Enhancing the accuracy of monitoring effective tiller counts of wheat using multi-source data and machine learning derived from consumer drones

Abstract

The effective tiller number of wheat (ETNW) is one of the three main factors affecting wheat yield. Most traditional methods for counting tillers are manual, which is inefficient and challenging to implement on a wide scale. Existing remote sensing techniques for tiller counting are typically focused on individual plants or small plot areas, often utilizing high-resolution sensors for close-range monitoring. This approach limits the scalability and applicability for large-scale field environments. Moreover, most studies on wheat tiller monitoring have concentrated on specific crop varieties or single-variable conditions, with little research conducted on estimating or monitoring ETNW in large-scale field scenarios using consumer-grade unmanned aerial vehicle (UAV) platforms. To address these issues, we propose a multi-modal fusion-driven machine learning method to enhance the performance of wheat tillering monitoring. This study was conducted using two experimental setups to ensure robust and comprehensive data collection. Experiment 1 (Exp.1) focused on water and nitrogen coupling conditions, while Experiment 2 (Exp.2) included both nitrogen-deficient and nitrogen-sufficient treatments. Each experiment involved multiple wheat varieties to account for genotypic variability. UAV data, including multispectral and RGB imagery, was collected across different growth stages under varying irrigation and nitrogen conditions to ensure the generalizability of the proposed model. This method employs spectral correlation analysis (SCA) to select features strongly correlated with the number of tillers and subsequently fuses these features to construct a multi-modal machine learning model based on vegetation index (VI), color index (CI), multispectral texture features (TF1), and RGB texture features (TF2) derived from UAV data. Three machine learning models are applied: Random Forest Regression (RFR), Partial Least Squares Regression (PLSR), and Support Vector Regression (SVR). The results demonstrated that the vegetation index-driven machine learning model (VI-RFR) achieved the best performance, with an R² of 0.85, RMSE of 348.60, and NRMSE of 0.367, followed by the color index model (CI-RFR, R² = 0.83, RMSE = 313.18, NRMSE = 0.330), the multispectral texture feature model (TF1-RFR, R² = 0.81, RMSE = 357.12, NRMSE = 0.376), and the RGB texture feature model (TF2-RFR, R² = 0.80, RMSE = 373.69, NRMSE = 0.393). Compared to using a single data source, data fusion significantly improved model accuracy, particularly when complementary data sources were combined. Specifically, the VI&CI combination achieved an R² improvement of 6.4 %-19.1 %, an RMSE reduction of 2.8 %-9%, and an NRMSE decrease of 1.56 %-8.24 %. The VI&TF1 combination exhibited an R² increase of 7.8 %–32.6 %, an RMSE reduction of 8.5 %-28.8 %, and an NRMSE decrease of 2.20 %-8.24 %. The VI&TF2 combination showed an R² increase of 3.8 %-40.5 %, an RMSE reduction of 2.8 %-27.6 %, and an NRMSE decrease of 2.68 %-21.61 %. The CI&TF1 combination resulted in an R² improvement of 5.2 %-25.6 %, an RMSE reduction of 7.4 %-24.1 %, and an NRMSE decrease of 2.04 %-19.41 %. The CI&TF2 combination achieved an R² increase of 12.5 %–33.3 %, an RMSE reduction of 5.4 %–23.4 %, and an NRMSE decrease of 1.17 %-18.94 %. The TF1&TF2 combination achieved an R² improvement of 13.9 %–33.3 %, an RMSE reduction of 10.9 %-27.6 %, and an NRMSE decrease of 5.01 %-21.61 %. However, increasing the number of data sources does not necessarily lead to higher model accuracy. The VI&CI&TF1 fusion model (RFR, R² = 0.90, RMSE = 288.39, NRMSE = 0.304) demonstrated the best performance, surpassing the combination of all four features. The machine learning models exhibited systematic variations, with the performance ranking as follows: RFR > SVR > PLSR, regardless of the data fusion strategy employed. Moreover, the best model (VI&CI&TF1-RFR) demonstrated adaptability across different growth stages, irrigation conditions, and nitrogen fertilizer applications. Notably, the model maintained robustness even in extreme environments with complete nitrogen deficiency. The findings of this study provide technical support for crop phenotyping, variety selection, and precision agriculture management.

Plant phenotyping relevance

UAVのマルチスペクトル・RGB画像から小麦の有効分げつ数を推定する特徴融合・機械学習手法を開発し、複数条件・品種で性能評価しており、植物表現型の取得・推定が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.