Unmanned aerial vehicle (UAV) remote sensing has emerged as a crucial approach in precision agriculture, due to the high timeliness, low cost of data acquisition, and superior spatial resolution. Physical and chemical parameters of crops can be estimated to enable the dynamic monitoring of crop growth using UAV multispectral imagery. However, the high spatial resolution of UAV imagery often leads to the misalignment between ground sampling points and corresponding image pixels, even the accuracy of inversion models. This study aims to investigate the optimal spatial window for the UAV-based multispectral inversion of rice chlorophyll content. A DJI Phantom 4-M UAV was employed to obtain the multispectral images from a rice experimental field in the National Agricultural Science and Technology Park of the Jinggangshan in Xingqiao Town, Ji'an City, Jiangxi Province, China. The images were also collected during the boosting, heading, and maturation stages of rice growth, with a uniform resolution of 2.7 cm per pixel. The UAV was equipped with a multispectral sensor consisting of five spectral bands (450 nm (blue), 560 nm (green), 650 nm (red), 730 nm (red edge), and 840 nm (near-infrared)) and a TimeSync time synchronization system with the centimeter-level positioning accuracy. The ground chlorophyll content (SPAD) measurements were obtained concurrently with the UAV multispectral data acquisition using a SPAD-502Plus chlorophyll content meter. The sample area was selected at the center of each rice paddy, where three to five rice plants were sampled. The SPAD value was measured for the upper, middle, and lower sections of each rice leaf. The average value of each rice plant was then determined to represent the SPAD value of the sample. Additionally, the latitude and longitude of the sampling points were recorded using network RTK services. Spatial windows of varying sizes and shapes were employed to process the acquired images. Various vegetation indices were computed using the processed images. The correlation coefficients between the vegetation indices generated with different windows and the ground-measured SPAD values were examined. The spatial window corresponding to the vegetation indices with the highest correlation coefficients was identified as the optimal spatial window. Subsequently, the vegetation indices were selected with the ground-measured SPAD values. The support vector machines (SVM), random forests (RF), extreme learning machines (ELM), generalized linear models (GLM), and multiple linear stepwise regression models (MLSR) were constructed to evaluate the inversion accuracy of SPAD values at distinct rice growth stages. The results showed that: 1) The correlation coefficients between various vegetation indices and SPAD values were significantly improved after processing with a spatial window. In circular spatial windows, the optimal window radius was determined as 35, 25, and 25 pixels for each growth stage, respectively. In square spatial windows, the optimal side length was 71, 41, and 61 pixels for each growth stage, respectively. The results of the square window were very similar to those of the circular window. 2) The support vector machine model demonstrated the highest efficacy in retrieving rice SPAD values. The peak inversion accuracy was achieved during the boosting stage with a coefficient of determination of 0.718, a root mean square error (RMSE) of 1.849, and a mean absolute error of 1.465. The UAV-based multi-spectral data during the boosting stage were input into the SVM model to spatially invert rice SPAD values. The resulting map effectively reflected the fertilization treatment of the experimental field, thereby offering guidance for agricultural production. This finding can be expected to serve as a valuable reference for the selection of spatial windows in the inversion of biochemical parameters of various crops. Furthermore, a technical foundation was established for the UAV-based multispectral monitoring of crop growth, thereby contributing to the advancement of precision agriculture.

The precise estimation of crop yields is essential for global food security, particularly in the face of challenges like climate change, population growth, and food distribution inequalities. Despite the widespread use of machine learning techniques combined with remote sensing data for large-scale yield prediction, the integration of crop spatial position information and local models remains underexplored. This is particularly significant given the spatial nature of crop yield prediction, where spatial factors are highly influential. Previous studies, predominantly conducted on an annual or full-growth season basis, have not provided precise predictions for each phenological stage of maize growth. Consequently, these studies fall short in pinpointing the most effective prediction time for maize yield and understanding the impact of environmental factors at each stage. This research delves into two key questions: 1) Does the inclusion of spatial location information in the geographic weighted random forest (GWRFR) model improve yield prediction accuracy over the traditional random forest model? 2) Among different phenological stages of maize, which stage provides the optimal window for yield prediction? To address these issues, this study employed multi-source remote sensing data in conjunction with machine learning algorithms, and predicted maize yield at the county level in the United States. This study investigated the relationship between yield prediction and the spatial location of sample points, assessing the relevance of including latitude and longitude as independent variables. Further, the study introduced the local GWRFR model for maize yield prediction and compared its modeling performance with the global random forest (RF) model. In addition, the study examined two methodological approaches for determining the best prediction time. The first approach, referred to as the accumulated environmental variables (AEV) approach, integrated data from various phenological periods. The second approach, known as the current stage variables (CSV) approach, used data exclusively from the specific growth stage under analysis. The seven key growth stages of maize included planted, emerged, silking, dough, dent, mature and harvest, providing a comprehensive view of the crop's lifecycle. Through a comprehensive evaluation of the results from both schemes, this study identified the optimal prediction time for maize yield. The findings indicate that incorporating latitude and longitude into the model enhanced yield prediction accuracy. Without these spatial factors, the RF model achieved an coefficient of determination (R²) of 0.83 and root mean squared error (RMSE) of 994.75 kg/hm², while including them improved these metrics to an R² of 0.85 and RMSE of 890.88 kg/hm². This provides preliminary evidence that including spatial factors can enhance maize yield prediction accuracy. Moreover, the local GWRFR model further improved prediction accuracy (R²=0.87, RMSE=864.21 kg/hm²), outperforming the traditional RF model and effectively addressing the non-stationarity of spatial data. In terms of optimal prediction time, the scheme where the environmental variables accumulate over phenological stages showed increasing accuracy from the first stage (planted) up to the fourth stage (dough), peaking at R²=0.90 and RMSE of 748.39 kg/hm², and then stabilized. In contrast, the scheme utilizing only current stage variables improved accuracy from the first stage up to the third stage (silking), reaching its peak (R²=0.88, RMSE=827.85 kg/hm²) before decreasing. This suggests the best prediction time was around dough stage, approximately 2-3 months before harvest. Additionally, the strong correlation observed between early prediction results and those covering the entire growth season underscores the reliability of maize yield predictions made during the dough stages. In conclusion, this study introduces a novel method for large-scale crop yield prediction, integrating spatial data and phenological stages with advanced modeling techniques. The findings significantly contribute to enhancing food security and stabilizing the global food supply chain. This research not only provides critical insights for agricultural practices but also sets a foundation for future studies in crop yield prediction, potentially extending to other crops and regions, and incorporating a broader range of environmental factors.