Abstract: The fraction of absorbed photosynthetically active radiation (FPAR) is a key biophysical parameter characterizing the light interception capacity and photosynthetic potential of crop canopies. Its accurate estimation is essential for quantifying canopy light interception in maize, diagnosing crop growth status, and supporting precision agricultural management. Existing UAV-based FPAR studies have primarily relied on single-source spectral information, which is insufficient to fully characterize the coupled effects of canopy biochemical status, surface spatial heterogeneity, and three-dimensional structure on radiation absorption. In this study, we developed a maize FPAR estimation framework that integrates UAV-derived multi-source remote sensing features with a lightweight Transformer model for maize FPAR retrieval. Field experiments were conducted during the maize large trumpet, tasseling, and grain-filling stages in 2024 and 2025. Water, nitrogen fertilization, and cultivar treatments were imposed to increase variability in canopy growth conditions. A total of 1,296 valid ground-based FPAR observations were collected, together with synchronous UAV multispectral and RGB imagery. Thirteen vegetation indices and 40 gray-level co-occurrence matrix (GLCM) texture features were extracted from the multispectral images, while canopy structural variables, including canopy height and canopy cover, were derived from the RGB images. Based on the fusion of these three categories of features, a lightweight Transformer model tailored for multimodal feature inputs was designed and compared with benchmark models, including partial least squares regression, support vector regression, random forest, and deep neural networks. In addition, ablation experiments involving single-feature inputs, dual-feature fusion, and three-feature fusion were conducted to systematically evaluate the effects of different feature configurations on model performance. Experiments based on independent experimental-plot grouping were further performed to assess the spatial generalization capability of the proposed model. Correlation analysis indicated that canopy structural features were the most effective source of information for characterizing maize FPAR. Among these features, canopy cover maintained a stable correlation with FPAR across the three growth stages. For the three individual growth stages, the lightweight Transformer achieved test-set R² values of 0.835, 0.808, and 0.770 at the large trumpet, tasseling, and grain-filling stages, respectively, outperforming the baseline models, including partial least squares regression, support vector regression, random forest, and deep neural networks. These results demonstrate that the lightweight Transformer model can effectively capture cross-feature dependencies among spectral, textural, and structural input features. When samples from the three growth stages over two years were combined for model development, the Transformer still achieved the highest accuracy, with a test-set R² of 0.903. Under a more stringent validation strategy based on independent experimental-plot grouping, the Transformer maintained test-set R² values of 0.829, 0.787, and 0.775 across the three growth stages, respectively, indicating that the model retained a certain degree of spatial generalization capability under independent plot partitioning. The ablation experiments further confirmed the necessity of multi-source feature fusion. Among all single-feature inputs, structural features achieved the highest accuracy across the three growth stages, whereas the individual use of spectral or textural features resulted in relatively low model accuracy and was insufficient to support robust FPAR retrieval. For dual-feature combinations, those incorporating structural information consistently outperformed those without structural information, indicating that structural information played a critical role in the feature system. In contrast, the joint input of spectral, textural, and structural features achieved the best performance across all growth stages. These results suggest that maize FPAR is jointly regulated by multidimensional canopy attributes, including spectral characteristics, spatial texture, and geometric structure, and that the lightweight Transformer model can effectively integrate these complementary UAV remote sensing features. The proposed method provides a practical methodological framework for high-throughput estimation of maize FPAR and offers methodological support for continuous monitoring of canopy photosynthetic traits and precision agriculture decision-making.