# A novel fusion of Sentinel-1 and Sentinel-2 with climate data for crop phenology estimation using Machine Learning

**Authors:** Shahab Aldin Shojaeezadeh, Abdelrazek Elnashar, Tobias Karl David Weber  
**arXiv:** [2409.00020](https://arxiv.org/abs/2409.00020)  
**Date:** 2024-08-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.00020v2/

## Key points

- Understanding crop phenology is crucial for effective agricultural management strategies.
- Efforts have been made to detect crop phenology using Remote Sensing (RS) and high resolution weather data.
- A Machine Learning (ML) LightGBM model was trained to predict 13 phenological stages for eight major crops across Germany at a 20-meter scale.
- The model achieved promising outcomes with predicted phenology having an R2 > 0.43 and a low Mean Absolute Error of 6 days across all phenological stages and crops.
- Combining radar sensors with climate data led to significant improvements in the model's performance.
- The fusion approach of Sentinel-1 and Sentinel-2 data with climate information has great potential for enhancing crop phenology estimation accuracy and supporting informed agricultural decisions.

## Summary

In the field of agriculture, understanding crop phenology is crucial for decision makers to effectively plan and adapt agricultural management strategies. With the increasing availability of big Earth observation data, efforts have been made to accurately detect crop phenology using Remote Sensing (RS) and high resolution weather data. However, many existing studies have focused on large scale predictions or developed methods that do not fully support crop modeler communities in leveraging Sentinel-1 and Sentinel-2 data fused with high resolution climate data. To address this gap, a Machine Learning (ML) LightGBM model was trained to predict 13 phenological stages for eight major crops across Germany at a 20-meter scale. Observed phenologies from the German national phenology network (German Meteorological Service; DWD) between 2017 and 2021 were used for training and validation. A thorough feature selection analysis was conducted to identify the best combination of RS and climate data for detecting phenological stages. The results showed promising outcomes at a national scale, with predicted phenology achieving a reasonable precision of R2 > 0.43 and a low Mean Absolute Error of 6 days across all phenological stages and crops. The spatio-temporal analysis demonstrated the model's transferability across different spatial and temporal contexts in Germany. By combining radar sensors with climate data, the model showed significant improvements that can benefit various practical applications. Furthermore, insights from previous studies highlighted the importance of data volume in influencing ML model accuracy, as well as the impact of specific feature sets on predicting different phenological stages of crops. For instance, radar backscatter coefficients combined with optical raw band data were found to be valuable for identifying early stages in winter wheat crops. Overall, this novel fusion approach of Sentinel-1 and Sentinel-2 data with climate information holds great potential for enhancing crop phenology estimation accuracy, supporting informed agricultural decisions, contributing to sustainable food production practices, and aiding in crop model calibrations and evaluations to meet the growing global food demand.