Abstract
Extracting lake boundaries from remote sensing imagery is a critical yet challenging task due to complex environmental noise and irregular target shapes. While prompt-based models like the Segment Anything Model (SAM) have shown promise, their reliance on manual interaction and heavy computational backbones limits their utility in automated, resource-constrained remote sensing applications. This paper introduces LEPrompter, a novel two-stage prompt enhancement framework that leverages a "teacher-student" paradigm. By utilizing automatically generated prompts during training to guide feature learning and subsequently removing them during inference, LEPrompter achieves state-of-the-art accuracy with zero additional computational overhead at deployment.
Introduction
The precise extraction of lake boundaries from satellite imagery is often hindered by reflections, cloud cover, and intricate surrounding terrains such as snow or mountainous regions. Traditional automated methods frequently struggle with these "noisy" inputs. Although recent advancements in interactive segmentation, notably SAM, provide powerful prompting capabilities, they are not inherently designed for fully automated semantic segmentation in remote sensing. SAM requires manual point or box inputs and lacks specific semantic labels. Furthermore, its massive architecture is often unsuitable for efficient deployment.
To bridge this gap, the researchers proposed LEPrompter, which integrates the benefits of prompt-based learning into a lightweight, automated pipeline.
Methodology: The LEPrompter Framework
The core of LEPrompter is a two-stage training strategy designed to instill "prompt knowledge" into a standard segmentation backbone without requiring those prompts during final use.
1. Prompt-Based Training (Teacher Stage)
In the initial training phase, the model acts as a "student" guided by a "teacher" (the prompts). The framework incorporates a lightweight Prompt Encoder and Prompt Decoder, which add a mere 1.23M parameters. These modules process various prompt types—points, boxes, or masks—derived from ground truth data and fuse them with image features extracted by the vision encoder. This guidance forces the model to attend to the specific morphological features of lakes.
2. Prompt-Free Training (Student Stage)
After an optimal period (identified as 50,000 steps), the prompts are deactivated. The model must then perform extraction independently. Having been "shown" the relevant features by the prompts in the first stage, the model retains a superior ability to identify lake boundaries even in the absence of explicit hints.
3. Prompt-Free Inference
During deployment, the prompt-related modules are completely bypassed. This ensures that the performance gains realized during training are preserved while maintaining the inference speed and efficiency of a standard, prompt-less segmentation network.
Benchmark and Datasets
The study evaluates LEPrompter on two primary remote sensing datasets:
- SW (Surface Water) Dataset: Comprising 17,596 images at 256x256 resolution.
- QTPL (Qinghai-Tibet Plateau Lake) Dataset: Consisting of 6,773 high-resolution images.
A significant contribution of this work is the establishment of a new benchmark for automated prompt generation. By applying morphological operations—including DBSCAN clustering, dilation, erosion, and contour finding—to the ground truth, the researchers generated five prompt types: random points, center points, bounding boxes, filled masks, and unfilled masks.
Experimental Results
LEPrompter was tested across several backbones, including SegFormer, PoolFormer, SegNeXt, and the specialized LEFormer.
- Performance Gains: On the SW dataset, LEFormer enhanced with LEPrompter achieved an mIoU of 91.53%, outperforming baseline models significantly.
- Generalization: The framework's utility extends beyond remote sensing. When applied to medical imaging datasets (CVC-ClinicDB for polyps and ISIC2018 for skin lesions), LEPrompter improved mIoU by over 1%, demonstrating its robustness across domains.
- Efficiency: The training overhead is minimal (0.95G FLOPs), and the inference overhead is zero, making it ideal for real-time applications.
Conclusion
LEPrompter demonstrates that prompts can serve as a powerful supervisory signal during training rather than just an interactive tool. By using prompts as "training wheels," the model learns to capture intricate details that standard supervision might miss. Future research will focus on generating these prompts in unsupervised or "in-the-wild" scenarios and exploring the use of "negative prompts" to further minimize false positives in highly noisy environments.
Summary updated