AKGMamba model estimates seed maize yield from Sentinel-2 time series

Chinese researchers have developed AKGMamba, a deep learning model that estimates seed maize plot yields from Sentinel-2 imagery. It combines 12 vegetation indices with six crop growth parameters (including leaf area index and fAPAR) retrieved through the SL2P processor on Google Earth Engine, and applies a flowering-synchrony decay rule at its output.

Tested on 853 yield-labelled plots in Zhangye, Gansu Province, across 2021, 2023 and 2024, it reached an R2 of 0.78 and RMSE of 793.73 kg/ha. Transformer scored 0.65, LSTM 0.54 and XGBoost 0.35. Adding growth parameters to spectral features cut error by 48.9 percent.

Cross-year tests were weaker, with R2 of 0.48 when testing on 2024. All plots were drip-irrigated, and the authors note that meteorological drivers are missing. The work appears in Artificial Intelligence in Agriculture.

Why it matters: The study shows that pairing satellite time series with radiative-transfer-derived crop parameters and agronomic rules can improve yield estimates for specialised crops such as seed maize. Its irrigated, single-region data limit direct transfer to rain-fed African systems.