Unsupervised Segmentation Of Smallholder Fields In Mozambique Using PlanetScope Imagery.pdf

This study investigates unsupervised segmentation techniques for mapping smallholder fields in northern Mozambique, where farms averaging less than 2 hectares are prevalent. Using PlanetScope satellite imagery with 3.7-meter spatial resolution, researchers compared three methods: mean shift, multiresolution segmentation, and Simple Non-linear Iterative Clustering (SNIC).

Key Findings:

- Evaluation Metrics: The study employed four supervised metrics—Area Fit Index (AFI), Quality Rate (QR), Oversegmentation (OS), and Undersegmentation (US)—to assess segmentation accuracy.

- Outperforming Methods: Multiresolution segmentation demonstrated the best performance in delineating smallholder fields, followed by mean shift and then SNIC.

- Relevance: Accurate mapping of smallholder farms is crucial for understanding rural dynamics, environmental management, and poverty alleviation efforts in Mozambique and similar regions globally.

Methods:

- Mean Shift: A non-parametric algorithm grouping pixels based on spatial and attribute space convergence.

- Multiresolution Segmentation: A region-based approach merging smaller objects into larger segments while optimizing internal heterogeneity at different scales.

- SNIC (Simple Non-linear Iterative Clustering): Built upon SLIC, SNIC uses a grid to define clusters and calculates distances between pixels and cluster centers considering both spatial and color factors.

Conclusion:

The study highlights the effectiveness of multiresolution segmentation for identifying smallholder fields from PlanetScope imagery. This finding paves the way for more comprehensive regional-scale mapping efforts, contributing to better understanding of smallholder dynamics in Mozambique and similar agricultural regions.