Reproducible Spectral-Spatial Vegetation Stratification from a Single Jilin-1 Scene under Controlled Validation

Very-high-resolution (VHR) satellite imagery may support individual palm delineation and field-calibrated assessment of vegetation condition, but such programs require costly multi-farm imagery, tree-level annotations, and field measurements. This study asks whether a single Jilin-1 scene, acquired on 6 October 2024, contains enough reproducible spectral-spatial structure to justify that investment. The scene serves only as a technical test bed because palm occurrence within it was not verified. An exploratory Stage 1 pipeline (Brovey preprocessing, four spectral indices, three-cluster MiniBatch K-Means pseudo-labels, and a U-Net) reached 0.9806 agreement with its own labels, but redundant features, whole-scene clustering, overlap between clustering features and model inputs, and reuse of validation data limited interpretation. A controlled Stage 2 removed an exact duplicate (green-NIR normalized contrast) and an algebraic duplicate (ratio vegetation index), clustered on NDVI and GNDVI, and selected two clusters from a five-seed analysis (silhouette 0.59472; pairwise adjusted Rand index 0.99034). Training, validation, and held-out test regions were separated by guard patches, and the scaler and clustering model were fitted on training data only. In the held-out region, a bands-only Logistic Regression reproduced the two-stratum reference with 0.99401 agreement, compared with 0.99192 ± 0.00153 for a bands-only U-Net over three seeds. The U-Net did not exceed Logistic Regression in agreement, although its macro-F1 and mean intersection over union were higher in point estimate. Supplying NDVI and GNDVI to the U-Net improved every metric, which provides direct evidence of target-feature overlap. The scene therefore contains strong, reproducible, scene-relative spectral structure that raw bands recover in a geographically separated region, and a deep network is not required for high pseudo-label reproduction. Because Brovey fusion degraded reduced-resolution radiometric fidelity, these results are conditional on that preprocessing choice. The strata are not validated indicators of moisture, stress, disease, or palm condition, and a multi-farm, multi-date study with palm annotations and field measurements remains necessary.