<p>The improvement in existing remote sensing detectors often relies on complex structures and deeper convolutional layers, leading to a significant increase in computational complexity. While previous works have designed various novel lightweight convolutions to improve target feature representation, these structures often overlook the unique prior knowledge of remote sensing scenes when applied to aerial detection tasks, resulting in decreased detection performance. Specifically, this prior knowledge includes: (1) Accurate detection of aerial objects often requires extensive contextual information. (2) The range of contextual information needed for different categories varies greatly. To address this, our paper fully considers this prior knowledge and proposes the Adaptive Spatial Selection Kernel Network (ASSK-Net). We first use large kernel convolution decomposition operations to decouple a series of convolutions with smaller sizes and rapidly expanding dilation rates, constructing diverse receptive fields with different spatial coverage ranges while significantly reducing computation. Based on this, a spatial selection mechanism is used to effectively weight feature maps with different levels of information richness, enhancing the feature differences between targets and backgrounds. Finally, we design an adaptive class equilibrium loss function, increasing the loss weight with rare categories and other easily confused categories with similar classification scores, thus improving the discriminative capability for rare categories. Comprehensive experiments on datasets (such as DOTA, UCAS-AOD, and HRSC2016) indicate that our ASSK-Net achieves optimal performance, effectively balancing computational complexity and detection accuracy.</p>