Turning satellite angles into signals for smarter soybean yield forecasts

Satellite viewing angles can be more than a source of uncertainty—they can reveal useful clues for crop yield estimation. Researchers have developed M-Net, a physics-guided deep learning framework that combines multi-angular canopy information with solar-induced chlorophyll fluorescence to improve soybean yield estimates. The findings highlight the value of directional information in satellite-based crop monitoring, beyond spatial resolution alone.

 

Accurate crop-yield forecasting supports food-security planning, agricultural management, and commodity markets. Satellite vegetation indices describe canopy structure, while solar-induced chlorophyll fluorescence (SIF) more directly reflects photosynthetic activity. However, both signals change with solar and sensor geometry, complicating comparisons across viewing angles. Conventional processing often normalizes these angular effects as noise, potentially discarding information about three-dimensional canopy architecture. Meanwhile, purely data-driven deep learning can lack physical interpretability and robustness when fusing heterogeneous observations from multiple sensors and scales. The challenge is to turn these angular differences into useful information while bringing observations from different sensors together.

Researchers from China Agricultural University, Pacific Northwest National Laboratory, KU Leuven, and Southwest Jiaotong University reported (DOI: 10.34133/remotesensing.1059) the work in the Journal of Remote Sensing, formally published online on August 3, 2026. The study tackles a practical limitation in satellite yield monitoring: different sun–sensor geometries can create inconsistent fluorescence and reflectance signals. Rather than removing those differences, the team designed M-Net to interpret them as information about canopy structure and crop development, enabling more robust integration of physiological and structural indicators.

M-Net’s key innovation is its physics-guided, angle-aware design. Instead of treating directional variability as error, the framework uses bidirectional reflectance distribution function (BRDF)-derived features to represent how soybean canopies appear under different illumination and viewing angles. It then combines these structural cues with SIF-based physiological information through attention and temporal-learning modules. Across 613 U.S. soybean-growing counties, M-Net’s best-performing tested configuration—combining SIF, multi-angular MODIS vegetation indices, and leaf area index—achieved R² = 0.69. Multi-angular MODIS structural information achieved R² = 0.64, outperforming higher-resolution Sentinel-2 observations at fixed angles (R² = 0.56), showing that angular information can be more valuable than spatial resolution alone for regional soybean yield estimation at scale.

The study covered 613 counties across 29 U.S. states from 2019 to 2023; 2019–2022 data were used for training and validation, while 2023 served as an independent test. TROPOMI observations showed strong geometric variability: median solar and viewing zenith angles were 23.9° and 44.0°, respectively, with viewing angles reaching about 60° at the 95th percentile. A GRU using structural features reached R² = 0.62, but performance declined after physiological information was added. M-Net instead benefited from multi-source fusion, reaching R² = 0.66 in direct baseline comparisons and up to R² = 0.69 in feature-ablation experiments. Its best configuration produced a mean RMSE of about 0.49 Mg/ha. Prediction error also fell as soybean purity increased, from 0.44 Mg/ha across all counties to 0.30 Mg/ha among 104 counties with more than 30% soybean cover. The network assigned greater temporal attention to late pod-filling and maturation stages, when prediction accuracy was highest in practice.

The team emphasized that the finding changes how directional satellite observations should be interpreted. As the authors wrote, “angular signatures are valuable signals to be decoded, not noise to be normalized.” Combining physical knowledge with deep learning, they argue, could move crop monitoring beyond purely correlational prediction toward more interpretable and transferable Earth-observation models.

The researchers assembled county-level time series from TROPOMI SIF, MODIS BRDF products, Sentinel-2 reflectance, leaf area index, meteorological variables, and USDA crop and yield records. BRDF modeling reconstructed directional vegetation and fluorescence-related features across multiple sun–sensor configurations. M-Net processed them through physics-guided angle encoding, multi-head attention, a bidirectional gated recurrent unit, and a final prediction head. Performance was compared with Random Forest, CNN, LSTM, and GRU baselines across five independent trials using R² and RMSE for robust overall model comparison.

The framework could support more reliable regional yield forecasts, agricultural risk assessment, crop insurance, food-supply planning, and precision-management services. Future work must validate M-Net beyond the U.S. soybean belt across other crops, climates, and management systems. The authors also highlight higher-resolution fluorescence missions such as FLEX, improved SIF downscaling, fuller viewing-geometry representation, and stronger radiative-transfer constraints within the network. These advances could help create globally transferable, physics-guided monitoring systems that extract more value from current and future Earth-observation satellites worldwide.

###

References

DOI

10.34133/remotesensing.1059

Original Source URL

https://doi.org/10.34133/remotesensing.1059

Funding information

This work was supported by the National Key R&D Program of China (2024YFF0810500), the National Natural Science Foundation of China (No. 42271361), the Chinese Universities Scientific Fund, the Pinduoduo-China Agricultural University Research Fund (No. PC2024B01006), ESA-MOST Dragon 6 Cooperation (No. 95441), and the 2024 Key Project of the Independent Innovation Research Fund for Graduate Students of China Agricultural University.

About Journal of Remote Sensing

The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.