Mid-season parcel-based crop type mapping using multi-source time series of big Earth observations and deep learning

Accurate and timely crop type mapping offers a wide range of benefits, from supporting annual crop inventories and agricultural monitoring to enabling sustainable resource management and a deeper understanding of landscape dynamics. This study proposes a practical framework for mid-season parcel-level crop classification utilizing models derived from past Earth Observation (EO) and training data, eliminating the necessity for additional training data collection for future inference. Utilizing Sentinel-1 SAR and Sentinel-2 optical data with a 21-day interval from April 1 to August 4, major crop types were mapped in the province of Quebec during the growing season spanning 5 years (2020–2024). Finally, the most accurate trained models were used to predict the crop types of 2025 to ensure the accuracy and transferability of the trained models. This study took into account several important crops, such as corn, soybeans, wheat, and hay. Two machine learning models and four deep learning models, including Transformers, were utilized for multitemporal crop mapping. Additionally, crop mapping was performed using Google’s AlphaEarth annual dataset and ML algorithms, and the results were compared with Sentinel-1 and Sentinel-2 data. Our findings indicated that the Transformers outperformed other models when using Sentinel-1 and Sentinel-2 data, achieving an overall accuracy of 0.97 on the validation data and 0.93 on the test data. Using Google’s AlphaEarth data, RF delivered a comparable accuracy, with overall accuracies of 0.96 and 0.92 for the validation and test data, respectively. The generated crop type maps for unseen 2025 data in this study show the practicality of the proposed framework to perform mid-season parcel-based crop mapping without the need for additional in-year field data.