Data and Code for: Decoupling the Impact of Spatial Resolution and Landscape Heterogeneity on Crop Classification via a Spatial-Spectral Simulation Framework

Accurate spatial information on crop type distribution is critical for precision agriculture and regional food security. However, in smallholder farming regions characterized by fragmented landscapes, the applicability of medium-resolution satellite data (e.g., Landsat, Sentinel-2) remains controversial due to pervasive mixed pixels. Existing comparative studies often rely on operational multi-source imagery, where the pure impact of spatial resolution is deeply confounded by environmental noise (e.g., atmospheric residuals, phenological shifts), limiting a mechanistic understanding of area estimation errors. To address this, we developed a spatial-spectral decoupling simulation framework based on high-resolution GF-6 (2 m) imagery to rigorously isolate the spatial resolution effect. Across a gradient of agricultural landscapes, from highly fragmented to homogeneous, we quantified the performance of simulated (10 m, 30 m) and operational (Sentinel-2, Landsat-9) datasets using Random Forest classification and a grid-based spatial consistency validation. The results reveal a fundamental bipolar error mechanism dictated by the coupling of spatial resolution and landscape heterogeneity: (1) Spatial resolution acts as a rigid physical filter. In highly fragmented landscapes, coarse resolution (30 m) caused the statistical extinction of minor crops (e.g., spring maize), leading to severe underestimation with a 100% omission error. (2) Conversely, in homogeneous landscapes, the boundary aggregation effect at 30 m resolution resulted in a massive 7.8-fold overestimation of non-dominant crops. (3) We identified 10 m as the critical safety threshold, which maintains high statistical consistency with the 2 m baseline. Furthermore, operational 30 m data suffered from a deadly coupling of spatial and spectral noise, rendering it reliable only for dominant crops in continuous fields. Ultimately, this study proposes a landscape-adaptive hierarchical decision matrix for sensor selection. These mechanistic insights provide a robust theoretical foundation and actionable guidelines for optimizing earth observation systems in globally fragmented agricultural landscapes.