Multispectral UAV imagery and machine learning for morphological differentiation of two tropical C4-forage grasses in an integrated crop-livestock system in the State of Goias, Brazil

<b>Context.</b> Mixed tropical pastures containing both bunchgrasses (<i>Megathyrsus maximus</i>) and lightly sod-forming grasses (<i>Urochloa bryzantha</i>) present significant management challenges due to their contrasting growth habits and grazing requirements. Traditional visual assessment methods are inadequate for large-scale monitoring of grass growth forms. <b>Aims.</b> To develop and evaluate an automated system for detecting and mapping bunchgrass distribution in tropical pastures using multispectral UAV imagery and machine learning algorithms, enabling precision grazing management. <b>Methods.</b> Multispectral imagery (green, red, red edge, NIR bands) was acquired using a DJI Mavic 3M drone over 6.30 hectares of mixed pasture. We extracted 65 features, including spectral, vegetation indices, height, and metrics for Gray-Level Co-occurrence matrix (GLCM) texture. Three machine learning algorithms (Random Forest – RF, SVM, XGBoost) were compared using 200,000 annotated pixels. Feature selection experiments evaluated accuracy-efficiency trade-offs. <b>Key results.</b> XGBoost achieved superior performance (F1-score = 0.69), representing 9.73 % improvement over RF. Red edge-based GLCM texture features were the most important predictors. Feature-optimized models using only 15 features retained 91.3 % of full model performance while reducing processing time by 77 %. Fast spectral-only models achieved 400× speedup with 67.9 % performance retention. Spatial analysis revealed 28.14 % bunchgrass coverage across the study area. <b>Conclusions.</b> Machine learning, particularly XGBoost with texture features, effectively discriminates between tropical grass growth forms. Feature selection enables practical deployment without substantial accuracy loss. <b>Implications.</b> This methodology enables objective assessment of pasture composition, supporting precision grazing strategies, targeted fertilization, and the sustainable intensification of tropical livestock systems. The approach is scalable for operational deployment across extensive ranching operations.<br>For instructions download attached word files above.

<b>研究背景</b>：同时包含丛生禾草（<i>Megathyrsus maximus</i>）与弱铺地型禾草（<i>Urochloa bryzantha</i>）的混合热带牧场，因二者生长习性与放牧需求差异显著，面临诸多管理难题。传统目视评估方法难以支撑牧草生长型的大规模监测工作。<b>研究目的</b>：开发并评估一套基于多光谱无人机（UAV）影像与机器学习算法的自动化系统，用于热带牧场丛生禾草的分布检测与制图，以实现精准放牧管理。<b>研究方法</b>：本研究采用DJI Mavic 3M无人机，在6.30公顷的混合牧场区域采集了涵盖绿光、红光、红边与近红外（NIR）波段的多光谱影像。共提取65类特征，包括光谱特征、植被指数、株高特征以及灰度共生矩阵（GLCM）纹理特征。基于20万个标注像素，对比了随机森林（RF）、支持向量机（SVM）与极限梯度提升树（XGBoost）三种机器学习算法的性能。通过特征选择实验，评估了模型精度与运算效率之间的权衡关系。<b>主要结果</b>：极限梯度提升树（XGBoost）表现最优，F1值达0.69，较随机森林（RF）提升9.73%。基于红边波段的灰度共生矩阵纹理特征为最重要的预测因子。仅使用15个特征的优化模型保留了全特征模型91.3%的性能，同时将处理时间缩短77%。仅采用光谱特征的轻量化模型实现了400倍的运算加速，且保留了67.9%的模型性能。空间分析结果显示，研究区域内丛生禾草覆盖率为28.14%。<b>研究结论</b>：机器学习方法，尤其是结合纹理特征的极限梯度提升树模型，可有效区分热带牧草的生长型。特征选择技术能够在不造成显著精度损失的前提下，实现模型的实际部署应用。<b>研究意义</b>：本研究方法可实现牧场群落组成的客观评估，支撑精准放牧策略、靶向施肥以及热带畜牧系统的可持续集约化发展。该方法具备良好的可扩展性，可在大规模牧场运营中实现落地应用。操作指南请下载上方附件中的Word文件。