Remote sensing images are rich in multi-scale objects, dense high-frequency textures, and distinct structural boundaries. Under low-bitrate compression, these images often suffer from distortion, over-smoothing, and loss of texture details due to quantization errors and strict bit budget constraints. This degradation collectively impairs rate–distortion performance and perceptual quality. Although existing deep learning-based image compression methods perform well at moderate bitrates, maintaining texture fidelity and structural consistency in complex remote sensing scenes under stringent bitrate constraints remains a significant challenge. Therefore, efficient low-bitrate remote sensing image compression is crucial for reducing storage costs, enhancing transmission efficiency, and supporting real-time downstream applications.

A generative modeling framework for remote sensing image compression is proposed that integrates adaptive convolution, a latent space diffusion model, and a latent residual prediction mechanism. The architecture consists of an encoder, a decoder, and a quantization and entropy model. Adaptive convolution is embedded in the encoder to modulate feature extraction based on local characteristics, thereby more effectively representing multi-scale objects and heterogeneous textures and improving latent compactness within a limited bit budget. A diffusion model is introduced into the latent space to learn more expressive latent distributions, enhancing the modeling of diverse texture patterns and complex structures. During reconstruction, this model alleviates over-smoothing at low bitrates and facilitates plausible detail recovery. In addition, a latent residual prediction module explicitly compensates for quantization errors by estimating correction terms from latent variables and injecting them into the reconstruction pathway. This process suppresses quantization-induced pseudo-textures and improves the recovery of edges and fine structures. The framework is trained end-to-end to balance bitrate and reconstruction quality, and its performance is evaluated from both rate–distortion and perceptual-consistency perspectives.

Experiments conducted on the Dataset for Object Detection in Aerial Images(DOTA) and UC-Merced datasets demonstrate that HiLD-RS consistently outperforms conventional codecs and representative learned baselines. On DOTA, HiLD-RS achieves superior rate–distortion performance compared to strong learned baselines (e.g., MGMNet, Cheng2020, and ELIC), delivering approximately 6.1%–40.8% average bitrate savings (BD-rate reductions) and 0.27–2.16 dB average quality improvements in BD-peak signal-to-noise ratio (PSNR) over overlapping operating ranges. For instance, HiLD-RS achieves 33.67 dB at 0.1751 bpp, whereas ELIC achieves 32.38 dB at 0.1988 bpp, corresponding to an 11.9% bitrate reduction while providing a 1.29 dB PSNR gain. Furthermore, HiLD-RS improves multi-scale structural similarity from 15.6543 to 16.8613 and reduces learned perceptual image patch similarity from 0.2411 to 0.2385, indicating simultaneous improvements in structural similarity and perceptual quality. Compared with traditional codecs such as Better Portable Graphics(BPG) and JPEG2000, HiLD-RS yields an even greater reduction of approximately 60% in BD-rate with approximately 4–4.6 dB higher PSNR. Overall, these results suggest that combining a latent diffusion prior with explicit decoder-side compensation can concurrently improve fidelity and perceptual quality under low-bitrate constraints, enabling more stable preservation of thin structures and high-frequency texture details.

HiLD-RS is an end-to-end framework for low-bitrate remote sensing image compression that integrates latent space diffusion modeling and decoder-side residual compensation. By jointly leveraging adaptive convolution, latent diffusion modeling, and residual compensation, the method effectively mitigates detail loss and quantization artifacts, substantially improving reconstruction quality for complex remote sensing scenes. The approach demonstrates strong generalization across various bitrates and scene types, consistently surpassing mainstream methods under identical settings. Performance varies with diffusion-step configurations and scene characteristics, highlighting the importance of scenario-adaptive parameter selection.