基于无人机高光谱图像的花生苗期株数监测

Estimation of the quantity of peanut seedling plant based on UAV hyperspectral imagery

关键词:  无人机  花生  高光谱图像  群体特征  株数  覆盖度

Key words:  UAV  peanut  hyperspectral imaging  group characteristics  number of plants  coverage

摘要:

【目的】 花生是我国重要的油料作物之一，从无人机遥感影像中准确识别花生植株，实现快速、无损、准确的花生株数统计，对花生田间管理和高产具有重大意义。【方法】 文章以广东省河源市灯塔盆地花生种植基地为研究对象，利用无人机获取花生种植区域的高光谱图像（Hyperspectral Images，HSI）数据，利用提取花生平均覆盖度的方法近似估算花生株数。通过分析390~1 000 nm范围内高光谱图像特征，选取分割提取花生植株的有效特征波段，构建了包括光谱特征（Vegetation Indices，VIs）和群体形态特征（Morphological Parameters，MPs）在内共20个特征参数，并利用皮尔逊相关分析选择了12个参数。然后基于偏最小二乘回归（Partial Least Square Regression，PLSR）、随机森林回归（Random Forest Regression，RFR）和支持向量回归（Support Vector Regression，SVR）3种算法，分别利用光谱特征、群体形态特征、联合光谱特征和群体形态特征构建花生平均覆盖度估测模型，并确定估算花生平均覆盖度的最优估测模型。【结果】 （1）在可见光区域，绿色植株与背景区域的差异明显且不存在重叠区域，671 nm处光谱反射率可有效区分花生植株与背景区域。（2）在使用PLSR、RFR和SVR 3种算法训练的模型中，SVR模型估测结果的精度最佳。（3）在单特征模型中，基于群体形态特征指数的模型精度优于基于最优光谱指数所构建的模型。在多特征模型中，联合光谱特征和群体形态特征所构建的模型精度更高、稳定性更强，PLSR、RFR和SVR 3种模型的R²分别达到0.56、0.58和0.75。其中，SVR模型的RMSE低至0.29株。【结论】 基于SVR算法、光谱特征和群体形态特征所构建的模型能有效监测田间条件下的花生覆盖度水平，适用于基于无人机的花生苗期株数的实时快速检测，可为花生的精准管理生产提供一种快捷高效的技术手段。

Abstract:

Purpose Peanut is one of the important oil crops in china. The ability to accurately identify individual peanut plants within a field and obtain precise plant counts is essential for optimizing field management practices and ultimately enhancing yield.Method This research was conducted in a peanut planting base located in the Dengta Basin of Heyuan City，Guangdong Province，China. Hyperspectral images（HSI）covering the peanut fields were acquired using a UAV platform. Recognizing the challenges of directly identifying individual plants within dense canopies，we developed a novel approach to estimate plant counts by predicting the average peanut coverage within the canopy. To achieve this，we first conducted a thorough analysis the spectral characteristics of peanut plants across the 390-1000 nm spectral range. This analysis aimed to pinpoint specific spectral bands that effectively differentiated peanut plants from the surrounding background，aiding in the segmentation of peanut plants from the imagery. A comprehensive set of 20 features was then meticulously extracted from the hyperspectral data. These features encompassed both vegetation indices（VIs），which capture unique spectral reflectance patterns of vegetation，and morphological parameters（MPs），which quantify the shape and structural characteristics of plant canopies. To streamline model development and focus on the most informative features，Pearson correlation analysis was employed to select the 12 most relevant features from the initial set of 20. With the selected features in hand，three powerful machine learning algorithms，namely partial least square regression（PLSR），random forest regression（RFR），and support vector regression（SVR），were trained and rigorously evaluated to predict the average peanut coverage，which could then be translated into plant count estimates. Each algorithm was trained and evaluated using three distinct feature combinations：spectral indices alone，morphological parameters alone，and a combined set of both spectral and morphological features. This comprehensive approach allowed us to assess the individual contributions of spectral and morphological features，as well as their synergistic potential in predicting peanut coverage.Result （1）In the visible light region，green plants and background areas showed significant differences with no overlapping regions，and the spectral reflectance at 671 nm effectively distinguished peanut plants from the background. （2）Among the three algorithms used for model training （PLSR，RFR，and SVR），the SVR model exhibited the highest estimation accuracy. （3）In single-feature models，the model based on population morphological feature indices demonstrated higher precision compared to models constructed using optimal spectral indices. In multi-feature models，models combining spectral features and population morphological features showed improved accuracy and stability，with R² values of 0.56，0.58，and 0.75 for the PLSR，RFR，and SVR models，respectively. The SVR model achieved the lowest RMSE of 0.29 plants.Conclusion The model constructed using the SVR algorithm，spectral features，and population morphological features effectively monitors peanut coverage levels under field conditions，providing a quick and efficient technical approach for real-time detection of peanut plant numbers using unmanned aerial vehicles，and offers a basis for precise peanut production management.