Attention-guided UAV model trims maize maturity inputs and improves prediction

An unmanned aerial vehicle equipped with a four-band multispectral camera can estimate maize maturity from canopy imagery with fewer than one-quarter of the features used in a conventional pooled model, according to a 2025 field study in Shandong Province, China. Researchers combined 15 vegetation indices and 32 canopy-texture descriptors, then used a radial basis function attention mechanism to identify the 10 most informative predictors. When paired with a random forest, the selected features produced a test-set coefficient of determination, R², of 0.7713 and a root mean square error of 10.15 maturity-index points. The result came from 252 observations collected across six dates from two maize cultivars, and the authors stress that the findings are not yet a general test across sites or seasons.

The study began by building a maize maturity index that could serve as the target for machine-learning models. The authors combined canopy chlorophyll status, measured with a SPAD-502Plus meter; leaf area index, estimated from leaf length and width in three canopy strata; and grain moisture content, calculated from fresh and oven-dry grain weights. These indicators were standardized and analyzed with principal component analysis. The first component explained 71.31% of the total variance and had positive loadings for all three measures, capturing their shared decline as the crop matured. Its scores were reverse-scaled to produce a Canopy Maturity Index, or CMI, from 0 to 100, with larger values indicating greater maturity. The resulting index had a mean of 47.42 and a standard deviation of 20.80 and covered a broad range of the experimental maturity gradient.

Synchronized field measurements and drone flights were conducted on September 7, September 15, September 22, September 28, October 9 and October 14. The experiment comprised 42 plots of 54 square meters at the Zaoyuan Experimental Base in Zhangqiu District, Jinan. The two cultivars were Zhengdan 958, described as conventional, and Jingnongke 728, described as early-maturing, planted at 75,000 plants per hectare with 60 centimeters between rows. Flights took place under clear, calm conditions from 10:00 to 14:00, using a DJI Mavic 3 M at an altitude of 20 meters. The images produced an approximately 5-centimeter ground sampling distance, were mosaicked and radiometrically calibrated, and were divided into plot-level regions of interest for feature extraction.

The candidate inputs included spectral information that reflects canopy reflectance and spatial texture that describes how those reflectance patterns are arranged. Vegetation indices were calculated from green, red, red-edge and near-infrared bands, while gray-level co-occurrence matrices supplied 32 texture descriptors: eight measurements from each spectral band. Six regression algorithms were tested, including random forest, extreme gradient boosting, LightGBM, gradient-boosted decision trees, partial least-squares regression and support vector regression. On average across the six algorithms, the vegetation-index-only configuration achieved an R² of 0.6336 and RMSE of 12.18, while texture alone produced an R² of 0.5272 and RMSE of 13.59. Pooling all 47 predictors gave a mean R² of 0.6348 and RMSE of 12.07, indicating that spectral information carried most of the signal and that indiscriminate fusion brought only a modest gain.

To reduce redundancy, the team compared four ways of ranking features: Pearson correlation coefficients and three attention mechanisms named SENet, GateAttention and RBFAttention. Each method was allowed to select its 10 highest-ranked features from the same training set, and a sensitivity analysis showed that accuracy rose rapidly up to roughly eight to 10 predictors before flattening. RBFAttention performed best, with a mean cross-validated R² of 0.6480 and RMSE of 11.9247 across the six regressors. The corresponding Pearson results were 0.6271 and 12.3087. RBFAttention used a Gaussian radial basis kernel to model nonlinear relationships among features, and the researchers selected it for subsequent optimization. Its top-ranked variables were NDI, NDVIrededge, CCCI, RVI and near-infrared homogeneity, followed by CIrededge, CARI, GNDVI, CIgreen and MSAVI. Nine of the ten were vegetation indices, and only one was a texture measure.

The researchers then tested whether manually balancing spectral and structural information would improve on the attention-ranked subset. A configuration with eight vegetation indices and two texture features achieved a mean R² of 0.6682, while a balanced method selecting five features from each source reached 0.6592. A texture-dominated subset with two vegetation indices and eight texture features performed worst, at 0.6544. The original RBFAttention top-10 set remained strongest, with a mean R² of 0.6789 and RMSE of 11.4587. Because the default set was also selected without prescribing a source ratio, the authors advanced it alongside the more functionally diverse balanced set for final model tuning. Hyperparameters were optimized with Optuna using five-fold cross-validation and as many as 100 trials per model.

On the 76-sample test set, the RBFAttention-selected random forest achieved the best reported result: R² of 0.7713 and RMSE of 10.15 CMI points. It exceeded a vegetation-index-only baseline with R² of 0.7455 and RMSE of 10.38, a texture-only model with R² of 0.7023 and RMSE of 11.58, and an all-feature model with R² of 0.7570 and RMSE of 10.23. Crucially, the optimized model used 10 predictors rather than 47, a reduction of 78.7%. Its residuals averaged -1.89 with a standard deviation of 9.97, and a Shapiro-Wilk test found no significant departure from normality. In total, 57.9% of test predictions were within 10 CMI points of the measured values and 78.9% were within 15 points. Accuracy was strongest at intermediate maturity, where RMSE was 7.31, compared with 10.76 at low maturity and 13.06 at high maturity, where relatively few training samples were available.

Repeated cross-validation and audits of the data-splitting procedure tempered the initially strong single-split result. Across five repetitions of five-fold validation, the balanced-feature random forest had a mean R² of 0.6881 plus or minus 0.0928, while the RBFAttention random forest reached 0.6813 plus or minus 0.0919. The difference between them was not statistically significant after Bonferroni correction. Tree-based ensembles generally outperformed support vector regression. More importantly, a plot-grouped holdout that kept repeated observations from the same plot together produced R² of 0.709 plus or minus 0.022 and RMSE of 11.18 plus or minus 0.79, compared with 0.743 plus or minus 0.015 and 10.68 plus or minus 0.48 for random sample splits. The roughly 0.03 difference in R² suggests some optimism from the sample-level split, and the authors present the reported 0.7713 as an upper end of a validated range of approximately 0.71 to 0.77.

SHAP analysis showed that the compact model remained closely tied to recognizable crop physiology. The Canopy Chlorophyll Content Index, or CCCI, made the largest contribution, accounting for 24.04% of the total absolute SHAP contribution, followed by NDVIrededge at 20.90% and CIrededge at 19.40%. Together, these highly correlated red-edge-related indices accounted for 64.34% of the contribution and were best interpreted as a shared senescence signal rather than three independent effects. Low CCCI and red-edge values increased predicted maturity, matching the loss of chlorophyll during ripening, while near-infrared homogeneity had more complex, bidirectional effects. Applied to the drone imagery, the final model generated pixel-level maturity maps and distinguished spatial differences between the two cultivar blocks. The study nevertheless remains limited to one site, one season and two cultivars; temporal extrapolation, later-season sampling, and performance under other management and environmental conditions still require multi-site, multi-year validation.

Subject of Research: UAV multispectral imaging and attention-based feature selection for estimating maize maturity from canopy physiological indicators.

Article Title: UAV-based maize maturity estimation: Multi-source feature fusion with attention mechanism optimization

Article References: Wang, X., Wang, X., Gong, K., Song, X., Yang, Z., Zhang, X., Zhang, X., & Zheng, J. (2026). UAV-based maize maturity estimation: Multi-source feature fusion with attention mechanism optimization. Smart Agricultural Technology, 15, Article 102657. https://doi.org/10.1016/j.atech.2026.102657

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102657

Keywords: maize maturity, UAV remote sensing, multispectral imagery, attention mechanisms, feature selection, random forest, crop monitoring, principal component analysis, SHAP interpretation, precision agriculture, UAV-based, maize