Time-series data inversion of soil moisture content in root zone of kiwifruit using BiLSTM
-
摘要:
根域土壤水分是决定猕猴桃树健康生长与产量的关键因素,尤其在果实膨胀期,土壤水分的动态监测尤为重要。针对传统监测方法无法监测土壤水分持续变化,该研究以眉县猕猴桃实验站为研究区域,采用无人机和地面传感器采集植被光谱反射率及土壤水分数据(共60 d,
1440 组数据),构建猕猴桃根域土壤含水率的反演模型。通过Pearson和Spearman相关系数筛选了9种植被指数作为模型输入,比较了前馈神经网络(feedforward neural network,FFNN)、长短期记忆网络(long short-term memory,LSTM)及双向长短期记忆网络(bidirectional long short-term memory,BiLSTM)的表现。FFNN由于无法吸收时间序列信息,其在测试集上的表现较差,决定系数为0.269,均方根误差为3.56%。而LSTM和BiLSTM模型利用多日历史数据显著提高预测精度,其中BiLSTM表现最佳,测试集决定系数为0.624,均方根误差为2.45%。研究表明,基于时序模型的土壤水分反演方法可以用于猕猴桃果园果实膨大期的精准监测,也为其他果园作物的水分管理提供一定的理论支持。Abstract:Soil moisture content in the root zone is one of the most crucial factors in determining the healthy growth and yield of kiwifruit trees, particularly in dynamic monitoring of the soil moisture during fruit expansion. However, traditional monitoring cannot capture the continuous variation in soil moisture. In this study, a time-series data inversion was carried out on the soil moisture content in the root zone of kiwifruit. The fruits were taken from the Meixian kiwifruit experimental station in Baoji City, Shaanxi Province, China. Both unmanned aerial vehicles (UAV) equipped with multispectral sensors and ground-based moisture sensors were utilized to collect the spectral reflectance and soil moisture data over a period of 60 days, respectively, resulting in a total of 1440 datasets. The spectral data was processed to initially extract 20 vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), and Green Normalized Difference Vegetation Index (GNDVI). The dataset was then refined to identify the most critical features for the soil moisture inversion. Pearson and Spearman correlation coefficients were employed to determine the nine most relevant vegetation indices. The indices were also optimized to reduce the model complexity for the high predictive power. Subsequently, the optimal indices were used as the inputs for three machine learning models: the Feedforward Neural Network (FFNN), the Long Short-Term Memory (LSTM) network, and the Bidirectional Long Short-Term Memory (BiLSTM) network. Among them, the FFNN served as a baseline to compare with the temporal models, due to the temporal independencies in the data. The experimental setup involved training and testing the three models using the dataset. The FFNN model was trained with the input features representing the selected vegetation indices without any temporal information. In contrast, the LSTM and BiLSTM models were designed to utilize the multi-day historical data, thus capturing the temporal dependencies in the soil moisture dynamics. The LSTM model was particularly effective in retaining the information over longer sequences, due to its memory cell architecture. While the BiLSTM extended the capability to consider both forward and backward temporal dependencies, thus providing a more comprehensive understanding of temporal relationships. The results indicated that the LSTM and BiLSTM models with the multi-day historical data significantly improved the prediction accuracy. Specifically, the LSTM model was achieved in the test dataset coefficient of determination (R2) value of 0.548 and a root mean square error (RMSE) of 2.68%, indicating the more effective temporal dependencies, compared with the FFNN. The FFNN model exhibited a relatively low performance on the test dataset, with an R² value of 0.269 and an RMSE of 3.56%, due to its inability to capture time-series information. Among them, the BiLSTM performed the best, thus achieving a test dataset R² value of 0.624 and an RMSE of 2.45%. This superior performance of the BiLSTM was attributed to both past and future contexts within the sequence, which enhanced its predictive capability for soil moisture dynamics. The better performance of temporal models was achieved in the dynamic prediction of soil moisture, compared with the non-temporal models. The LSTM and BiLSTM were incorporated with the temporal data for more accurate modeling of the complex interactions between vegetation indices and soil moisture content. Particularly, the precise prediction of soil moisture during fruit expansion was crucial to optimize the yield and quality. In conclusion, the time-series LSTM and BiLSTM models were proved to effectively monitor the soil moisture in the period of the kiwifruit fruit expansion. This approach can also offer valuable theoretical support to the water practices in orchards. The promising potential can integrate the multispectral UAV remote sensing data with deep learning in smart agriculture. Advanced temporal modeling can be expected to enhance the monitoring precision of soil moisture in sustainable and efficient crop production.
-
Keywords:
- UAV /
- kiwifruit /
- soil moisture /
- multispectral /
- remote sensing /
- feedforward neural network /
- long short-term memory network
-
0. 引 言
猕猴桃是陕西省的重要地标产品,种植面积和产量均居陕西前列。在规模化发展中,水分管理是关键“因素”之一[1-3]。有效的水分管理在果实膨大期起到关键作用。猕猴桃对水分需求较高,土壤水分不足会严重影响果实生长和单果质量,导致产量与品质下降,削弱市场竞争力[4-5]。传统的土壤水分采集手段主要有直接取样法、水分传感器法和重量法[6],不仅数据采集效率较低,且对土壤有一定程度的破坏,从而无法大规模的对果园进行有效的水分管理,因此需要找到一种连续、准确、灵活的土壤水分监测系统。
近年来,大规模的土壤水分监测主要是通过无人机遥感和卫星遥感获取微波、热红外和光学遥感数据结合机器学习去反演土壤水分[7]。其中微波遥感技术在土壤水分监测领域取得了显著进展,特别是合成孔径雷达(synthetic aperture radar,SAR)技术的发展,使其在探测一定深度的根域土壤水分方面展现出巨大的潜力,但微波遥感常搭载卫星无法高精度且连续的监测特定区域的土壤水分[8]。相比之下,无人机光学遥感技术因其高效、便捷、覆盖范围广等特点已广泛应用果园的土壤水分监测。靳亚红等[9]利用无人机多光谱遥感数据,通过皮尔逊相关系数和XGBoost模型反演玉米和小麦的土壤含水率。喷灌条件下的最优光谱窗口较畦灌小,玉米灌浆期的R2分别为0.80(喷灌)和0.83(畦灌),RMSE为1.27%和0.98%;小麦苗期的R2为0.76(喷灌)和0.79(畦灌),RMSE分别为1.68%和0.85%。该研究为基于无人机遥感的土壤水分监测提供了有效的方法。在果园方面,DENG等[10]通过多层感知器模型结合冠层植被指数评估猕猴桃果园的土壤水分,R2为
0.744; ZHU等[11]通过利用无人机多光谱影像特征与集成学习方法,结合最优波段组合算法,评估了猕猴桃冠层反射率和植被覆盖度特征对根域土壤水分的敏感性。这些为果园中精确的土壤水分监测提供了一种可行的技术路径。同时深度学习的模型改进也成为研究的关键方向。潘时佳等[12]使用无人机搭载多光谱相机获取的数据,通过MLP网络和CVCRNet模型,探索了植被指数与土壤含水率的关系。综上所示,尽管当前利用无人机多光谱成像技术应用在土壤水分监测中已经取得了显著进展[13-15],但大多数现有方法主要集中在图像信息的处理和模型的改进,然而土壤水分监测是一种动态变化,则需要通过输入信息的时间序列来动态反演果园土壤水分的变化。该研究针对目前存在的问题,特别是在时序问题的处理和土壤水分监测的连续性方面的不足,以眉县猕猴桃果实膨大期为研究对象,提出了一种基于无人机遥感技术与时序模型结合的猕猴桃根域土壤水分时序反演方法,获取根域土壤水分的动态变化情况,同时可以为其他的果园作物的水分管理提供一定的参考。
1. 材料和方法
1.1 试验区域
试验地点位于陕西省宝鸡市眉县西北农林科技大学猕猴桃农业气象科技小院,北纬107°59′,东经34°7′,海拔 643.22 m。四季分明,夏季炎热,冬季寒冷,多年平均降水量为 609.5 mm,年均日照时数为
2015.2 h。猕猴桃果树每年6月初进入果实生长期,冠层生长完全覆盖地面,对水分和养料的需求显著增加。区域总面积约6000 m2 ,试验区域面积约 300 m2 ,品种为徐香猕猴桃,共 4 行 6 列,株距约 3 m,行距约 4 m,分布如图1所示,从早春萌芽前1个月开始到萌芽后约2个月进行剪枝管理,同时果园内架设钢丝以供猕猴桃藤蔓生长缠绕。试验于 2023 年 6 月 7 日开始, 60 d连续监测。为了保证试验的一致性,测量时间为 11:30—12:00 ,试验期间根据土壤表面视觉观察情况进行灌溉。每棵猕猴桃果树的每日遥感数据和土壤含水率为一组基础数据,共有基础数据
1440 组,将原始数据集随机划分,其中 80% 的数据作为训练集,20% 的数据作为测试集。1.2 地面土壤水分数据收集
土壤水分测量采用土壤温度水分变送器(485型),其原理利用传感器测量土壤中的温度和水分,通过数据转换器将测量结果以模拟信号(如4~20 mA)传输至记录仪。土壤含水率的计算式:
土壤含水率(%)=(土壤水分体积土壤体积)×100% (1) 首先,选择尽量靠近待测猕猴桃主干距离 1 m 内且深度0.4 m处埋置监测仪探针,该区域位于猕猴桃果树根域深度的水分活跃层(0.2~0.8 m),该层受灌溉和降雨影响显著,且深度大于0.2 m,受大气环境影响较小。每棵猕猴桃果树埋置一根探针,该相对位置如图1所示。每次测量重复3次,去除异常值后取平均值,作为被测果树的根域土壤水分。试验期间各组平均土壤含水率范围为 22.2%~42.5%。分别在6月17日、6月28日、7月03日、7月14日分别有降水出现,实际根域土壤水分变化情况如图2(选择试验区域内5棵猕猴桃根域土壤含水率做为代表)。
1.3 无人机遥感数据采集与预处理
1.3.1 遥感数据采集
遥感图像的采集采用搭载RTK模块的DJI Mavic 3多光谱版无人机,共集成 4 个多光谱镜头,分别对应绿波段(560±16)nm、红波段(650±16)nm、红边波段(730±16)nm、近红外波段(860±26)nm,为确保区域图像边缘的完整性,航拍区域覆盖整个试验区域适当扩大。每次飞行持续 276 s ,覆盖面积为 964 m2 ,并与地面土壤水分观测同步进行。飞行高度 15 m ,巡航速度 5 m/s ,航拍旁向重叠率和航向重叠率均为 80% ,拍摄方向为竖直向下。在无阴影遮挡的位置放置MAPIR多光谱校准标定板,以确保数据的准确性。获得的原始多光谱图像通过Pix4D Mapper进行几何校正和辐射校正,生成试验地多光谱正射影像。
1.3.2 光谱反射率提取
生成多光谱正射影像以后需要对单株猕猴桃多光谱感兴趣区域范围进行平均光谱反射率提取,但由于猕猴桃果园的冠层信息分布与大田作物存在显著差异。其冠层结构复杂,冠层信息与根域土壤水分在空间上无法直接对应。对于小麦、玉米等大田作物(图3a)其冠层数据与根域土壤的位置几乎垂直对应,而猕猴桃果树的冠层覆盖范围较大,且存在相互交叉现象(图3b),传统的采样方式难以准确获取目标区域的植被指数,导致与根域土壤水分的相关性降低。
为了确定单棵猕猴桃树的冠层分布范围并解决冠层相互交叉的问题,冠层范围采样法融合了冬季骨架信息与夏季冠层叶片信息,对猕猴桃冠层进行多级综合分析,从而提取单株植株的冠层范围。骨架预测和图像合并技术,即使在冠层密集且存在交叉的果园中,也能有效地分离出单个冠层[16]。最后,通过对传统矩形采样法(图4a)和冠层范围采样法(图4b)进行对比,验证了两种方法在提取平均光谱反射率计算出的植被指数与根域土壤水分相关性方面的效果差异,从而选取更优的采样方法去提取四波段平均光谱反射率(图4)。
1.3.3 植被指数(vegetation index,VI)特征提取
在提取单株猕猴桃的四波段平均反射率以后,为得到猕猴桃根域土壤含水率与遥感数据之间的确切关系,依据表1中的计算式初步选择20种对土壤水分较为敏感的植被指数[17-23]作为猕猴桃根域土壤水分模型的原始数据。
表 1 植被指数计算式Table 1. Vegetation index formula植被指数Vegetation index(VI) 计算式Formula 参考文献 Reference 绿色指数Green index(GI) ρGRE/ρGREρREDρRED [17] 修改型土壤调整植被指数Modified soil adjusted vegetation index(MSAVI) 0.5×(2×ρNIR+1−√(2×ρNIR+1)2−8(ρNIR−ρRED)) [17] 绿度归一化植被指数Green normalized difference vegetation index(gNDVI) (ρNIR−ρGRE)/(ρNIR−ρGRE)(ρNIR+ρGRE)(ρNIR+ρGRE) [18] 归一化植被指数Normalized difference vegetation index(NDVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED)(ρNIR+ρRED) [18] 优化土壤调节植被指数Optimized soil adjusted vegetation index(OSAVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED+0.5)(ρNIR+ρRED+0.5) [18] 重归一化植被指数Renormalized difference vegetation index(RDVI) (ρNIR−ρGRE)/(ρNIR−ρGRE)(√ρNIR+ρGRE)(√ρNIR+ρGRE) [18] 土壤调整植被指数Soil adjusted vegetation index(SAVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED+L)(ρNIR+ρRED+L)×(1+L)* [19] 简单比值指数Simple ratio index(SR) ρNIR/ρNIRρREDρRED [19] 绿色叶绿素植被指数Green chlorophyll vegetation index(GCVI) ρNIR/ρGRE−1 [20] 植物衰老反射指数Plant senescence reflection index(PSRI) (ρRED−ρGRE)/ρREG [20] 红边归一化差异植被指数Red-edge normalized difference vegetation index(NDVIre) (ρNIR−ρREG)/(ρNIR+ρREG) [21] 修正型红边简单比值指数Modified red-edge simpleratio index(MSRre) (ρNIR/ρREG−1)/(√ρNIR/ρREG+1)(√ρNIR/ρREG+1) [21] 红边叶绿素指数Red-edge chlorophyll index(CIre) ρNIR/ρREG−1 [21] 红边简单比率指数Red-edge simple ratio index(SRre) ρNIR/ρREG [21] 双波段增强植被指数Two-band enhanced vegetation index(EVI2) 2.5×(ρNIR−ρRED)/2.5×(ρNIR−ρRED)(ρNIR+2.4ρRED+1)(ρNIR+2.4ρRED+1) [22] 陆地叶绿素指数MERIS terrestrial chlorophyll index(MTCI) (ρNIR−ρREG)/(ρNIR−ρREG)(ρREG+ρRED)(ρREG+ρRED) [22] 归一化差值水体指数Normalized difference water index(NDWI) (ρGRE−ρNIR)/(ρGRE−ρNIR)(ρGRE+ρNIR)(ρGRE+ρNIR) [23] 全球环境监测指数Global environment monitoring index(GEMI) eat×(1−0.25×eat)−(ρRED−0.125)/−(ρRED−0.125)(1−ρRED)(1−ρRED)eat=(2×(ρ2NIR−ρ2RED)+1.5×ρNIR+0.5ρRED)/+1.5×ρNIR+0.5ρRED)(ρNIR+ρRED+0.5)(ρNIR+ρRED+0.5) [23] 红边三角植被指数Red-edge triangular vegetation index(RTVICore) 100×(ρNIR−ρREG)−10×(ρNIR−ρGRE) [24] 修正型三角植被指数Modified triangular vegetation index(MTVI2) 1.5×(1.2×(ρNIR−ρGRE)−2.5×(ρRED−ρGRE))×√(2×ρNIR+1)2−(6×ρNIR−5√ρRED)−0.5 [24] 注:ρGRE为绿波段平均反射率;ρRED为红波段平均反射率;ρREG为红边波段平均反射率;ρNIR为近红外波段平均反射率。L为土壤亮度校正因子,根据覆盖量此处取L=0.90。 eat为中间变量,通过近红外和红光波段的值按照特定公式计算得出。
Note: ρGRE is the average reflectance of green band; ρRED is the average reflectance of the red band; ρREG is the average reflectance of the red-edged band; ρNIR is the average reflectance of near infrared band. L is the correction factor for soil brightness, L=0.90 is taken here according to the coverage amount. eat is the intermediate variable, which is calculated by the values of near-infrared and red wavelengths according to a specific formula.1.3.4 植被指数(VI)特征筛选
为筛选出与根域土壤含水率具有显著相关性的植被指数特征,并有效提升模型的预测性能和稳定性,同时减少过拟合风险,综合运用了Pearson相关系数和Spearman秩相关系数进行特征筛选[24]。在Pearson和Spearman分析中,筛选标准设定为P ≤ 0.05,表示相关性具有统计显著性。具体而言,Pearson相关系数用于评估植被指数与根域土壤水分之间的线性关系,而Spearman秩相关系数则用于检测非线性关系。通过同时应用这两种方法,能够确保所选特征在不同类型的关系下均具有显著统计学意义,从而减少依赖单一方法可能带来的偏误。各特征的相关性系数绝对值越大,表明该特征对模型的重要性越高,有助于提高模型的稳定性与预测准确性。
1.4 模型构建
1.4.1 基于非时序网络的根域土壤含水率反演模型
为对比非时序网络与时序网络的性能差异,首先引入传统的前馈神经网络(feedforward neural network,FFNN)进行反演。根据第 1.3.4 节筛选出植被指数作为模型输入,以对应的根域土壤含水率为网络输出,网络基本结构如图5 所示。
为了获得最优模型,确保模型的损失函数值下降到较低水平并收敛,设定训练批次(batch size)为 10,选择 Adam 优化器,初始学习率设置为 0.001,学习率衰减范围设置为 0.9~0.999。
选取均方误差(mean square error,MSE)作为损失函数,其计算式:
MSE=N∑i(yturei−yprei)2N (2) 式中yprei表示样本i的估测值;yturei表示样本i的真实值,N表示变量的个数。
使用不同参数选择下的最优迭代结果,对训练集和测试集样本的根域土壤含水率估算结果进行对比,从而确定建立的非时序FFNN网络结构网络类型为全连接网络,第一层节点数50,中间层结构(100,5),输出层节点数为1,激活函数为ReLu。
1.4.2 基于时序网络的根域土壤含水率反演模型
为更好地捕捉猕猴桃果园中土壤水分的时序变化规律,本节在FFNN网络结构基础上进行改进,将其输入层替换为LSTM(long short-term memory)层或BiLSTM(bidirectional long short-term memory)层[17-23]。LSTM和BiLSTM本身是专为处理时间序列数据而设计的模型,能够有效地捕捉土壤水分在时间维度上的动态变化。LSTM层能够记住并传递长时间间隔内的重要信息,而BiLSTM层则通过同时从过去和未来的时间点提取信息,进一步提升了模型的时序捕捉能力,该节为了避免过拟合在原基础上中增加了多个Dropout层,以0.5的概率随机舍弃部分全连接层参数,建立的LSTM网络(图6a)和BiLSTM网络(图6b)结构。
对于时序反演模型主要需要定义两个超参数:历史关注步长(input_steps)和预测序列长度(predict_steps),其中,input_steps为模型输入的数据天数;predict_steps 为模型需要预测的时间步数,input_steps设置为3~7,predict_steps 设置为 1。
利用滑动窗口方法在时序数据上构建样本,如图7 所示。具体过程如下,以 input_steps 为 3作为模型输入,以当日的根域土壤含水率为输出为例:
构建第一个样本:从第一个时间点开始,取前 3 d的植被指数作为输入特征,第三天的根域土壤含水率作为目标值:features = [(V1),(V2),(V3)],target = (D3)滑动窗口移动:features = [(V2),(V3),(V4)], target = (D4)最终获得 58 组数据。
1.5 训练平台与精度评估
该试验在 Windows 10 专业版操作系统(Microsoft)上进行。系统硬件配置包括 NVIDIA GeForce RTX
3060 GPU、AMD Ryzen 7 5800H 处理器以及 32GB DDR4 内存。软件环境方面,使用 Anaconda3.0 搭建 TensorFlow 框架,并在 Python 3.8.13 版本下进行网络的构建与训练。为了评估模型的预测精度,采用决定系数和均方根误差作为评价指标,其计算式:R2=∑(ypre−ˉy)2∑(ytrue−ˉy)2 (3) RMSE=√∑(ypre−ytrue)2N (4) 式中ytrue表示真实值;ypre表示估测值,ˉy表示全部真实值的均值。
2. 结果与分析
2.1 不同光谱反射率采样法结果对比
为了具体分析不同采样法的显著性,将分析结果汇总得到表2,在使用Pearson和Spearman相关系数进行评估时,传统矩形采样法有8种植被指数表现出显著性,而冠层范围采样法则有9种植被指数具有显著性,在这些植被指数中,冠层范围采样法与土壤含水率的相关性较高,因此该研究冠层采样法进行非时序模型构建。选择9种(SR、NDVI、GI、OSAVI、MSAVI、SAVI、PSRI、EVI2和MTCI)植被指数作为非时序模型的原始输入。此外,为了减少试验结果的偶然性,还将在9种的基础上额外选取6种(GCVI、GEMI、MTVI2、gNDVI、RDVI和NDWI)仅在线性相关性分析中具有显著性的植被指数作为对比输入(方案2),从而确定非时序与时序模型的最终输入。
表 2 基于不同采样方法的20种植被指数Pearson与Spearman相关性分析表Table 2. Pearson and Spearman correlation analysis table of 20 planting indices based on different sampling methods植被指数 传统采样法
Traditional sampling method冠层采样法
Canopy sampling methodVegetation index Pearson Spearman Pearson Spearman SR 0.29* 0.19* 0.33* 0.20* SRre 0.18* −0.05 0.12 0.13 NDVI 0.24* 0.19* 0.30* 0.20* gNDVI 0.11 0.08 0.21* 0.12 NDVIre 0.16* −0.05 0.092 0.13 CIg 0.17* 0.08 0.25* 0.12 Clre 0.18* −0.05 0.12 0.13 GI 0.29* 0.26* 0.26* 0.16* OSAVI 0.24* 0.18* 0.30* 0.20* RDVI 0.14* 0.04 0.19* 0.10 MSAVI 0.22* 0.19* 0.29* 0.20* SAVI 0.24* 0.18* 0.30* 0.20* PSRI −0.21* −0.25* −0.17* −0.15* MSRre 0.17* −0.05 0.12 0.13 EVI2 0.26* 0.18* 0.31* 0.20* MTCI 0.20* −0.01 0.17* 0.16* GEMI −0.19* −0.06 −0.20* −0.10 MTVI2 0.21* 0.04 0.19* 0.14 NDWI −0.11 −0.08 −0.21* −0.12 RTVICore 0.17* 0.07 0.12 0.12 注:*表示在P≤ 0.05时,该植被指数具有显著性。
Note: * indicates that the vegetation index is significant when P≤0.05.2.2 非时序网络反演结果
首先,分别使用FFNN网络对9种植被指数输入和15种植被指数输入进行非时序模型构建,筛选出在测试集样本中RMSE最小的迭代结果作为最终训练结果(见表3)。在训练集上,使用15种植被指数时,模型的R2为0.384,RMSE为3.15%;而使用9种植被指数时,R2为0.432,RMSE为3.56%。该结果表明9种植被指数数量作为输入的反演效果更加稳定,故该研究选择9种植被指数作为时序与非时序模型的最终输入。
表 3 不同数量植被指数非时序网络根域土壤水分反演精度Table 3. Precision of root soil moisture retrieval using non- temporal networks with different vegetation index numbers植被指数数量
Vegetation index numbersR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test set9 0.432 0.269 2.99 3.56 15 0.384 0.254 3.15 3.61 2.3 时序网络反演结果
为验证时间序列信息对根域土壤含水率反演的影响,该研究构建了以多日植被指数为输入的数据集,并结合LSTM与BiLSTM进行建模。历史关注步长设置为3~7天,分别进行迭代训练,结果如表4所示。
表 4 不同历史关注步长下时序网络最佳迭代次数Table 4. Optimal iteration times for time-series networks under different history focuses on step size历史关注步长History focuses on step size LSTM BiLSTM 3 470 590 4 505 610 5 595 590 6 300 570 7 320 410 在训练集和测试集(表5)上,对比不同历史关注步长对应的测试结果发现,LSTM 网络和BiLSTM 网络均在历史关注步长为 5 天时效果最佳。
表 5 不同历史关注步长下时序网络的反演精度Table 5. Inversion accuracy of temporal networks with different history focuses on step size网络
Network历史关注步长
History focuses on step sizeR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test setLSTM 3 0.473 0.413 2.88 2.97 4 0.661 0.402 2.35 2.80 5 0.842 0.548 1.59 2.68 6 0.609 0.538 2.46 2.80 7 0.722 0.515 2.12 2.81 BiLSTM 3 0.552 0.537 2.66 2.67 4 0.676 0.488 2.33 2.65 5 0.815 0.624 1.74 2.45 6 0.797 0.587 1.82 2.70 7 0.775 0.485 1.91 2.93 在最优历史关注步长下,对LSTM和BiLSTM网络的模型表现(图8)。图8a显示,LSTM网络在迭代次数小于300时,训练集和测试集的损失函数值显著下降,模型逐渐收敛;相比之下,图8b表明,BiLSTM网络的损失函数值下降趋势更明显,在360至530次迭代间达到最佳效果,且测试集损失函数值更低,模型收敛效果更佳。
结合散点图(图9)分析,在训练集上,LSTM的拟合效果略优于BiLSTM,R2为0.842,RMSE为1.59%;BiLSTM 的R2为 0.815,RMSE为1.74%。然而,在测试集上,BiLSTM 网络表现出更好的泛化能力,R2为0.624,RMSE为 2.45%,优于LSTM网络(R2为 0.548,RMSE为 2.68%)。
2.4 基于不同模型反演效果分析
为了体现不同时间尺度下的时序与非时序反演效果,该研究构建传统前馈神经网络(FFNN)、长短期记忆网络(LSTM)和双向长短期记忆网络(BiLSTM)在最优历史关注步长的反演结果对比,结果如表6所示。
表 6 不同方法对于根域土壤含水率的反演结果Table 6. Inversion results of soil moisture content in root domain by different methods网络
Network历史关注步长
History focuses
on step sizeR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test setFFNN 5 0.650 0.528 2.40 2.78 LSTM 5 0.842 0.548 1.59 2.68 BiLSTM 5 0.815 0.624 1.74 2.45 结合上表,时序网络模型在测试集上的拟合精度均高于非时序网络模型,具体表现为:
1)FFNN模型:由于模型无法吸收时间序列信息,FFNN模型在测试集上的R2值仅为0.528,RMSE值为2.78%,明显逊色于LSTM和BiLSTM模型。这主要归因于FFNN模型无法有效利用长时间段内的历史信息,限制了其反演精度。
2)LSTM模型:在最优历史关注步长时,测试集上的R2值为0.548,RMSE值为2.68%。这表明LSTM模型能够有效捕捉土壤水分变化的时间序列特征,具有较高的预测准确性。
3)BiLSTM模型:在相同步长下,测试集上的R2值为0.624,RMSE值为2.45%。尽管BiLSTM模型在某些情况下R2值略低于LSTM模型,但其在测试集上的表现更为优异,显示出在处理复杂时间序列数据时具有更强的鲁棒性和泛化能力。
3. 结 论
该研究提出以猕猴桃根域土壤水分为研究对象,通过Pearson与Spearman相关性法筛选植被指数得到最优指标组合,结合传统前馈神经网络(FFNN)、长短期记忆网络(LSTM)和双向长短期记忆网络(BiLSTM)构建猕猴桃果园的根域土壤水分反演模型,并通过采集田间数据证明其可靠性,获得主要结论如下:
1)冠层范围采样法在植被指数与根域土壤水分的线性与非线性相关性上优于传统矩形采样法,能够选择更多显著性相关的植被指数,从而提高模型反演精度。
2)在不同的历史关注步长下,时序网络(LSTM和BiLSTM)相比传统FFNN模型表现出显著的优势。特别是BiLSTM模型,具有更强的泛化能力,在测试集上R2为0.624,RMSE为2.45%,优于LSTM。证明了时序网络能够有效捕捉土壤水分变化的时间序列特征。
-
表 1 植被指数计算式
Table 1 Vegetation index formula
植被指数Vegetation index(VI) 计算式Formula 参考文献 Reference 绿色指数Green index(GI) ρGRE/ρGREρREDρRED [17] 修改型土壤调整植被指数Modified soil adjusted vegetation index(MSAVI) 0.5×(2×ρNIR+1−√(2×ρNIR+1)2−8(ρNIR−ρRED)) [17] 绿度归一化植被指数Green normalized difference vegetation index(gNDVI) (ρNIR−ρGRE)/(ρNIR−ρGRE)(ρNIR+ρGRE)(ρNIR+ρGRE) [18] 归一化植被指数Normalized difference vegetation index(NDVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED)(ρNIR+ρRED) [18] 优化土壤调节植被指数Optimized soil adjusted vegetation index(OSAVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED+0.5)(ρNIR+ρRED+0.5) [18] 重归一化植被指数Renormalized difference vegetation index(RDVI) (ρNIR−ρGRE)/(ρNIR−ρGRE)(√ρNIR+ρGRE)(√ρNIR+ρGRE) [18] 土壤调整植被指数Soil adjusted vegetation index(SAVI) (ρNIR−ρRED)/(ρNIR−ρRED)(ρNIR+ρRED+L)(ρNIR+ρRED+L)×(1+L)* [19] 简单比值指数Simple ratio index(SR) ρNIR/ρNIRρREDρRED [19] 绿色叶绿素植被指数Green chlorophyll vegetation index(GCVI) ρNIR/ρGRE−1 [20] 植物衰老反射指数Plant senescence reflection index(PSRI) (ρRED−ρGRE)/ρREG [20] 红边归一化差异植被指数Red-edge normalized difference vegetation index(NDVIre) (ρNIR−ρREG)/(ρNIR+ρREG) [21] 修正型红边简单比值指数Modified red-edge simpleratio index(MSRre) (ρNIR/ρREG−1)/(√ρNIR/ρREG+1)(√ρNIR/ρREG+1) [21] 红边叶绿素指数Red-edge chlorophyll index(CIre) ρNIR/ρREG−1 [21] 红边简单比率指数Red-edge simple ratio index(SRre) ρNIR/ρREG [21] 双波段增强植被指数Two-band enhanced vegetation index(EVI2) 2.5×(ρNIR−ρRED)/2.5×(ρNIR−ρRED)(ρNIR+2.4ρRED+1)(ρNIR+2.4ρRED+1) [22] 陆地叶绿素指数MERIS terrestrial chlorophyll index(MTCI) (ρNIR−ρREG)/(ρNIR−ρREG)(ρREG+ρRED)(ρREG+ρRED) [22] 归一化差值水体指数Normalized difference water index(NDWI) (ρGRE−ρNIR)/(ρGRE−ρNIR)(ρGRE+ρNIR)(ρGRE+ρNIR) [23] 全球环境监测指数Global environment monitoring index(GEMI) eat×(1−0.25×eat)−(ρRED−0.125)/−(ρRED−0.125)(1−ρRED)(1−ρRED)eat=(2×(ρ2NIR−ρ2RED)+1.5×ρNIR+0.5ρRED)/+1.5×ρNIR+0.5ρRED)(ρNIR+ρRED+0.5)(ρNIR+ρRED+0.5) [23] 红边三角植被指数Red-edge triangular vegetation index(RTVICore) 100×(ρNIR−ρREG)−10×(ρNIR−ρGRE) [24] 修正型三角植被指数Modified triangular vegetation index(MTVI2) 1.5×(1.2×(ρNIR−ρGRE)−2.5×(ρRED−ρGRE))×√(2×ρNIR+1)2−(6×ρNIR−5√ρRED)−0.5 [24] 注:ρGRE为绿波段平均反射率;ρRED为红波段平均反射率;ρREG为红边波段平均反射率;ρNIR为近红外波段平均反射率。L为土壤亮度校正因子,根据覆盖量此处取L=0.90。 eat为中间变量,通过近红外和红光波段的值按照特定公式计算得出。
Note: ρGRE is the average reflectance of green band; ρRED is the average reflectance of the red band; ρREG is the average reflectance of the red-edged band; ρNIR is the average reflectance of near infrared band. L is the correction factor for soil brightness, L=0.90 is taken here according to the coverage amount. eat is the intermediate variable, which is calculated by the values of near-infrared and red wavelengths according to a specific formula.表 2 基于不同采样方法的20种植被指数Pearson与Spearman相关性分析表
Table 2 Pearson and Spearman correlation analysis table of 20 planting indices based on different sampling methods
植被指数 传统采样法
Traditional sampling method冠层采样法
Canopy sampling methodVegetation index Pearson Spearman Pearson Spearman SR 0.29* 0.19* 0.33* 0.20* SRre 0.18* −0.05 0.12 0.13 NDVI 0.24* 0.19* 0.30* 0.20* gNDVI 0.11 0.08 0.21* 0.12 NDVIre 0.16* −0.05 0.092 0.13 CIg 0.17* 0.08 0.25* 0.12 Clre 0.18* −0.05 0.12 0.13 GI 0.29* 0.26* 0.26* 0.16* OSAVI 0.24* 0.18* 0.30* 0.20* RDVI 0.14* 0.04 0.19* 0.10 MSAVI 0.22* 0.19* 0.29* 0.20* SAVI 0.24* 0.18* 0.30* 0.20* PSRI −0.21* −0.25* −0.17* −0.15* MSRre 0.17* −0.05 0.12 0.13 EVI2 0.26* 0.18* 0.31* 0.20* MTCI 0.20* −0.01 0.17* 0.16* GEMI −0.19* −0.06 −0.20* −0.10 MTVI2 0.21* 0.04 0.19* 0.14 NDWI −0.11 −0.08 −0.21* −0.12 RTVICore 0.17* 0.07 0.12 0.12 注:*表示在P≤ 0.05时,该植被指数具有显著性。
Note: * indicates that the vegetation index is significant when P≤0.05.表 3 不同数量植被指数非时序网络根域土壤水分反演精度
Table 3 Precision of root soil moisture retrieval using non- temporal networks with different vegetation index numbers
植被指数数量
Vegetation index numbersR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test set9 0.432 0.269 2.99 3.56 15 0.384 0.254 3.15 3.61 表 4 不同历史关注步长下时序网络最佳迭代次数
Table 4 Optimal iteration times for time-series networks under different history focuses on step size
历史关注步长History focuses on step size LSTM BiLSTM 3 470 590 4 505 610 5 595 590 6 300 570 7 320 410 表 5 不同历史关注步长下时序网络的反演精度
Table 5 Inversion accuracy of temporal networks with different history focuses on step size
网络
Network历史关注步长
History focuses on step sizeR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test setLSTM 3 0.473 0.413 2.88 2.97 4 0.661 0.402 2.35 2.80 5 0.842 0.548 1.59 2.68 6 0.609 0.538 2.46 2.80 7 0.722 0.515 2.12 2.81 BiLSTM 3 0.552 0.537 2.66 2.67 4 0.676 0.488 2.33 2.65 5 0.815 0.624 1.74 2.45 6 0.797 0.587 1.82 2.70 7 0.775 0.485 1.91 2.93 表 6 不同方法对于根域土壤含水率的反演结果
Table 6 Inversion results of soil moisture content in root domain by different methods
网络
Network历史关注步长
History focuses
on step sizeR2 RMSE/% 训练集
Training set测试集
Test set训练集
Training set测试集
Test setFFNN 5 0.650 0.528 2.40 2.78 LSTM 5 0.842 0.548 1.59 2.68 BiLSTM 5 0.815 0.624 1.74 2.45 -
[1] 曹永研,杨玮,王懂,等. 基于水分和粒度的土壤有机质特征波长提取与预测模型[J]. 农业机械学报,2022,53(1):241-248. CAO Yongyan, YANG Wei, WANG Dong, et al. Characteristic wavelength extraction and prediction model of soil organic matter based on water content and grain size [J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(S1): 241-248. (in Chinese with English abstract)
[2] 介飞龙,费良军,李山,等. 基于根域层水分状态的旱区净灌溉需水量模型构建和应用[J]. 农业工程学报,2022,38(13):105-113 JIE Feilong, FEI Liangjun, LI Shan, et al. Construction and application of net irrigation water requirement model for arid regions based on water state in root zone[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2022, 38(13): 105-113. (in Chinese with English abstract)
[3] LI Z , ZHAO L , ZHAO Z , et al. Estimating the transpiration of kiwifruit using an optimized canopy resistance model based on the synthesis of sunlit and shaded leaves[J]. Agricultural Water Management, 2024, 306: 109193.
[4] 何新林,王佳鑫,龚萍,等. 南疆矮化密植高产优质香梨节水灌溉模式筛选[J]. 农业工程学报,2024,40(2):249-262. HE Xilin, WANG Jiaxin, GONG Ping, et al. Screening of water-saving irrigation models for high-yield and high-quality fragrant pear with dwarf dense planting in southern Xinjiang[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(2): 249-262. (in Chinese with English abstract)
[5] EMNA A, MEHREZ Z, ZOHRA C L, et al. Sensitivity of surface soil moisture retrieval to satellite-derived vegetation descriptors over wheat fields in the Kairouan plain[J]. European Journal of Remote Sensing, 2023, 56(1):2260555.
[6] 马红章,孟庆涛,刘素美. TAU-OMEGA遥感辐射模型改进与参数化[J]. 农业工程学报,2024,40(11):292-298. MA Hongzhang, MENG Qingtao, LIU Sumei. Improvement and parameterization of TAU-OMEGA remote sensing radiation model [J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(11) : 292-298. (in Chinese with English abstract)
[7] 李文佳,王瑞燕,徐家豪,等. 西北干旱区农田灌溉事件遥感检测[J]. 农业工程学报,2024,40(17):120-128. LI Wenjia, WANG Ruiyan, XU Jiahao, et al. Remote sensing detection of farmland irrigation events in arid Northwest China [J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(17) : 120-128. (in Chinese with English abstract)
[8] 刘欣,王树果,尹培霖,等. 不同植被覆盖下宇宙射线中子法土壤水分反演[J]. 农业工程学报,2024,40(22):98-107. LIU Xin, WANG Shuguo, YIN Peilin, et al. Soil water inversion by cosmic ray neutron method under different vegetation cover [J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(22): 98-107. (in Chinese with English abstract)
[9] 靳亚红,吴鑫淼,甄文超,等. 基于采样点光谱信息窗口尺度优化的土壤含水率无人机多光谱遥感反演[J]. 农业机械学报,2024,55(1):316-327. JIN Yahong, WU Xinmiao, ZHEN Wenchao, et al. UAV multi-spectral remote sensing retrieval of soil water content based on window scale optimization of spectral information of sampling points [J]. Transactions of the Chinese Society for Agricultural Machinery, 2024, 55(1): 316-327. (in Chinese with English abstract)
[10] DENG J, PAN S, ZHOU M, et al. Optimum sampling window size and vegetation index selection for low-altitude multispectral estimation of root soil moisture content for Xuxiang Kiwifruit[J]. Agricultural Water Management, 2023, 282 : 108297.
[11] ZHU S, CUI N, LI G, et al. Enhancing precision of root-zone soil moisture content prediction in a kiwifruit orchard using UAV multi-spectral image features and ensemble learning[J]. Computers and Electronics in Agriculture, 2024, 221: 108943.
[12] 潘时佳,吴津乐,程梅,等. 基于改进CNN的猕猴桃根域土壤水分反演方法[J]. 农业工程学报,2024,40(11):85-91. PAN Shijia, WU Jinle, CHENG Mei, et al. Soil water retrieval method in kiwifruit root domain based on improved CNN[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2024, 40(11): 85-91. (in Chinese with English abstract)
[13] 张军,邓俊涛,倪国威,等. 基于植被指数的猕猴桃根域土壤水分反演影响因素研究[J]. 农业机械学报,2022,53(12):223-230. ZHANG Jun, DENG Juntao, NI Guowei, et al. Study on Influencing factors of soil water inversion in Kiwifruit root domain based on vegetation index [J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(12): 223-230. (in Chinese with English abstract)
[14] 张智韬,王海峰,韩文霆,等. 基于无人机多光谱遥感的土壤含水率反演研究[J]. 农业机械学报,2018,49(2):173-181. ZHANG Zhitao, WANG Haifeng, HAN Wenting, et al. Soil water content inversion based on UAV multi-spectral Remote sensing[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(2): 173-181. (in Chinese with English abstract)
[15] RUO-CHEN L, CHANG-WAI W, Shu-yan L, et al. Quality prediction of tractor rotary tillage based on BiConvLSTM with self-attention[J]. Computers and Electronics in Agriculture, 2023, 206: 107643.
[16] PANG S, NIU Z, DENG J, et al. Dense orchard landscape mapping based on image merging with skeleton prediction[J]. International Journal of Applied Earth Observation and Geoinformation, 2023, 125:103578.
[17] 余国雄,王卫星,谢家兴,等. 基于物联网的荔枝园信息获取与智能灌溉专家决策系统[J]. 农业工程学报,2016,32(20):144-152. doi: 10.11975/j.issn.1002-6819.2016.20.019 YU Guoxiong, WANG Weixing, XIE JiaXing, et al. Information acquisition and intelligent irrigation expert decision system of litchi orchard based on Internet of Things[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2016, 32(20): 144-152. (in Chinese with English abstract) doi: 10.11975/j.issn.1002-6819.2016.20.019
[18] MANUEL V, CAROLINE A, LUIS J R, et al. Remote detection of cyanobacterial blooms and chlorophyll-a analysis in a eutrophic reservoir using sentinel-2[J]. Sustainability, 2021, 13(15): 8570-8570. doi: 10.3390/su13158570
[19] 张瑜,张立元,Zhang Huihui,等. 玉米作物系数无人机遥感协同地面水分监测估算方法研究[J]. 农业工程学报,2019,35(1):83-89. ZHANG Yu, ZHANG Liyuan, ZHANG Huihui, et al. Research on estimation method of maize crop coefficient by UAV remote sensing in collaboration with ground water monitoring[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2019, 35(1): 83-89. (in Chinese with English abstract)
[20] THIAGO F, ANNE S, BILL H, et al. Differentiation of six grassland/forage types in three Canadian rcoregions based on spectral characteristics[J]. Remote Sensing, 2022, 14(9): 2121. doi: 10.3390/rs14092121
[21] KAMONPORN U, SARAWUT N, NOPHEA S, et al. Discrimination of Mangrove Stages Using Multitemporal Sentinel-1 C-Band Backscatter and Sentinel-2 Data—A Case Study in Samut Songkhram Province, Thailand[J]. Forests, 2022, 13(9): 1433-1433. doi: 10.3390/f13091433
[22] CECILIA C, YADVINDER M, JIN W, et al. Leaf aging of Amazonian canopy trees as revealed by spectral and physiochemical measurements.[J]. The New Phytologist, 2017, 214(3): 1049-1063. doi: 10.1111/nph.13853
[23] SANA A, HASAN J K, AHMED F P, et al. Exploring dynamic response of agrometeorological droughts towards winter wheat yield loss risk using machine learning approach at a regional scale in Pakistan[J]. Field Crops Research, 2023, 302:109057.
[24] ZHU S, CUI N, JIN H, et al. Optimization of multi-dimensional indices for kiwifruit orchard soil moisture content estimation using UAV and ground multi-sensors[J]. Agricultural Water Management, 2024, 294:108705.