geoskill-super-resolution-dl

深度学习超分辨率 | Deep Learning Super-Resolution

Upscales low-resolution remote sensing imagery to high resolution while enhancing detail, outputting the super-resolved raster together with quality metrics (PSNR/SSIM).

Core model: SRCNN (Super-Resolution Convolutional Neural Network, Dong et al. 2014, ECCV/TPAMI) — a 3-layer fully convolutional network (9×9 / 5×5 / 5×5 kernels, 1→64→32→1 channels, ReLU) that takes the bicubic-upsampled result as input and learns the HR − bicubic residual. The network architecture, training objective (pixel-wise MSE) and inference pipeline match the original SRCNN.

All training/inference runs on a CUDA GPU (torch >= 2.x). Pre-trained weights srcnn_weights.pt ship with the skill (trained on synthetic ground truth); if the weights are missing, they are automatically trained on the GPU at first run and cached to disk.

Synthetic mode runs the self-consistent "high-resolution ground truth → downsampling → super-resolution" experiment, directly quantifying restoration quality and comparing against the bicubic baseline with --psnr_bicubic_only.

Dependencies / 依赖

pip install numpy rasterio scipy torch  # torch requires CUDA build for GPU inference

Usage / 使用方法

Example 1 (synthetic data, offline)

python geoskill-super-resolution-dl.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

Example 2: 3x upscaling of a real low-resolution image

python geoskill-super-resolution-dl.py --input low.tif --scale 3 --amount 0.7 --output-dir ./out

Output / 输出

Data Source / 数据源 / Source

A local single-band GeoTIFF, or --synthetic (downsample the high-resolution ground truth to obtain the low-resolution input; the ground truth is retained for evaluation).

Privacy / 隐私声明 / Privacy

• Runs offline by default; --synthetic mode requires no network at all.
• All processing is done locally; no user data is uploaded.

License / License

MIT

---

---

name: geoskill-super-resolution-dl description: '基于 SRCNN（Dong 2014）卷积神经网络的 2x/3x/4x 影像超分辨率，在 CUDA GPU 上训练与推理，输出高分辨率栅格与 PSNR/SSIM 评估'

深度学习超分辨率 | Deep Learning Super-Resolution

把低分辨率遥感影像放大到高分辨率并增强细节，输出超分栅格与质量评估指标（PSNR/SSIM）。

核心模型：SRCNN（Super-Resolution Convolutional Neural Network, Dong et al. 2014, ECCV/TPAMI） —— 3 层全卷积网络（9×9 / 5×5 / 5×5 核，1→64→32→1 通道，ReLU），以双三次上采样结果为输入、 学习 HR - bicubic 残差；网络结构、训练目标（pixel-wise MSE）与推理流程与原始 SRCNN 一致。 所有训练/推理在 CUDA GPU 上执行（torch >= 2.x）；随 skill 附带预训练权重 srcnn_weights.pt（在合成真值上训练得到）；若权重缺失则在首次运行时自动用 GPU 训练并落盘缓存。

合成模式执行"高分辨率真值 -> 降采样 -> 超分"自洽实验，可直接量化恢复质量，并与 --psnr_bicubic_only 双三次基线对比。

依赖

pip install numpy rasterio scipy torch  # torch 需 CUDA 版以使用 GPU 推理

使用方法

示例 1（合成数据，离线）

python geoskill-super-resolution-dl.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

示例 2：真实低分影像 3 倍放大

python geoskill-super-resolution-dl.py --input low.tif --scale 3 --amount 0.7 --output-dir ./out

输出

数据源 / Source

本地单波段 GeoTIFF，或 --synthetic（高分辨率真值降采样得到低分辨率输入，真值留存用于评估）。

隐私声明 / Privacy

• 默认离线运行，--synthetic 模式完全无网络。
• 所有处理在本地完成，不上传任何用户数据。

License

MIT