Enhancing YOLOv11 with Multi-Head Channel-Attentive Feature Enhancement Module for Aerial Image Classification

Unmanned aerial vehicles (UAVs) are revolutionizing applications from urban planning to disaster response by capturing vast quantities of high-resolution imagery. Yet translating these images into reliable object-level insights remains challenging, as small targets, complex backgrounds, occlusions, and variable lighting lead to missing critical details. In this paper, we introduce a novel aerial image classification framework that couples the state-of-the-art YOLOv11 backbone with a custom Feature Enhancement Module (FEM), designed to refine coarse feature representations via multi-head channel attention, dilated convolutions, and a global MLP fusion branch. We evaluate our approach on the Sikkim Aerial Images for Object Detection (SAIOD) dataset. The proposed model achieves an overall classification accuracy of 95.33%, with per-class accuracies peaking at 97.0% for Building and 96.5% for Car, demonstrating its ability to emphasize semantically salient structures and suppress background clutter. Convergent loss curves and a confusion matrix with minimal off-diagonal entries further confirm robust generalization and high true-positive rates across all classes.