.
- Synthetic Datasets: The method achieved perfect scores (1.0) for Accuracy (ACC) and Normalized Mutual Information (NMI) on complex shapes like Spiral and Two Moons, where $k$-means failed due to its assumption of spherical clusters.
- UCI Benchmarks: On the Robot Navigation dataset, the proposed method reached an ACC of 0.7428, a significant improvement over the 0.3616 achieved by traditional DPC.
- Plant Segmentation: In practical applications using the Leaf Counting Dataset, the method demonstrated high robustness. It successfully outlined plant structures even when soil brightness varied significantly, avoiding the "muddy" edges and misclassifications common in other clustering approaches.
Conclusion
By combining high-dimensional kernel mapping with a robust two-stage allocation strategy, this improved DPC method effectively mitigates the cascading errors inherent in traditional density-based clustering. While the current implementation requires some manual parameter tuning for the decision graph, future work will focus on automating centroid selection to create a fully autonomous segmentation pipeline for precision agriculture.
Summary updated