Pixel, Object, or Parcel?

The spatial unit of an orchard map is not a technical detail that can be postponed until the end. It determines what the map means. A pixel map answers which locations look similar to the training data, an object map answers which segmented shapes share a pattern, and a parcel map answers which field should be treated as one operational unit. Those questions overlap, but they are not interchangeable, especially when a ten-metre pixel contains tree canopy, grass, a road edge, and bare ground at the same time.

Pixel-based classification is often the easiest place to start because it uses the data directly and can preserve fine variation. Its weakness is the familiar salt-and-pepper effect, in which a single orchard becomes a patchwork of labels. Object-based methods reduce that noise by grouping neighbouring pixels before extracting spectral, temporal, textural, or radar features. Research in arid fruit-growing regions found strong results from multi-scale segmentation combined with Random Forest, while a 2025 study in Linyi found that the pixel workflow produced a slightly higher overall accuracy than the object workflow even though object-based results gave more coherent boundaries. These findings point to a useful distinction: the model with the highest OA is not automatically the model that produces the most useful field map. See the Tarim Basin study and the Linyi study.

For field operations, the parcel is usually the right place to make the final decision, but the pixel and object layers should not disappear. A parcel-level result can report the proportion of usable pixels, the area showing a change, the strength of that change, and the uncertainty around the estimate. The map can then point to a priority zone without pretending that every pixel inside the boundary has the same condition. This layered approach keeps the analysis faithful to the imagery while making the result easier for a crew to act on.