Partnering with a Global Leader to Power Accurate Land Cover Intelligence at Scale

About the customer

A leading geospatial analytics company uses AI and satellite data to provide actionable insights for global enterprises and governments. Their advanced platform transforms large datasets, including satellite imagery, mobile signals, and IoT feeds, into clear and predictive intelligence that enhances operational efficiency and strategic advantage. From infrastructure monitoring and supply chain visibility to analyzing population movements, they facilitate faster, data-driven decision-making across various industries, improving agility and real-world responsiveness.

Goals

Providing high-quality and consistent ground truth annotations using electro-optical satellite imagery is essential for developing an effective land use and land classification model. The goal is to speed up the development and deployment of AI models by ensuring quick turnaround times, rigorous quality control, and cost-effective labeling. This approach aims to reduce the time needed for production while enhancing model accuracy on a large scale.

Challenge

To accurately label diverse land use and land cover (LULC) classes across several hundred square kilometers of satellite and aerial imagery, developing a high-performance classification model that relies on precise annotations and consistent semantic labeling is essential. This task is complicated by the variability in land types, seasonal changes, spectral signatures, and terrain characteristics in remote sensing data. Achieving high-quality and consistent labeling is crucial for ensuring the model's accuracy and ability to generalize across different geographic areas.

Key challenges include distinguishing between visually similar land cover classes, such as grassland and cropland, managing mixed-use regions, addressing occlusions like shadows or cloud cover, and accurately mapping irregular boundaries. Pixel-accurate annotations are necessary, with a strict focus on boundary delineation and class consistency to uphold labeling integrity across diverse scenes.

Solution

Labelbees offered comprehensive semantic segmentation and classification annotation using a human-in-the-loop pipeline. This process integrates our extensive knowledge base, domain experts, customized workflows, and rigorous quality control measures. Our team efficiently annotated thousands of geospatial tiles, allowing the client to expedite model training and shorten development time.

Labelbees has developed a comprehensive land use and land cover ontology based on global standards while incorporating localized knowledge. This ontology is organized into eleven Level 1 categories, twenty-five Level 2 subcategories, and ten Level 3 sub-subcategories, featuring over a hundred unique ontological entries and in-depth expertise in landscape classification.

To address the complexities of the project, our in-house labeling team utilized subject matter experts, machine learning data specialists, and data scientists to ensure high-quality outcomes throughout the entire pipeline. By following a rigorous, custom-built data labeling workflow, Labelbees provided a tailored solution that included:

• Development and review of ontologies
• Identification and curation of edge cases
• Collection of custom metadata
• Detailed labeling guidelines and documentation
• Specialized labeling strategies for mixed classes and seasonal variants
• An iterative feedback loop
• Insights and reporting on the dataset 
• Dataset Insights & Reporting

This approach allowed us to maintain high standards and adapt to the needs of the project effectively.

Result

Labelbees' expertise in geospatial data annotation significantly improved the performance of an advanced land use and land cover classification model. The model attained exceptional classification accuracy across diverse terrains, facilitating applications such as urban planning, environmental monitoring, deforestation tracking, agricultural analysis, and disaster risk assessment.

This approach helped avoid costly rework and delays, leading to first-pass approval. We consistently achieved over 98% quality approvals on the first attempt through pixel-level image interpretation and precise classification. The model demonstrated high performance even with minimal datasets.

• Boosted downstream model accuracy and performance
• Accelerated deployment of geospatial AI capabilities
• Reduced rework, delays, and annotation costs
• Enabled high-confidence urban planning, agriculture, and disaster mitigation analytics