Random Forest, XGBoost, CNN, or Transformer?

Remote-sensing papers can make model choice sound like a race toward the newest architecture, but orchard mapping is usually won or lost earlier, in the quality of the time series, the field labels, and the way the validation split is designed. A Random Forest can be the right answer when labels are scarce and the main question is which spectral or phenological features matter. XGBoost is useful when nonlinear interactions and probability estimates matter. A segmentation network becomes more attractive when boundaries and spatial context are central, while a recurrent model or Transformer earns its complexity when the sequence itself carries information that should not be flattened into a large stack of bands.

The research record supports a layered approach rather than a single winner. Hierarchical ConvRNN work showed that a model can first learn the broad category “orchard” and then use that representation to distinguish rarer classes such as apples and pears. More recent Transformer work uses Landsat and Sentinel-2 sequences for timely, within-season crop mapping, while the 2025 apple study combined temporal feature selection with SegNet, CBAM, and Focal Loss to improve segmentation. Those methods answer different questions, so comparing them only by one headline accuracy can be misleading. See the hierarchical time-series study, the 2025 Transformer study, and the apple orchard study.

For a real project, the most defensible sequence is to build a strong feature-based baseline, add a spatial model only when the map shows a boundary problem, and add a temporal deep model only when the baseline cannot represent the sequence adequately. Every improvement should be tested with the same field-level split, the same held-out year or region when possible, and an ablation that shows whether the gain came from the architecture, the extra bands, the temporal window, or the labels. Complexity is justified when it makes the result more portable or more useful, not when it merely makes the method harder to explain.