Plastic Greenhouse Extraction Based on Index-Guided Three-Stage Large-Factor Remote Sensing Image Super-Resolution Reconstruction

Fine-scale, large-area plastic greenhouse (PG) mapping generally relies on costly high-resolution imagery with limited spatial coverage. Although freely available Sentinel-2 data offer wide swaths and frequent revisits, their insufficient spatial resolution limits PG extraction accuracy. Direct super-resolution (SR) reconstruction from 10 m or coarser imagery to the submeter level is highly ill posed because PG structures must be inferred from severely mixed pixels. To address these challenges, we propose a three-stage 12× SR framework guided by a spatially enhanced Agricultural Plastic Greenhouse Index (APGI). The framework decomposes reconstruction into successive 4×, 2×, and 1.5×-scale transformations and progressively recovers PG spatial arrangements and local boundaries through an RRDB generator and learnable upsampling operations. APGI values, directional gradients, and gradient magnitude are encoded as structural priors to modulate shallow, backbone, and upsampling features. Intermediate-scale supervision, interstage residual refinement, and tone consistency constraints are further introduced to improve the structural stability and spectral consistency of the reconstructed imagery. The model is trained using strictly co-registered Sentinel-2 and Jilin-1 image pairs, and its practical value is further evaluated through a downstream PG segmentation task. Spatially blocked five-fold cross-validation is used, with four folds for training and one fold for testing in each run, to prevent overlapping or adjacent patches from entering different subsets. Experiments conducted in Weifang, China, and Almería, Spain, demonstrate that progressive reconstruction and APGI guidance improve PG boundary continuity, reduce extraction errors, and enhance downstream segmentation accuracy. Accordingly, the framework is positioned as task-driven conditional image generation that recovers a recognition-relevant PG structure, while textures below the Sentinel-2 information limit remain learned inferences rather than direct observations.