迁移学习(transfer learning)是一种机器学习方法,其核心思想在于利用在一个或多个源任务(source tasks)上学到的知识来帮助目标任务(target task)的学习过程[17],尤其适用于目标任务数据稀缺或标注成本高昂的场景。迁移学习打破了传统机器学习中“每个任务从头开始学习”的局限,通过知识的迁移和再利用,显著提高了学习效率和学习效果。迁移指将一个在丰富训练数据(源域)上训练的模型迁移到另一个具有较少或无注释训练数据任务的过程[18]。当训练的标注样本不足时,可以用已训练好的模型将少量样本迁移至新的领域(目标域),在获得检测能力的情况下,较大程度地减少对类似分布样本的标注数据的依赖[19]。迁移学习方法在图像、语音和自然语言处理领域已被广泛应用[20]。Li等[21]分别采用降维和迁移学习方法,提出了两种具有较强泛化能力的校正迁移任务方法,分别是选择性剪枝与精细化策略(selective pruning and refinement strategy,SPRS)方法和基于标准正态变量变换(standard normal variate,SNV)的Aug-TrAdaBoost.R2(augmented transfer adaptive bossing for regression)方法。在基于近红外光(near infrared,NIR)的茶叶光谱数据集中实现了不同种类茶叶之间以及同种类茶叶不同批次之间的模型迁移。Peng等[22]针对小样本场景下的神经网络模型训练问题,通过对先验知识的应用来满足模型迁移的需求。
在收集原始光谱数据时,除了枯叶的固有特征外,还可能包含一些干扰信息,如散射、噪声和基线漂移,这些都可能影响含水率的反演精度。因此,对原始光谱数据进行预处理是必要的步骤[24]。本研究采用了标准正态变量变换(standard normal variate,SNV)和S-G(savitzky-golay)平滑两种预处理方法对光谱进行预处理,SNV消除光谱散射影响造成的误差,去除无关变量,降低光谱维度空间[25];S-G平滑算法是一种利用滑动窗口进行卷积运算的方法,可以有效降低光谱中的噪声和干扰信号,同时保留光谱的光滑特征[26]。处理后的图像如图3和图4所示。
在地表枯叶含水率研究领域,模型效果的评价指标主要有3个:分别是MSE、MAE和决定系数(coefficient of determination,R2)。MSE是一种常用的衡量模型预测值与实际观测值之间差异的指标,用于评估模型在给定数据上的拟合程度。MSE(式中记为MSE)是通过计算预测值与实际观测值之间差异的平方的平均值得到。
LIANGS.Surface dead fuel loading and combustion energy release of five typical subtropical forest types[D].Changsha:Central South University of Forestry & Technology,2024.
XUW H, WUC, YANGL,et al.Comparative study on surface litter load and fire intensity of Pinus armandii and Pinus yunnanensis in Northeastern Yunnan Province[J].Journal of Southwest Forestry University(Natural Sciences),2019,39(5):151-156.
XIEZ X, HUH Q, YANGX G,et al.Estimating model of typical forest litter moisture content based on the field spectrum in Daxing'anling of China[J].Chinese Journal of Ecology,2017,36(11):3321-3328.
SUNL, LIUQ, HUT X.Advances in research on prediction model of moisture content of surface dead fuel in forests[J].Scientia Silvae Sinicae,2021,57(4):142-152.
HUH Q, LUOS S, LUOB Z,et al.Forest fuel moisture content and its prediction model[J].World Forestry Research,2017,30(3):64-69.
[13]
王婕.安宁地区主要树种死可燃物含水率及火险等级研究[D].北京:北京林业大学,2018.
[14]
WANGJ.Research on moisture content of dead forest fuel and fire danger rating of maior tree secies in Annin area[D].Beijing:Beijing Forestry University,2018.
ZHANGY L, TIANL L, XIANGM,et al.Effects of indoor simulated air temperature and relative humidity on the equilibrium moisture content and time lag of the fuelbed of Quercus mongolica [J].Chinese Journal of Ecology,2022,41(10):2072-2080.
[17]
张小丹.基于卷积神经网络的近红外光谱定量分析模型建立与转移研究[D].太原:中北大学,2022.
[18]
ZHANGX D.Research on the establishment and transfer of near-infrared spectroscopy quantitative analysis model based on convolutional neural network[D].Taiyuan:North University of China,2022.
LIUY M, WANGQ H, CHENY Z,et al.Non-Destructive near-infrared spectroscopy of physical and chemical indicator of pork meat[J].Spectroscopy and Spectral Analysis,2024,44(5):1346-1353.
ZHUS H, WUZ W, LIZ J,et al.Moisture dynamics and modeling of ground surface fine dead combustibles in Pinus massoniana forest in Southern Jiangxi,China[J].Scientia Silvae Sinicae,2024,60(5):158-168.
LIUK, WANGY F, PENGZ Q,et al.Hyperspectral soil nutrient feature band extraction method based on machine learning[J].Acta Optica Sinica,2024:1-18.
[25]
王冉.基于卷积神经网络与迁移学习的高光谱图像分类[D].赣州:赣南师范大学,2023.
[26]
WANGR.Hyperspectral image classification based on convolutional neural network and transfer learning[D].Ganzhou:Gannan Normal University,2023.
[27]
PENGB, ZHANGJ W, XINGJ,et al.Measuring moisture content of dead fine fuels based on the fusion of spectrum meteorological data[J].Journal of Forestry Research,2023,34:1333-1346.
[28]
叶颖慧.森林死可燃物含水率在线测量技术研究[D].哈尔滨:东北林业大学,2019.
[29]
YEY H.Research on on-line measurement technology of forest dead fuel moisture content[D].Harbin:Northeast Forestry University,2019.
[30]
JIANGH, DENGJ, ZHUC.Quantitative analysis of aflatoxin B1 in moldy peanuts based on near-infrared spectra with two-dimensional convolutional neural network[J].Infrared Physics & Technology,2023,131:104672.
[31]
郑文瑞.基于迁移学习的土壤速效磷近红外预测方法研究[D].合肥:安徽农业大学,2022.
[32]
ZHENGW R.Research on near infrared prediction methods of soil available phosphorus based on transfer learning[D].Hefei:Anhui Agricultural University,2022.
[33]
魏玉震.基于光谱和光谱成像技术的茶叶含水率检测机理和方法研究[D].杭州:浙江大学,2019.
[34]
WEIY Z.Moisture content detection of tea leaves based on spectral and spectral imaging technologies[D].Hangzhou:Zhejiang University,2019.
[35]
沈欢超.基于人工智能与近红外光谱技术的烟草质量控制研究[D].杭州:浙江大学,2023.
[36]
SHENH C.Studies on quality control of tobacco based on artificial intelligence and near infrared spectroscopy[D].Hangzhou:Zhejiang University,2023.
[37]
胡燕芳.基于小样本学习的图像分类方法研究与应用[D].赣州:江西理工大学,2022.
[38]
HUY F.Research and application on image classification method based on few-shot learning[D].Ganzhou:Jiangxi University of Science And Technology,2022.
[39]
LIX L, LIZ, YANGX F,et al.Boosting the generalization ability of Vis-NIR-spectroscopy-based regression models through dimension reduction and transfer learning[J].Computers and Electronics in Agriculture,2021,186:106157.
[40]
PENGX, LIY, WEIX,et al.RGB-NIR image categorization with prior knowledge transfer[J].Journal on Image and Video Processing,2018,149:1.
[41]
JIANGH, XUEY, CHENQ.Quantitative analysis of residues of chlorpyrifos in corn oil based on Fourier transform near-infrared spectroscopy and deep transfer learning[J].Infrared Physics & Technology,2023,133(6):104814.
ZHUR F, XUY D, SUNW B,et al.Research on nutritional components of Pennisetum Rich.forage by near infrared spectroscopy model based on artificial neural network[J].Acta Agrestia Sinica,2024,32(2):527-534.
WANGZ Q, LID P, LIUQ,et al.Water content detection of Camellia oleifera seeds based on temperature correction and visible/near infrared spectroscopy[J].Food & Machinery,2022,38(12):127-132.
XUS Y, LIUZ Y, HUANGY.Hyperspectral non-destructive detection of nitrogen,phosphorus and potassium content of watermelon seedling leaves based on Self-Attention-BiLSTM network[J].Transactions of the Chinese Society for Agricultural Machinery,2024,55(98):243-252.
[48]
TANA, WANGY, ZHAOY,et al.Near infrared spectroscopy quantification based on Bi-LSTM and transfer learning for new scenarios[J].Spectrochimica Acta Part A:Molecular and Biomolecular Spectroscopy,2022,283:121759.