Abstract
Farmland abandonment has become a critical issue in China over the last two decades, posing significant challenges to food security and land-use planning. Traditional field investigations are inefficient for large-scale monitoring, while standard pixel-based remote sensing techniques often yield poor results due to the heterogeneous spectral signatures of abandoned land. This paper evaluates an object-oriented classification (OOC) technique using SPOT-5 satellite imagery. By grouping pixels into meaningful objects based on color, size, and shape, the OOC method achieved an overall accuracy of 91.43% and a Kappa coefficient of 0.8721, significantly outperforming traditional pixel-based methods.
Introduction
The rapid urbanization of China has led to a significant increase in abandoned farmlands. Monitoring these changes is vital for agricultural policy, yet abandoned land lacks a uniform spectral "signature," often appearing as a messy mix of weeds, bare soil, and water. In traditional pixel-based classification, each pixel is analyzed in isolation, leading to high "noise" and low accuracy because the spectral characteristics of abandoned fields often overlap with other land cover types like woodland or active vegetable plots.
Methodology
The study focused on the Pingqiao District of Xinyang City, Henan Province, a major agricultural region. The researchers utilized SPOT-5 satellite imagery captured on November 8, 2010, which provided both high-resolution panchromatic and multispectral data.
The study compared two primary approaches:
- Pixel-based Supervised Classification: Six standard algorithms were tested, including Maximum Likelihood, Mahalanobis Distance, and Spectral Angle Mapper. These methods rely solely on the spectral mean of training areas to assign classes to individual pixels.
- Object-oriented Classification (OOC): This approach involves an initial segmentation step to group pixels into homogeneous objects, followed by classification based on a combination of spectral signatures, texture, shape, and spatial context.
Results and Discussion
The quantitative analysis revealed a substantial performance gap between the two methods. The OOC approach reached an overall accuracy of 91.43%, compared to 82.34% for the best-performing pixel-based method (Maximum Likelihood). For the specific task of identifying abandoned farmland, OOC achieved 78.94% accuracy, whereas pixel-based methods were significantly less reliable, with some dropping as low as 18% for this specific category.
Qualitatively, pixel-based maps suffered from heavy "salt-and-pepper" noise. Because OOC mimics human logic—noticing boundaries and spatial relationships—it produced maps with cleaner boundaries and results that were much closer to reality. The core insight of the study is that abandoned land is better defined by its structure and texture than by its color alone.
Conclusion
Object-oriented classification provides a more accurate and scalable solution for monitoring complex land-use changes like farmland abandonment. While the 78.94% accuracy for abandoned land is a major improvement, future work will explore integrating multi-sensor data and ancillary GIS layers, such as Digital Elevation Models (DEM) and proximity to infrastructure, to further refine classification rules and improve detection in highly varied landscapes.
Summary updated