Abstract: Lodging is a major factor limiting wheat yield and harvesting efficiency. Wheat lodging can reduce crop yield by 10% to 40% and harvesting efficiency by 20% to 30%. Timely and accurate monitoring of lodged areas is therefore essential for mitigating production losses and ensuring food security. At present, wheat lodging detection mainly relies on traditional machine learning and deep learning methods. Deep learning has become the mainstream approach because of its ability to automatically learn hierarchical feature representations from data. However, existing deep learning segmentation networks still face several challenges in complex farmland scenes. In particular, they have difficulty extracting complex texture features from lodged wheat canopies, show limited adaptability to the irregular geometric deformations of lodging patterns, and often produce blurred or discontinuous boundaries because of the loss of high frequency details during feature encoding and decoding. To address these problems, this study used UAV visible light images acquired in a major wheat production area as the data source and proposed E-UKAN based on U-KAN. To strengthen feature representation throughout the network, E-UKAN introduces targeted improvements into the encoder, bottleneck, and decoder. First, an Adaptive Gabor-Wavelet Feature Enhancement Module (AGW-FEM) was designed in the encoder. The module dynamically captures multidirectional texture features and adaptively extracts texture responses in different directions, while wavelet processing supplements high frequency details. This design improves the network’s ability to distinguish complex lodging textures and reduces the interference caused by the similar visual appearance of lodged and standing wheat. Second, a Frequency-Aware Dynamic Bottleneck (FADB) was introduced at the bottleneck. Through multilevel frequency decomposition, the module decouples high and low frequency information so that texture details and spatial structure can be represented more effectively. In addition, a constrained dynamic convolution mechanism adjusts sampling positions according to local feature characteristics, thereby improving its geometric adaptability to irregular lodging patterns. Finally, a Frequency-Spatial Dual-domain Boundary Refinement Decoder (FSD-BRD) was designed in the decoder. By integrating wavelet transform, gated convolution, multiscale high frequency edge information, and spatial contextual information, the decoder captures high frequency details and models long range dependencies. It thereby alleviates boundary blurring caused by detail loss and improves boundary continuity, particularly in occluded regions and transition regions with partial lodging. Experiments were conducted on the wheat lodging dataset collected in this study to evaluate the effectiveness of E-UKAN. The experimental results showed that E-UKAN achieved a mean Intersection over Union, Dice coefficient, Pixel Accuracy, Precision, and Recall of 90.70%, 96.97%, 95.81%, 97.72%, and 96.23%, respectively. Compared with the original U-KAN baseline, these values increased by 6.40, 2.09, 3.01, 3.78, and 0.38 percentage points, respectively. It has also outperformed the other mainstream comparison networks while slightly reducing the parameter count, demonstrating the effectiveness of the three proposed modules in texture enhancement, geometric adaptation, and boundary refinement. Overall, E-UKAN provides an effective method for the fine segmentation and timely and accurate monitoring of wheat lodging in complex farmland scenes, providing a new solution for wheat lodging detection.