Abstract: Wheat is one of the primary food crops for global food security. However, temperature fluctuations can affect wheat growth, physiological characteristics, and yield, particularly against global warming. Alternatively, organic fertilization can be expected to regulate soil structure and microbial activity under environmental stress from high temperatures. Since the interaction between global warming and fertilization can result in phenotypic variations in winter wheat, conventional monitoring cannot fully meet the needs of large-scale cultivation in recent years. Therefore, it is often required for highly efficient and accurate high-throughput phenotypic classification to monitor crop growth and cultivation under a dynamic environment. This study aims to investigate phenotypic variations in winter wheat under different cultivation using hierarchical multimodal fusion. Four treatments were established, including no warming with inorganic fertilizer (T0SF), no warming with organic fertilizer (T0OF), warming with inorganic fertilizer (T1SF), and warming with organic fertilizer (T1OF). UAV remote sensing was then employed to acquire visible light and multispectral imagery under stable illumination. A hierarchical multimodal fusion was proposed for high-throughput phenotypic classification (HMF-HTPC). Winter wheat phenotypes were classified under warming and fertilization. Five vegetation index images were selected, including the green normalized difference vegetation index (GNDVI), leaf chlorophyll index (LCI), normalized difference red edge index (NDRE), normalized difference vegetation index (NDVI), and optimized soil adjusted vegetation index (OSAVI). UAV visible light images were also obtained for hierarchical multimodal feature representation and fusion. Visible structural information was combined with spectral responses. The framework was improved to recognize phenotypic features under different treatments. A series of experiments were conducted to validate the effectiveness of the model using single- and multi-modality deep learning. The results showed that single-modality deep learning models, including ResNet, MobileNet, and ViT, ResNet, achieved a classification accuracy of 77.50%. Single-modality approaches shared the limited feature representation because they relied on visible light imagery, thereby restricting classification performance. Among all multimodal deep learning models, the best performance was achieved with a classification accuracy of 84.58% using visible light and vegetation index images. Simple fusion strategies also restricted the effective exploitation of complementary information from multiple data sources. In contrast, the HMF-HTPC model with hierarchical multimodal fusion was realized to integrate features from visible light and vegetation index images at multiple levels. Classification accuracies of 91.67% and 91.25% were achieved on the 2024 and 2026 heading-stage datasets, respectively, and 96.67% on the 2026 grain-filling-stage dataset. Stable and high performance of classification was maintained over years and phenological stages, indicating strong robustness and generalization. In addition, a systematic evaluation was also conducted on the classification performance of visible light imagery with vegetation index images. The LCI and NDRE vegetation indices contributed most to the model performance, and their combination enhanced the accuracy of winter wheat phenotypic classification. Overall, this finding can provide an effective technical framework for high-throughput phenotypic classification of winter wheat, particularly for crop cultivation and soil fertilization under global warming.