Machine Learning-Based Data Fusion to Enhance Spectral Capabilities of Planet Data with Sentinel-2

Abstract

Soybean is an essential crop to fight global food insecurity and ensure economic vigor around the world. While the agricultural community has pushed genetic improvements in soybean to increase the yield, fewer efforts have been made to ensure increased amount of protein, oil and other seed composition traits. Since crop phonology across the soybean growth cycle ensures the nutrient accumulation in its seeds, remote sensing offers a unique opportunity to estimate seed composition traits. This study introduces a novel spectral fusion technique called multiheaded kernel-based spectral fusion (MKSF) that combines the higher spatial resolution of PlanetScope (PS) and increased spectral bands from Sentinel 2 (S2) satellites together. The study also focuses on using the additional spectral bands and different statistical machine learning models to estimate soybean seed composition traits, e.g., protein, oil, sucrose, starch, ash, fiber, and yield. The MKSF technique was trained using PS and S2 image pairs from different growth stages in the soybean growing cycle, and predicted the potential VNIR1 (705nm), VNIR2 (740nm), VNIR3 (783nm), SWIR1 (1610nm), and SWIR2 (2190nm) bands from the PS images. Our results indicate that VNIR3 prediction performance was the highest (R2 of 0.90), followed by VNIR2, VNIR1, SWIR1, and SWIR2. Among the seed composition traits, sucrose yielded the highest predictive performance. Overall, the random forest regression outperformed other models and the performance level varied with different growth stages. Finally, the feature importance analysis reveals the importance of MKSF-generated vegetation indices as the topmost important feature was found from the fused image bands.