Estimating the above-ground biomass of maize using soil background segmentation in UAV multispectral imagery

Above-ground biomass (AGB) can be one of the key indicators to assess crop photosynthetic product accumulation and final yield. It is often required to accurately monitor the maize AGB for the precise crop cultivation and national food security. Unmanned aerial vehicle (UAV) remote sensing can provide a more flexible and effective technical means for the accurate AGB estimation at the field scale, due to the high spatial resolution, strong maneuverability, and timely imaging. To address the issue of soil background interference when estimating maize AGB from UAV multispectral imagery, this study proposed a "segment first, model later" approach. A field test was conducted at the Wangxingjian Family Farm in Dafeng District, Yancheng City, Jiangsu Province, China. Firstly, The green normalized difference vegetation index (GNDVI) was utilized to segment vegetation from soil, and 5 bands of reflectance (RBs), 14 vegetation indices (VIs), and 40 texture features (TFs) were extracted exclusively from the vegetation areas. Pearson correlation coefficients between each feature and AGB were calculated to select feature variables for four input combinations: RBs, VIs, TFs, and their fusion (RBs+VIs+TFs). Finally, two machine learning algorithms, random forest (RF) and partial least squares regression (PLSR), were employed to construct AGB estimation models. The results demonstrated that: 1) When using single feature types, VIs provided the best predictive performance (e.g., for fresh AGB with RF: R²=0.860, RMSE=72.408 g/m², rRMSE=22.874%, and MAPE=25.574%; for dry AGB with RF: R²=0.844, RMSE=28.161 g/m², rRMSE=21.166%, and MAPE=23.171%). In contrast, the RBs showed the lowest performances (e.g., for fresh AGB with PLSR: R²=0.526, RMSE=133.325 g/m², rRMSE=42.118%, and MAPE=31.910%; for dry AGB with PLSR: R²=0.423, RMSE=54.200 g/m², rRMSE=40.736%, and MAPE=26.517%). The performance of TFs was intermediate (e.g., for fresh AGB with RF: R²=0.827, RMSE=80.458 g/m², rRMSE=25.417%, and MAPE=34.689%; for dry AGB with RF: R²=0.825, RMSE=29.881 g/m², rRMSE=22.458%, and MAPE=27.428%). 2) The multifeatured fusion model (RBs+VIs+TFs) outperformed all single-feature models. The fused model for fresh AGB increased the R² by 0.356, 0.209, and 0.132, respectively, whereas, the RMSE was reduced by 66.841, 44.287, and 30.347 g/m², respectively, compared with the individual RBs, VIs, or TFs. In dry AGB, the R² increased 0.339, 0.179, and 0.012, with the RMSE reductions of 19.413, 11.303, and 0.877 g/m², respectively. 3) Overall, the RF algorithm demonstrated superior robustness compared to PLSR, especially when using single features (e.g., when estimating fresh and dry AGB with RBs, the R² differences between the two algorithms were 0.265 and 0.368, respectively). Multi-feature fusion (RBs+VIs+TFs) substantially narrowed the performance gap between the two algorithms. The optimal model for fresh AGB (PLSR) achieved R², RMSE, rRMSE, and MAPE of 0.882, 66.484 g/m², 21.003%, and 25.374%, respectively; the optimal model for dry AGB (RF) achieved R², RMSE, rRMSE, and MAPE of 0.844, 28.226 g/m², 21.215%, and 21.569%, respectively. In summary, the soil background interference was effectively eliminated, making this approach particularly suitable for accurate AGB estimation of crops with sparse canopy coverage. Therefore, these findings provide scientific support for the application of UAV imagery in precision agriculture monitoring.