Abstract

The Brazilian Institute of Geography and Statistics proposes using automated crop parcel recognition methods, with a reliance on Sentinel-2 satellite imagery. This choice is driven by the availability of free, high-frequency imaging, essential for large-scale agricultural monitoring, where manual monitoring and parcel delineation are unfeasible due to resource constraints.

This study presents a comprehensive pipeline that leverages Google Earth Engine (GEE) to facilitate the acquisition, processing, and analysis of Sentinel-2 satellite imagery for agricultural parcel detection. Our approach prominently features the integration of transfer learning methods, specifically designed to adapt models trained on European regions to the varied landscapes of Brazil.

By leveraging the extensive spectral and temporal information provided by Sentinel-2, we create a robust foundation for model training that can capture the intricate land cover patterns characteristic of both European and Brazilian environments. The strategic use of transfer learning allows us to address the challenges posed by domain shifts, thus maximizing the utility of pre-existing datasets of Europe for effective parcel detection in Brazilian regions. This methodology highlights the significance of cross-continental knowledge transfer in enhancing remote sensing applications, demonstrating the adaptability and generalizability of machine learning models across diverse geographical areas.

The pipeline begins by accessing the extensive GEE repository to retrieve Sentinel-2 images, filtering them based on temporal and spatial criteria to ensure minimal cloud cover and maximum land feature clarity. Preprocessing steps, including atmospheric correction, radiometric normalization, spatial resampling, and cloud masking, are applied to enhance image quality. Subsequently, the pipeline employs machine learning algorithms for semantic segmentation, using supervised classification methods with labeled training data from Europe datasets to create the first prediction results. The model further incorporates multi-temporal information, examining images from multiple time points to capture vegetation pattern variations critical for precise crop boundary identification. This dynamic approach improves the robustness and precision of parcel delineation, offering insights into crop phenology and health.

Post-processing steps refine parcel boundaries through morphological operations, and validation against ground truth data ensures result reliability. The outcome is an accurate, high-resolution map of agricultural parcels, compatible with Geographic Information Systems (GIS) for further analysis and application.

In conclusion, our pipeline utilizing Google Earth Engine and Sentinel-2 data represents a significant advancement in the field of parcel detection. Through the integration of transfer learning, we effectively adapt European models to the Brazilian context, showcasing the potential for cross-continental knowledge transfer in remote sensing applications. This study not only enhances the scalability and accessibility of parcel detection solutions but also contributes to the global discourse on sustainable land use practices. By leveraging the strengths of Sentinel-2 imagery and advanced machine learning techniques, we provide a robust and adaptable tool for addressing complex land management challenges on a global scale.