Abstract: [Objective]Small and medium-sized unmanned aerial vehicles (UAVs) are expected to play an increasingly important role in UAV front-end intelligence for geological applications. However, limitations such as low onboard computational performance and restricted battery capacity continue to constrain the deployment of intelligent models on UAV platforms. To address this challenge, this study integrates the multi-kernel lightweight convolutional model ultralight_unet micro model into geological interpretation tasks under complex environments, and evaluates its effectiveness in geological feature interpretation for front-end embedded systems. Distinct from traditional passive compression-based lightweight model approaches—such as pruning and quantization—and from existing lightweight networks that rely on single kernels or weak attention mechanisms, ultralight_unet employs an inherently lightweight multi-kernel architecture (MKIR/MKIRA) that enables more robust multi-scale geological feature extraction at extremely low computational cost.[Methods]Using Landsat-8 imagery from the Eastern Kunlun region, we conduct a systematic comparison between the ultralight_unet micro model and large-scale models such as U-Net and DeepLabv3plus, as well as mainstream lightweight networks including MobileNetV3 and Fast-SCNN. The comparison assesses performance across model parameters, floating-point operations, and interpretation accuracy to reflect deployment requirements typical of UAV front-end intelligence scenarios. [Results]Results show that the ultralight_unet micro model contains only 0.32M parameters and 0.77G FLOPs, representing 92–466× and 10–230× reductions compared with U-Net and DeepLabv3plus, respectively. It achieves an overall Pixel Accuracy (oPA) of 62.75%, a mean Intersection over Union (mIoU) of 40.82%, and an F1-score of 55.68%. Compared with SegNet, oPA, mIoU, and F1-score improve by 4.14%, 6.98%, and 6.92%, respectively. [Conclusion]Moreover, the ultralight_unet micro model demonstrates lower complexity and computational cost than MobileNetV3 and Fast-SCNN, while offering enhanced feature representation for remote sensing scenes characterized by weak geological textures and blurred boundaries. This provides a deployable lightweight solution for UAV-based geological feature interpretation on front-end devices.Although its accuracy remains below that of certain large-scale state-of-the-art multimodal networks, this study provides experimental evidence and methodological insights for the intelligent deployment of UAV geological equipment, and establishes a foundation for developing more advanced lightweight models tailored to specific tasks.