A hybrid deep clustering and machine learning-based decision support framework for modeling crop-livestock environmental suitability and spatial discrepancies

Abstract The development of China’s National Modern Agricultural Industrial Parks (NMAIPs) has provided valuable experience to guide regional agricultural structural adjustment. To systematically analyze and scale up the successful practices of crop–livestock spatial layouts, this study examines 335 NMAIPs established between 2017 and 2024.Based on seven natural environmental variables, a deep clustering model (VAE-GMM) was applied to classify the parks into representative environmental types. This classification establishes a standardized spatial reference frame. Concurrently, a LightGBM multi-label classifier was utilized to predict the theoretical suitability of various crop–livestock spatial configurations. Crucially, the study introduces a spatial discrepancy (Gap) metric to evaluate industrial expansion potential. This metric is explicitly calculated as the difference between the model-predicted theoretical suitability proportion and the actual occurrence frequencies within each environmental cluster. The results show that the parks can be grouped into five distinct environmental types with clear regional spatial patterns. The LightGBM prediction achieved a micro-average AUC of 0.75, effectively capturing natural constraints. Furthermore, the discrepancy analysis reveals a structural divergence between environmental suitability and real-world agricultural allocation. Quantifying this divergence highlights environmentally suitable yet underrepresented industries. By treating existing parks as reference cases under specific environmental baselines, this data-driven framework provides objective, transferable decision support for industrial selection and spatial planning in newly established agricultural parks.