Satellite embeddings for crop type classification: a comparative examination

Embedding datasets encode complex relationships among multiple sources of Earth observation data into a compact format. Here, we evaluated the utility of a 10-m global Satellite Embedding product (SE) for classifying crop types in central California for the year 2020. We compared the classification accuracy of a random forest model based exclusively on the SE layer to an existing random forest model with multiple imagery inputs. Our results showed the SE-based classification had higher agreement with the reference dataset (California Department of Water Resources crop map) than the classification based on Landsat and National Agricultural Imagery Program inputs (94.7% versus. 91.9% overall accuracy, respectively). The performance of individual crop types was consistent across models, ranging from high agreement for rice (98.4% versus 98% accuracy) to lower agreement for pasture, grain, and fallow/young perennial classes (< 65% accuracy in both models). The SE-based workflow used three times less cloud-based computational resources and represented substantial savings of predictor development time. Geospatial embedding products can aid classification efforts by reducing predictor development time and processing demands while maintaining classification accuracy.