Crop classification in Google Earth Engine: leveraging Sentinel-1, Sentinel-2, European CAP data, and object-based machine-learning approaches

Abstract

In contemporary agriculture and environmental management, the need for precise and accurate crop maps has never been more vital. Although object-based (OB) methods within Google Earth Engine (GEE) improve accuracy and output quality in contrast to pixel-based approaches, their application to crop classification remains relatively rare. Therefore, this study aimed to develop an OB classification methodology for crops located in central Italy’s Lake Trasimeno area. This methodology employed spectral bands, spectral indices (Normalized Difference Vegetation Index and Modified Radar Vegetation Index), and textural information (Gray-Level Co-occurrence Matrix) derived from Sentinel-2 L2A (S2) and Sentinel-1 GRD (S1) data within the GEE platform. Moreover, European Common Agricultural Policy (CAP) data associated with cadastral parcels were employed and served as ground information during the training and validation stages. The CAP crop classes were aggregated into three levels (Level 1–3 crop types, Level 2–5 crop types, and Level 3–7 crop types). Subsequently, optimized Random Forest (RF) classifiers were applied to map crops effectively. Feature selection analysis highlighted the importance of certain textural features. Additionally, findings demonstrated high overall accuracy results (89% for Level 1, 86% for Level 2, and 82% for Level 3). It was found that winter crops achieved the highest F-score at Level 1, while specific subclasses, such as winter cereals and warm-season cereals, excelled at Level 2. Overall, this study provides a promising approach for improved crop mapping and precision agriculture in the GEE environment.

Keywords

• Machine learning
• random forest
• IACS
• CAP
• winter cereals
• summer crops