Abstract: To enable rapid，non‑destructive monitoring of the soil plant analysis development（SPAD）value of field‑grown potatoes，this research employed unmanned aerial vehicle（UAV） hyperspectral imaging to construct a quantitative detection model during critical growth phases.UAV hyperspectral imagery captured during the tuber initiation and enlargement stages was processed using mathematical transformations.Characteristic bands correlating with the SPAD value were identified using the competitive adaptive reweighted sampling（CARS），uninformative variable elimination（UVE），and random frog（RF） algorithms. Subsequently，partial least squares regression（PLSR），support vector regression（SVR），and back propagation（BP）neural networks were utilized to formulate models for estimating the potato SPAD values.It was observed that the characteristic bands derived from distinct feature selection algorithms varied slightly，with the CARS algorithm demonstrating efficiency in extracting sensitive spectral features，reducing hyperspectral data dimensions，and enhancing model precision.Compared to models constructed with alternative algorithmic combinations，the 1/R‑CARS‑SVR model displayed superior predictive capabilities，yielding R ²values of 0.88 for the training set and 0.84 for the validation set，and consistent root mean square error（RMSE） values of 0.39 for both.The 1/R‑CARS‑SVR model was utilized to perform point‑by‑point SPAD value computations across the study area，and a detailed inversion map was generated.It was found that SPAD value in tuber expansion stage was generally higher than that in tuber formation stage.This map offered a visual representation of potato growth conditions for managerial decision‑making，contributing to the theoretical framework and methodological approach for the surveillance of potato growth dynamics.

Key words: Potato, SPAD, UAV hyperspectrum, CARS, SVR

摘要： 为实现大田马铃薯SPAD值的快速、无损检测，采用无人机高光谱成像技术建立马铃薯关键生长时期SPAD值的定量检测模型。在大田条件下获得了马铃薯块茎形成期和彭大期的无人机高光谱影像，并对高光谱数据采用数学变换方法处理；接着，通过竞争性自适应重加权采样（CARS）、无信息变量去除（UVE）和随机蛙跳（RF）算法筛选与SPAD值相关的特征波段；随后，使用偏最小二乘回归（PLSR）、支持向量回归（SVR）和反向传播（BP）3种机器学习算法建立马铃薯SPAD值估算模型。不同算法筛选的特征波段存在差异，CARS算法能有效识别敏感光谱特征，降低特征数量，提升估测精度。基于不同数学变换和特征波段筛选算法建立的马铃薯SPAD值估测模型中，1/R-CARS-SVR模型对马铃薯SPAD值具有较强的估测能力，模型建模集和验证集R²分别为0.88和0.84，RMSE 均为0.39。采用1/R-CARS-SVR模型逐点计算研究区马铃薯SPAD值，绘制了SPAD值反演图，发现块茎膨大期的SPAD值普遍高于块茎形成期。

关键词: 马铃薯, SPAD, 无人机高光谱, 竞争性自适应重算法, 支持向量机