Enabling AI-Powered River Monitoring through Image Annotation Services

The client was developing an AI-powered solution to automatically detect and monitor obstructions in river-flow paths, supporting proactive environmental monitoring and water resource management for municipal authorities and environmental agencies.

To train this machine learning model to accurately identify the obstructions (fallen trees, accumulated debris, overgrown vegetation, etc.) that impede water flow, leading to increased flooding risks, ecosystem disruption, and navigation hazards, the client had collected aerial images of river paths using their drone fleet and required precise and fast image annotation on this growing dataset.

The client required image annotation services for approximately 75,000 high-resolution aerial images captured across multiple river systems under varying environmental conditions. The annotation work needed to be performed exclusively within the client's proprietary annotation platform to ensure seamless integration with their existing data pipeline.

Specific requirements included:

• Bounding box annotation services: Precise rectangular bounding boxes around obstructions blocking river flow paths.
• Binary image classification system: Objects categorized as either "Single" (individual fallen trees or isolated obstacles) or "Group" (multiple trees, debris clusters, or aggregated obstructions).
• Image labeling accuracy: Minimum 96% accuracy rate for acceptance.
• Delivery schedule: Processing 1,500-2,000 images per week with bi-weekly quality audits.
• Platform proficiency: Our team was required to complete training on the client's annotation software and adhere to their specific annotation guidelines and taxonomy.

The scale and complexity of image labeling for aerial river monitoring presented unique obstacles. From deciphering what constitutes an actual obstruction in overhead imagery to ensuring that all the team maintained identical classification standards across thousands of images, the project demanded systematic problem-solving at every stage.

Visual Ambiguity in Images Captured by Drones

It was challenging to distinguish actual obstructions from non-obstructive debris (logs on riverbanks or branches in shallow areas away from the main flow) that visually resembled blockages but didn't impede water movement. Water reflections, changing seasons, and varying lighting added complexity, requiring annotators to understand water flow patterns and object positioning to accurately identify genuine obstructions in the river’s path.

Ensuring Consistent Image Annotation across a Large Dataset

With 75,000 images spanning different river systems, weather conditions, and seasons, ensuring annotation consistency across multiple annotators was critical. Edge cases, such as partially submerged objects or debris clusters, required clear guidelines and regular team discussions to ensure everyone applied the same standards when classifying obstructions as "Single" or "Group."

Determining Appropriate Bounding Box Coverage

Fallen trees and scattered debris have irregular shapes that don't fit neatly into rectangular bounding boxes. Annotators had to decide how tightly or loosely to draw each box, carefully capturing the entire object without including too much unnecessary background, especially for long branches, partially visible objects, or items at the edges of the image (where an annotator had limited context to judge by).

Adapting to the Client’s Proprietary Annotation Platform for Image Labeling

The client's proprietary image and video annotation platform had unique interface elements, keyboard shortcuts, and workflow requirements that differed from standard annotation tools like CVAT or LabelStudio. Our team needed to master these platform-specific features while simultaneously maintaining productivity targets and quality benchmarks during the initial project phase.