Optimization of comprehensive wheat growth index system and monitoring model based on LAI Authors: Mingguang Diao, Shuning Liang, Jianing Chen, Chuyan Zhang, Yong Liu Published: 2025 ABSTRACT Wheat growth monitoring plays a vital role in agricultural decision-making and food security. This study aims to develop an accurate and efficient monitoring method for wheat growth by integrating satellite remote sensing and machine learning techniques. Based on preprocessed Sentinel-2 satellite images and measured wheat leaf area index (LAI) data, a set of 11 vegetation indices—such as NDVI, NDRE, and RVI—were selected and ranked through Pearson correlation analysis. A comprehensive index system was then constructed by selecting the top eight indices using a stepwise optimization approach. Three machine learning models—Linear Regression (LR), Backpropagation Neural Network (BPNN), and XGBoost—were applied to evaluate the performance of the index system, with the Particle Swarm Optimization (PSO) algorithm employed to optimize each model. The results demonstrate that the PSO-optimized XGBoost model achieved the highest accuracy (R² = 0.94, MSE = 0.075), exhibiting strong stability and robustness to data fluctuations. These findings suggest that the proposed approach provides a reliable solution for wheat growth monitoring. Introduction Wheat is an important grain crop widely grown in China, and its growth monitoring is crucial for efficient, high-yield and high-quality agricultural production. However, wheat has a long growth cycle, is affected by a variety of factors, and its morphology varies significantly at different fertility stages, so efficient, accurate, and non-invasive monitoring methods are needed to guide crop management strategies. Currently, methods for monitoring wheat growth mainly involve manual observation and traditional technological monitoring. Manual observation is time-consuming and less accurate. Traditional technological monitoring mainly evaluates wheat growth through indicators like rainfall and soil conditions, achieving higher accuracy. However, it incurs high costs, lacks real-time crop information, and faces challenges in monitoring remote areas with difficult sampling. This study aims to achieve rapid, economical and high-precision monitoring of wheat growth conditions by constructing a comprehensive wheat growth indicator system that integrates remote sensing and artificial intelligence, so as to reduce labor costs, improve monitoring efficiency and enhance prediction accuracy. In response to the above situation, scholars both domestically and internationally have conducted extensive research and proposed numerous agronomic parameter indicators for characterizing wheat growth. Such indicators include Leaf Area Index (LAI), Chlorophyll Relative Content (SPAD), leaf nitrogen content, Absorbed Photosynthetically Active Radiation (APAR), Above-Ground Biomass (AGB), etc1. Existing studies have used chlorophyll content to predict the nitrogen nutrient status of crops, and 27 spectral parameters were extracted from hyperspectral images of crops to invert the SPAD of cotton leaves and obtained more accurate inversion results2. Some other scholars used spectral vegetation indices to construct a chlorophyll inversion model for wheat flowering, and found that the highest correlation was between the Normalized Vegetation Index (NDVI) and the chlorophyll content. Some researchers have employed Support Vector Machines for predicting physiological indicators such as SPAD and biomass in corn leaves, achieving high prediction accuracy3. Another study found that the visible to short-infrared wavelength band from 350 mm to 2500 mm of winter wheat canopy had a strong positive correlation with nitrogen content of leaves4. And a study analyzed and concluded that H-value of the image of cotton canopy was a reliable indicator for describing the information of cotton growth5. Some scholars have proved that the proposed red-trimmed triangular vegetation index (RTVI) is the best index for estimating crop canopy biomass6. And some scholars have found that the Red-Edge Chlorophyll Index (CIre) effectively enhances inversion precision7. In the study of crop growth monitoring, domestic and foreign researchers have carried out more theoretical and methodological studies by combining LAI and remote sensing techniques. Existing studies have established the prediction model of vegetation canopy spectral characteristics, spectral reflectance and leaf chlorophyll value and LAI, and better predicted the growth of overstory vegetation on torrefied alkaline soil8. Another study found that the LAI of oilseed rape crop could be accurately estimated by using the Red Edge Vegetation Index (REDVI)9. Some scholars have established regression models relating canopy coverage with LAI and Dry Matter Accumulation (DMA) using population images during the corn elongation and tassel emergence stages10. Certain scholars have demonstrated that the proposed Modified Triangular Vegetation Index (MTVI2) and Modified Chlorophyll Absorption Ratio Index (MCARI2) are the optimal factors for predicting crop LAI11. Another study incorporated Leaf Area Index (LAI) obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) into the World Food Studies Model (WOFOST) to predict the maturity date of winter wheat in Henan Province, revealing the highest LAI prediction accuracy during the flowering to grain-filling stages12. Moreover, other scholars have highlighted a significant correlation between Leaf Area Index (LAI) of rice crops and the Normalized Difference Vegetation Index (NDVI) and Ratio Vegetation Index (RVI)13. As artificial intelligence gains prominence, scholars worldwide have harnessed machine learning techniques in the agricultural domain, augmenting the precision of wheat growth prediction. Some scholars have reviewed the progress and knowledge gaps in the application of machine learning (ML) in life cycle assessment (LCA) and proposed directions for enhancing future LCA research through ML14. Certain studies have utilized Support Vector Machine (SVM) models to extract and monitor wheat growth using multi-source remote sensing images in the Xiong’an research area15. Additionally, another study has inverted summer maize Leaf Area Index (LAI) using eight vegetation indices highly correlated with LAI and plant height, employing the Gradient Boosting Tree algorithm16. Some scholars have used winter wheat biomass, plant height, SPAD, and other data to construct a comprehensive index system based on the coefficient of variation17. Other scholars have constructed prediction models for winter wheat Leaf Area Index (LAI) during three growth stages using seven spectral parameters and multiple regression, PLSR, RF, BPNN, and SVM algorithms, achieving high fitting accuracy18. Another study explained and predicted crop yields using a Bayesian ensemble model (BM), which offered better interpretability and predictive power compared to black-box machine learning methods19. However, the existing research has the following shortcomings: most studies only predict a single growth indicator or a single model, lacking systematic comparison and fusion optimization of model performance; insufficient synergy between traditional technical monitoring and artificial intelligence methods, making it difficult to balance real-time performance and accuracy; and lacking efficient monitoring strategies for large-scale fields under the constraints of cost and computing resources. To address these issues, this study chooses to combine remote sensing imagery and artificial intelligence technology for wheat growth monitoring. Regional sensing technology can accurately, efficiently, and over a large area to provide more comprehensive and detailed wheat growth assessment, which can help the relevant personnel to grasp the wheat planting status in real time20. Meanwhile, the development of artificial intelligence technology has contributed to the automation and high precision of wheat growth monitoring. Artificial intelligence models can interpret data from remote sensing images and automatically recognize patterns to predict crop growth. This study proposes a comprehensive approach that integrates multispectral UAV remote sensing data and a comparative ensemble of machine learning models (including RF, SVM, XGBoost, and BP neural networks) to evaluate and predict wheat growth indicators across multiple key stages. The novelty of this study lies in its model comparison strategy, multi-index prediction scheme, and its application of an optimized model ensemble for enhanced prediction accuracy and interpretability. Research data and methods Data sources This study used Sentinel-2 remote sensing image data from the Copernicus Data Space Ecosystem (https://dataspace.copernicus.eu/). Among them, the wheat Leaf Area Index (LAI) dataset originates from the canopy chlorophyll and ground validation dataset of winter wheat in Yucheng, Shandong21. The dataset contains a total of 107 samples, and the content of the dataset is shown in Table 1. Ground sampling was carried out during the wheat filling period from May 9 to 16, 2020. A total of 107 plots were set up, each with an area of 10 m × 10 m, which strictly corresponds to the 10 m spatial resolution of Sentinel-2. The center of the plot was positioned by high-precision GPS (error < 1 m) to ensure alignment with the satellite pixel boundary. LAI measurements were performed in an effective area of 6 m × 6 m in the center of the plot to avoid interference from edge mixed pixels and fully cover a single satellite pixel. Table 1Sample information.NumberSampling datePlot longitudePlot latitudePlot codeMeasured LAI12020/5/9116.55043736.775319W116.09922020/5/9116.55214436.774825X03_N_55.00032020/5/9116.55772736.774031W09-W104.55442020/5/9116.56513336.770740W085.07652020/5/9116.56912836.772604W074.33462020/5/9116.56023436.781318W065.62972020/5/9116.54423836.786077X01_S_55.00082020/5/9116.54250636.787166W04-W055.29492020/5/9116.54188636.787433W032.836102020/5/9116.54761136.785149W024.300112020/5/9116.54779736.786330W016.249122020/5/9116.54438336.787043X02_N_33.000132020/5/10116.54625836.769104W31-W325.167142020/5/10116.52142136.765311W29-W305.000152020/5/10116.52640336.764452W27-W285.000162020/5/10116.53093136.763486W23-W264.255172020/5/10116.53469036.760951W21-W225.212182020/5/10116.53326736.761199W19-W205.156192020/5/10116.53293736.759169W15-W184.747202020/5/10116.53262936.756912W13-W144.661212020/5/10116.53365736.749919W124.057222020/5/11116.58177036.824531W39-W406.066232020/5/11116.58099736.824671W37-W384.678242020/5/11116.57968736.824605W35-W364.912252020/5/11116.57841736.824698W33-W345.044262020/5/12116.57103636.825535W71-W725.172272020/5/12116.56945736.825629W69-W705.722282020/5/12116.56859936.825756W67-W684.875292020/5/12116.56790136.825670W65-W664.744302020/5/12116.56735036.825766W63-W645.815312020/5/12116.56664136.826104W61-W624.273322020/5/12116.58623636.829022W59-W604.909332020/5/12116.58695736.828984W57-W585.233342020/5/12116.58833536.828844W55-W563.844352020/5/12116.58904836.828959W53-W543.594362020/5/12116.59052536.828702W51-W525.839372020/5/12116.59166636.828580W49-W505.049382020/5/12116.59279836.828586W47-W485.174392020/5/12116.59366536.828660W45-W465.284402020/5/12116.59503036.828403W43-W444.549412020/5/12116.59596636.828336W41-W424.261422020/5/13116.58468036.825629W101-W1025.143432020/5/13116.57750436.832762W97-W985.078442020/5/13116.57800936.832748W95-W965.283452020/5/13116.57819036.832800W93-W945.534462020/5/13116.57861636.832690W91-W924.344472020/5/13116.57996636.832467W89-W904.918482020/5/13116.58028236.832581W87-W886.231492020/5/13116.58072036.832292W85-W864.040502020/5/13116.58075936.832289W83-W845.137512020/5/13116.58108436.832243W81-W825.047522020/5/13116.58133536.832263W79-W805.612532020/5/13116.58152436.832337W75-W785.167542020/5/13116.58231536.833043W73-W744.278552020/5/13116.57291336.833279W99-W1004.204562020/5/14116.57952536.828308Z-302.574572020/5/14116.57936036.827886Z-293.384582020/5/14116.57885736.828190Z-283.031592020/5/14116.57862036.827443Z-272.423602020/5/14116.57856436.827946Z-262.885612020/5/14116.57839136.827579Z-252.840622020/5/14116.57818336.828622Z-243.278632020/5/14116.57805436.828010Z-233.171642020/5/14116.57769936.828057Z-223.515652020/5/14116.57774736.827784Z-213.702662020/5/14116.58490036.825159Z-204.818672020/5/14116.58479636.825025Z-195.527682020/5/14116.58527936.825219Z-184.890692020/5/14116.58526736.824839Z-175.204702020/5/14116.58593336.825338Z-165.498712020/5/14116.58581036.824720Z-156.155722020/5/14116.58620136.825214Z-145.386732020/5/14116.58601136.824727Z-135.961742020/5/14116.58635736.825219Z-124.301752020/5/14116.58617936.824858Z-114.404762020/5/14116.58658536.824956Z-106.179772020/5/14116.58653136.824513Z-096.518782020/5/14116.58712036.824589Z-076.677792020/5/14116.58750236.824771Z-064.987802020/5/14116.58737436.824440Z-056.032812020/5/14116.58781236.824937Z-044.583822020/5/14116.58771636.824452Z-035.287832020/5/14116.58831236.824649Z-025.365842020/5/14116.58818636.824284Z-015.725852020/5/14116.58723236.825039Z-085.874862020/5/15116.57274236.828584Z-452.655872020/5/15116.57273136.828398Z-442.529882020/5/15116.57266036.828897Z-431.954892020/5/15116.57244236.829116Z-423.697902020/5/15116.57233936.828698Z-413.470912020/5/15116.57219836.828920Z-402.985922020/5/15116.57209236.828530Z-392.976932020/5/15116.58032436.828349Z-382.822942020/5/15116.58020036.828028Z-372.688952020/5/15116.57999736.828122Z-362.235962020/5/15116.57988836.827767Z-352.928972020/5/15116.57981536.828125Z-343.831982020/5/15116.57961436.827621Z-333.688992020/5/15116.57958136.828335Z-322.1781002020/5/15116.57942936.827853Z-312.4671012020/5/16116.57361236.829033Z-502.8141022020/5/16116.57343236.828506Z-492.6091032020/5/16116.57323236.829033Z-482.1011042020/5/16116.57311536.828710Z-472.2441052020/5/16116.57304136.828417Z-461.7981062020/5/16116.57396036.828807Z-522.6911072020/5/16116.57377836.828175Z-513.379 Table 1Sample information.NumberSampling datePlot longitudePlot latitudePlot codeMeasured LAI12020/5/9116.55043736.775319W116.09922020/5/9116.55214436.774825X03_N_55.00032020/5/9116.55772736.774031W09-W104.55442020/5/9116.56513336.770740W085.07652020/5/9116.56912836.772604W074.33462020/5/9116.56023436.781318W065.62972020/5/9116.54423836.786077X01_S_55.00082020/5/9116.54250636.787166W04-W055.29492020/5/9116.54188636.787433W032.836102020/5/9116.54761136.785149W024.300112020/5/9116.54779736.786330W016.249122020/5/9116.54438336.787043X02_N_33.000132020/5/10116.54625836.769104W31-W325.167142020/5/10116.52142136.765311W29-W305.000152020/5/10116.52640336.764452W27-W285.000162020/5/10116.53093136.763486W23-W264.255172020/5/10116.53469036.760951W21-W225.212182020/5/10116.53326736.761199W19-W205.156192020/5/10116.53293736.759169W15-W184.747202020/5/10116.53262936.756912W13-W144.661212020/5/10116.53365736.749919W124.057222020/5/11116.58177036.824531W39-W406.066232020/5/11116.58099736.824671W37-W384.678242020/5/11116.57968736.824605W35-W364.912252020/5/11116.57841736.824698W33-W345.044262020/5/12116.57103636.825535W71-W725.172272020/5/12116.56945736.825629W69-W705.722282020/5/12116.56859936.825756W67-W684.875292020/5/12116.56790136.825670W65-W664.744302020/5/12116.56735036.825766W63-W645.815312020/5/12116.56664136.826104W61-W624.273322020/5/12116.58623636.829022W59-W604.909332020/5/12116.58695736.828984W57-W585.233342020/5/12116.58833536.828844W55-W563.844352020/5/12116.58904836.828959W53-W543.594362020/5/12116.59052536.828702W51-W525.839372020/5/12116.59166636.828580W49-W505.049382020/5/12116.59279836.828586W47-W485.174392020/5/12116.59366536.828660W45-W465.284402020/5/12116.59503036.828403W43-W444.549412020/5/12116.59596636.828336W41-W424.261422020/5/13116.58468036.825629W101-W1025.143432020/5/13116.57750436.832762W97-W985.078442020/5/13116.57800936.832748W95-W965.283452020/5/13116.57819036.832800W93-W945.534462020/5/13116.57861636.832690W91-W924.344472020/5/13116.57996636.832467W89-W904.918482020/5/13116.58028236.832581W87-W886.231492020/5/13116.58072036.832292W85-W864.040502020/5/13116.58075936.832289W83-W845.137512020/5/13116.58108436.832243W81-W825.047522020/5/13116.58133536.832263W79-W805.612532020/5/13116.58152436.832337W75-W785.167542020/5/13116.58231536.833043W73-W744.278552020/5/13116.57291336.833279W99-W1004.204562020/5/14116.57952536.828308Z-302.574572020/5/14116.57936036.827886Z-293.384582020/5/14116.57885736.828190Z-283.031592020/5/14116.57862036.827443Z-272.423602020/5/14116.57856436.827946Z-262.885612020/5/14116.57839136.827579Z-252.840622020/5/14116.57818336.828622Z-243.278632020/5/14116.57805436.828010Z-233.171642020/5/14116.57769936.828057Z-223.515652020/5/14116.57774736.827784Z-213.702662020/5/14116.58490036.825159Z-204.818672020/5/14116.58479636.825025Z-195.527682020/5/14116.58527936.825219Z-184.890692020/5/14116.58526736.824839Z-175.204702020/5/14116.58593336.825338Z-165.498712020/5/14116.58581036.824720Z-156.155722020/5/14116.58620136.825214Z-145.386732020/5/14116.58601136.824727Z-135.961742020/5/14116.58635736.825219Z-124.301752020/5/14116.58617936.824858Z-114.404762020/5/14116.58658536.824956Z-106.179772020/5/14116.58653136.824513Z-096.518782020/5/14116.58712036.824589Z-076.677792020/5/14116.58750236.824771Z-064.987802020/5/14116.58737436.824440Z-056.032812020/5/14116.58781236.824937Z-044.583822020/5/14116.58771636.824452Z-035.287832020/5/14116.58831236.824649Z-025.365842020/5/14116.58818636.824284Z-015.725852020/5/14116.58723236.825039Z-085.874862020/5/15116.57274236.828584Z-452.655872020/5/15116.57273136.828398Z-442.529882020/5/15116.57266036.828897Z-431.954892020/5/15116.57244236.829116Z-423.697902020/5/15116.57233936.828698Z-413.470912020/5/15116.57219836.828920Z-402.985922020/5/15116.57209236.828530Z-392.976932020/5/15116.58032436.828349Z-382.822942020/5/15116.58020036.828028Z-372.688952020/5/15116.57999736.828122Z-362.235962020/5/15116.57988836.827767Z-352.928972020/5/15116.57981536.828125Z-343.831982020/5/15116.57961436.827621Z-333.688992020/5/15116.57958136.828335Z-322.1781002020/5/15116.57942936.827853Z-312.4671012020/5/16116.57361236.829033Z-502.8141022020/5/16116.57343236.828506Z-492.6091032020/5/16116.57323236.829033Z-482.1011042020/5/16116.57311536.828710Z-472.2441052020/5/16116.57304136.828417Z-461.7981062020/5/16116.57396036.828807Z-522.6911072020/5/16116.57377836.828175Z-513.379 Sample information. Field measurements of LAI were performed using an LAI-2200 plant canopy analyzer following standardized protocols. Measurements were taken during optimal lighting conditions - at sunrise/sunset on clear days or throughout the day during overcast conditions. Each plot was sampled along three transects with at least three measurements per transect, with the average value representing the plot-level LAI. Prior to measurement, senescent leaves and dew were removed to minimize interference, and a 180° view cap was used to prevent sun and shadow effects on the readings. The selection of 107 sample plots was designed to ensure statistical robustness while maintaining correspondence with individual Sentinel-2 pixels. The temporal sampling strategy was implemented to coincide with the critical grain-filling stage when wheat canopy characteristics show significant variation. Ground sampling occurred from May 9–16, 2020, the ideal verification method should be that the satellite transit time and the ground sampling time are on the same day. However, due to various restrictions, such consistency is rare. This study uses Sentinel-2 cloud-free images from April 29 and May 19 to cover this period. In order to coordinate the mismatched dates, the satellite data are interpolated to the ground observation date using the time series inverse distance weighting method, assuming that the CCC between the transit points changes linearly. According to the principle of inverse distance weighting of time series21, the satellite remote sensing data of these 2 days were time normalized with the measured wheat LAI data to obtain as shown in formula.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{B}_{Field}^{i}}^{\text{*}}=\left(1-{w}_{i}\right)\times\:{B}_{1}^{\text{*}}+{w}_{i}\times\:{B}_{2}^{\text{*}}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i}=\frac{{T}_{i}}{{T}_{2}-{T}_{1}}$$\end{document} Where,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{B}_{Field}^{i}}^{\text{*}}$$\end{document} is the multispectral data of day i in the sampling time of the measured wheat LAI dataset; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i}$$\end{document} is the weight index of day i, which is derived from Eq. (2); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{1}^{\text{*}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{2}^{\text{*}}$$\end{document} correspond to the multispectral data of the remote sensing images on April 29, 2020, and May 19, 2020, respectively; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{2}-{T}_{1}$$\end{document} represents the time interval between satellite’s two passes, which is 20 days. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i}$$\end{document} represents the time interval between ground sampling time and the previous satellite data. The Level-1 C remote sensing images, already orthorectified and geometrically corrected, required further radiometric and atmospheric corrections. Radiometric calibration converts image grayscale values to absolute radiance, enhancing signal accuracy. The specific calculation method is illustrated in Eq. (3).3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Le=Gain\times\:DN+Bias$$\end{document} In the formula, Le represents the radiance of the remote sensing image; Gain stands for the calibration coefficient gain; DN corresponds to the pixel brightness value of the remote sensing image, which is the grayscale value of the land feature; Bias is the offset value. This study used the Sen2cor plugin from the European Space Agency (ESA) to preprocess Sentinel-2 Level-1 C images, generating Level-2 A data. The Level-2 A data were resampled using Sentinel application platform (SNAP 10.0.0, https://step.esa.int/main/download/snap-download/previous-versions/) and saved in ENVI format, completing the preprocessing. Methodology In this study, on the basis of previous research, the leaf area index (LAI) was used to characterize the growth condition of wheat, and the optimal wheat growth monitoring model was constructed by using the measured wheat LAI dataset and hyperspectral feature parameters extracted from Sentinel-2 satellite remote sensing images, combined with machine learning techniques. The flowchart of the technical roadmap for this study is illustrated in Fig. 1which was generated using draw.io (https://www.drawio.com/), and can be divided into three main stages: (1) Data preprocessing stage: Sentinel-2 Level-1 C satellite imagery underwent radiometric calibration and atmospheric correction using Sen2Cor, followed by resampling and format conversion via SNAP 10.0.0 (https://step.esa.int/main/download/snap-download/previous-versions/). In parallel, the ENVI platform was used to extract vegetation indices (VIs) based on the location of LAI sampling points provided by the measured wheat LAI dataset. (2) Feature engineering stage: The processed dataset was further refined through post-processing and correlation analysis. Pearson correlation coefficients were calculated using a feature-by-feature approach to filter and select vegetation indices most relevant to LAI. The selected VIs were then integrated with the measured LAI values to construct a composite dataset for modeling. (3) Model construction and optimization stage: Using the integrated dataset, multiple machine learning models (including Linear Regression, Back Propagation Neural Network (BPNN), and XGBoost) were trained and evaluated using R² and MSE metrics via 10-fold cross-validation. The Particle Swarm Optimization (PSO) algorithm was further applied to optimize model parameters, and the best-performing model was selected as the final wheat growth monitoring model. Fig. 1Technical route flow. The figure shows the workflow for developing the wheat growth monitoring model using leaf area index (LAI) and Sentinel-2 hyperspectral features. Field-measured LAI data were combined with spectral indices and reflectance parameters extracted from Sentinel-2 images. After preprocessing and feature selection, machine learning algorithms, including Random Forest and Support Vector Machine, were used to build and evaluate the model. The optimal model was applied to predict the spatial distribution of LAI, providing insights into wheat growth conditions. Figure 1 was generated using draw.io (https://www.drawio.com/). Technical route flow. The figure shows the workflow for developing the wheat growth monitoring model using leaf area index (LAI) and Sentinel-2 hyperspectral features. Field-measured LAI data were combined with spectral indices and reflectance parameters extracted from Sentinel-2 images. After preprocessing and feature selection, machine learning algorithms, including Random Forest and Support Vector Machine, were used to build and evaluate the model. The optimal model was applied to predict the spatial distribution of LAI, providing insights into wheat growth conditions. Figure 1 was generated using draw.io (https://www.drawio.com/). Methodology for constructing comprehensive index system In this study, preprocessed Sentinel-2 satellite remote sensing images and measured wheat LAI dataset were employed. Firstly, wheat canopy spectral features were extracted, and 11 vegetation indices were screened. Pearson correlation analysis was conducted between the vegetation indices and wheat, revealing that most indices were correlated with LAI. Based on the correlation coefficients, a stepwise inclusion approach was employed to select vegetation indices. Ultimately, the top 8 vegetation indices were chosen as the comprehensive indicator system. The degree of reflection and radiative absorption of different bands of incident light by vegetation can reflect the type characteristics and growth status of vegetation. A large number of scholars5,6,23 have demonstrated that the Vegetation Index (VI) constructed from the reflectance of multiple visible light bands has a good sensitivity to the inversion of vegetation canopy information, and is an important parameter for observing the vegetation growth. Through a review of relevant literature, this study ultimately selected 11 candidate VI indexes that exhibited good performance in reflecting wheat growth conditions, as shown in Table 2. Table 2 provides the calculation formulas, functional descriptions and references of these indices to illustrate their applicability in wheat growth monitoring. Table 2Candidate vegetation index calculation methods and references.Vegetation IndexFormulaDescriptionNormalized Difference Vegetation Index(NDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:NDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{{\rho\:}_{Nir}+{\rho\:}_{R}}$$\end{document} Reflecting crop growth and nutritional information by comparing plant reflections in the near-infrared and red light bands[12]Normalized Difference Red Edge Index (NDRE) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:NDRE=\frac{{\rho\:}_{Nir}-{\rho\:}_{R-edge}}{{\rho\:}_{Nir}+{\rho\:}_{R-edge}}$$\end{document} Evaluate the health condition of vegetation by comparing the ratio of near-infrared reflection to red-edge reflection[9]Ratio Vegetation Index (RVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RVI=\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}$$\end{document} Estimating and monitoring the biomass of green plants through the ratio of near-infrared and red band reflectance[15]Differential Vegetation Index (DVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:DVI={\rho\:}_{Nir}-{\rho\:}_{R}$$\end{document} Monitoring Vegetation Ecological Environment through Calculation of Near-Infrared and Red Band Reflectance[18]Soil Adjusted Vegetation Index (SAVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:SAVI=\frac{(1+{l}_{1})\times\:({\rho\:}_{Nir}-{\rho\:}_{R})}{{\rho\:}_{Nir}+{\rho\:}_{R}+{l}_{1}}$$\end{document} Eliminating the influence of soil background on NDVI estimation by adjusting the soil regulation factor[13]Enhanced Vegetation Index (EVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:EVI=\frac{2.5\times\:({\rho\:}_{Nir}-{\rho\:}_{R})}{{\rho\:}_{Nir}+{6\rho\:}_{R}-{7.5\rho\:}_{B}+1}$$\end{document} Considering the influence of atmospheric and soil backgrounds, used for assessing vegetation coverage and growth status interference[14]Modified Simple Ratio (MSR) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:MSR=\frac{\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}-1}{\sqrt{\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}+1}}$$\end{document} Retrieving biophysical parameters using remote sensing data.[2]Green Normalized Vegetation Index (GNDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:GNDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{G}}{{\rho\:}_{Nir}+{\rho\:}_{G}}$$\end{document} Calculating vegetation coverage through the contrast between the green and near-infrared bands.[8]Optimized Soil Adjustment Vegetation Index (OSAVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:OSVAI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{{\rho\:}_{Nir}+{\rho\:}_{R}+{l}_{2}}$$\end{document} Taking soil factors into consideration, excluding soil effects, and reflecting the chlorophyll content of vegetation.[20]Triangular Vegetation Index (TVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:TVI=0.5\times\:(200\times\:\left({\rho\:}_{Nir}+{\rho\:}_{G}\right)+120\times\:\left({\rho\:}_{Nir}-{\rho\:}_{G}\right))$$\end{document} Describing the radiative energy absorbed by thepigments as a function of the relative difference between redand NIR reflectance in conjunction with the magnitude ofreflectance in the green region[1]Renormalized Difference Vegetation Index (RDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{\sqrt{{\rho\:}_{Nir}+{\rho\:}_{R}}}$$\end{document} Combining the advantages ofDVI and NDVI for low and high vegetation coverages,respectively[1]Note: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{Nir}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R-edge}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{G}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{B}$$\end{document} are near-infrared wavelength reflectance, red-edge wavelength reflectance, red-light wavelength reflectance, green-light wavelength reflectance, and blue-light wavelength reflectance, respectively; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} are the soil conditioning coefficients, and usually \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document} is taken as 0.5, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} is taken as 0.16. Table 2Candidate vegetation index calculation methods and references.Vegetation IndexFormulaDescriptionNormalized Difference Vegetation Index(NDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:NDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{{\rho\:}_{Nir}+{\rho\:}_{R}}$$\end{document} Reflecting crop growth and nutritional information by comparing plant reflections in the near-infrared and red light bands[12]Normalized Difference Red Edge Index (NDRE) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:NDRE=\frac{{\rho\:}_{Nir}-{\rho\:}_{R-edge}}{{\rho\:}_{Nir}+{\rho\:}_{R-edge}}$$\end{document} Evaluate the health condition of vegetation by comparing the ratio of near-infrared reflection to red-edge reflection[9]Ratio Vegetation Index (RVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RVI=\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}$$\end{document} Estimating and monitoring the biomass of green plants through the ratio of near-infrared and red band reflectance[15]Differential Vegetation Index (DVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:DVI={\rho\:}_{Nir}-{\rho\:}_{R}$$\end{document} Monitoring Vegetation Ecological Environment through Calculation of Near-Infrared and Red Band Reflectance[18]Soil Adjusted Vegetation Index (SAVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:SAVI=\frac{(1+{l}_{1})\times\:({\rho\:}_{Nir}-{\rho\:}_{R})}{{\rho\:}_{Nir}+{\rho\:}_{R}+{l}_{1}}$$\end{document} Eliminating the influence of soil background on NDVI estimation by adjusting the soil regulation factor[13]Enhanced Vegetation Index (EVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:EVI=\frac{2.5\times\:({\rho\:}_{Nir}-{\rho\:}_{R})}{{\rho\:}_{Nir}+{6\rho\:}_{R}-{7.5\rho\:}_{B}+1}$$\end{document} Considering the influence of atmospheric and soil backgrounds, used for assessing vegetation coverage and growth status interference[14]Modified Simple Ratio (MSR) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:MSR=\frac{\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}-1}{\sqrt{\frac{{\rho\:}_{Nir}}{{\rho\:}_{R}}+1}}$$\end{document} Retrieving biophysical parameters using remote sensing data.[2]Green Normalized Vegetation Index (GNDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:GNDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{G}}{{\rho\:}_{Nir}+{\rho\:}_{G}}$$\end{document} Calculating vegetation coverage through the contrast between the green and near-infrared bands.[8]Optimized Soil Adjustment Vegetation Index (OSAVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:OSVAI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{{\rho\:}_{Nir}+{\rho\:}_{R}+{l}_{2}}$$\end{document} Taking soil factors into consideration, excluding soil effects, and reflecting the chlorophyll content of vegetation.[20]Triangular Vegetation Index (TVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:TVI=0.5\times\:(200\times\:\left({\rho\:}_{Nir}+{\rho\:}_{G}\right)+120\times\:\left({\rho\:}_{Nir}-{\rho\:}_{G}\right))$$\end{document} Describing the radiative energy absorbed by thepigments as a function of the relative difference between redand NIR reflectance in conjunction with the magnitude ofreflectance in the green region[1]Renormalized Difference Vegetation Index (RDVI) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RDVI=\frac{{\rho\:}_{Nir}-{\rho\:}_{R}}{\sqrt{{\rho\:}_{Nir}+{\rho\:}_{R}}}$$\end{document} Combining the advantages ofDVI and NDVI for low and high vegetation coverages,respectively[1]Note: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{Nir}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R-edge}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{G}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{B}$$\end{document} are near-infrared wavelength reflectance, red-edge wavelength reflectance, red-light wavelength reflectance, green-light wavelength reflectance, and blue-light wavelength reflectance, respectively; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} are the soil conditioning coefficients, and usually \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document} is taken as 0.5, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} is taken as 0.16. Candidate vegetation index calculation methods and references. Describing the radiative energy absorbed by the pigments as a function of the relative difference between red and NIR reflectance in conjunction with the magnitude of reflectance in the green region[1] Combining the advantages of DVI and NDVI for low and high vegetation coverages, respectively[1] Note: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{Nir}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R-edge}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{R}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{G}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{B}$$\end{document} are near-infrared wavelength reflectance, red-edge wavelength reflectance, red-light wavelength reflectance, green-light wavelength reflectance, and blue-light wavelength reflectance, respectively; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} are the soil conditioning coefficients, and usually \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{1}$$\end{document} is taken as 0.5, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{2}$$\end{document} is taken as 0.16. The Pearson coefficient is commonly used to analyze the relationship between two continuous variables. Calculating the Pearson coefficient between two variables requires first computing their covariance and dividing it by the product of the standard deviations of the two variables. The specific formula is as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r=\frac{cov\left(X,Y\right)}{std\left(X\right)\times\:std\left(Y\right)}$$\end{document} Where X and Y represent two variables, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cov(X,Y)$$\end{document} represents the covariance of X and Y, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:std\left(X\right)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:std\left(Y\right)$$\end{document} represent the standard deviation of X and Y, respectively. For the 11 candidate vegetation indices and wheat LAI selected in this study, the Pearson correlation coefficients between each of the two indices were calculated separately using 107 sample data from the wheat measured LAI dataset. The results are presented in Fig. 2, where colors closer to the endpoints of the legend indicate stronger correlation between the two indices, while colors closer to the center of the legend indicate weaker correlation. Figure 2 was generated using Python 3.10.9 (https://www.python.org/downloads/). A correlation threshold of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|r\right|>0.5$$\end{document} was applied in this study to identify vegetation indices with meaningful relationships to LAI. As can be seen from Fig. 2, most of the vegetation indices selected in this study are correlated with LAI, with GNDVI having the highest correlation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|r\right|=0.8$$\end{document}), and others such as NDVI, NDRE, RVI, DVI, and OSAVI also have high correlation with LAI (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|r\right|$$\end{document} 0.7, 0.61, 0.68, 0.59, and 0.55, respectively). In addition, the vast majority of the candidate vegetation indices were not correlated with each other, but there existed a few combinations of vegetation indices with strong covariance, such as the combinations of RVI and NDVI, OSAVI and NDVI, GNDVI and NDRE, OSAVI and RVI, and RDVI and SAVI, which had correlation coefficients with an absolute value of close to 0.9, suggesting that there may exist a stronger covariance. This covariance affects the training of the model and reduces the importance of the features, thus decreasing the generalization performance. Therefore, it is necessary to screen the candidate vegetation indices. The 11 candidate vegetation indices selected in this study were sorted in ascending order based on their Pearson correlation coefficient absolute values with LAI, as illustrated in Fig. 3. Figure 3 was generated using Python 3.10.9 (https://www.python.org/downloads/). Adopting the method of adding features one by one, the indices were successively added to the feature set according to the sorting results. The updated feature set was then used to train LR, BPNN, and XGBoost models with default parameters. The coefficient of determination (R2) was adopted as the evaluation metric, utilizing a 10-fold cross-validation methodology. The mean R2 values of the models were recorded and summarized after each update, as presented in Fig. 4. Figure 4 was generated using Python 3.10.9 (https://www.python.org/downloads/). From the analysis of the figures, it was observed that in the LR, BPNN, and XGBoost models, the optimal number of feature set indices were 10, 9, and 8, respectively. The addition of the fourth vegetation index (NDRE) to the feature set led to a significant improvement in predictive performance for all three models. Upon introducing the tenth vegetation index (SAVI) to the feature set, the predictive performance of the BPNN and XGBoost models began to decline, except for a marginal enhancement in the LR model. Considering the performance of the three models and the correlation of the 11 vegetation indices with LAI, the top 8 vegetation indices (GNDVI, NDVI, RVI, NDRE, DVI, OSAVI, TVI, and EVI) were selected to construct the comprehensive indicator system for the input model. Fig. 2Pearson correlation coefficients of candidate vegetation indices with LAI. The figure presents the Pearson correlation coefficients between 11 candidate vegetation indices and wheat LAI using 107 sample data points. Colors closer to the legend endpoints represent stronger positive or negative correlations, while colors near the center indicate weaker correlations. This analysis identifies the indices most strongly related to wheat LAI, providing a basis for feature selection in model development. Figure 2 was generated using Python 3.10.9 (https://www.python.org/downloads/). Pearson correlation coefficients of candidate vegetation indices with LAI. The figure presents the Pearson correlation coefficients between 11 candidate vegetation indices and wheat LAI using 107 sample data points. Colors closer to the legend endpoints represent stronger positive or negative correlations, while colors near the center indicate weaker correlations. This analysis identifies the indices most strongly related to wheat LAI, providing a basis for feature selection in model development. Figure 2 was generated using Python 3.10.9 (https://www.python.org/downloads/). Fig. 3Positive order ranking of 11 candidate vegetation indices based on Pearson’s correlation. The figure shows the 11 candidate vegetation indices sorted in ascending order of the absolute values of their Pearson correlation coefficients with wheat LAI. This ranking highlights the strength of each index’s relationship with LAI, aiding in the selection of the most relevant indices for wheat growth monitoring. Figure 3 was generated using Python 3.10.9 (https://www.python.org/downloads/). Positive order ranking of 11 candidate vegetation indices based on Pearson’s correlation. The figure shows the 11 candidate vegetation indices sorted in ascending order of the absolute values of their Pearson correlation coefficients with wheat LAI. This ranking highlights the strength of each index’s relationship with LAI, aiding in the selection of the most relevant indices for wheat growth monitoring. Figure 3 was generated using Python 3.10.9 (https://www.python.org/downloads/). Fig. 4Model performance based on the incorporation of feature-by-feature method. The figure illustrates the mean R² values of LR, BPNN, and XGBoost models as vegetation indices were added to the feature set one by one in ascending order of their correlation with LAI. Models were trained with default parameters using 10-fold cross-validation, and R² values were recorded after each update. The results highlight the impact of feature addition on model performance and the differences in sensitivity across the three models. Figure 4 was generated using Python 3.10.9 (https://www.python.org/downloads/). Model performance based on the incorporation of feature-by-feature method. The figure illustrates the mean R² values of LR, BPNN, and XGBoost models as vegetation indices were added to the feature set one by one in ascending order of their correlation with LAI. Models were trained with default parameters using 10-fold cross-validation, and R² values were recorded after each update. The results highlight the impact of feature addition on model performance and the differences in sensitivity across the three models. Figure 4 was generated using Python 3.10.9 (https://www.python.org/downloads/). Optimization and selection of wheat growth monitoring model After constructing a wheat growth monitoring method based on a comprehensive indicator system, achieving accurate monitoring of wheat growth was pursued. This study conducted a performance comparison and model selection among three categories of models: logistic regression, backpropagation neural network, and extreme gradient boosting trees. By considering the performance of each model, the PSO-optimized XGBoost model, which exhibited the best performance, was ultimately chosen as the monitoring model. Logistic regression (LR) For wheat growth prediction, logistic regression can provide an intuitive probability interpretation, indicating the impact of different factors on wheat growth, helping decision makers to understand the contribution of each characteristic to the prediction results. Its formula is as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(y=1|x,\theta\:\right)=\frac{1}{1+{e}^{-{\theta\:}^{T}x}}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:y$$\end{document} is the label, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x$$\end{document} is the input feature, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} is the model parameter. Back propagation neural network (BPNN) The BPNN algorithm, as a mature neural network model, possesses excellent capabilities for approximating complex functions. It is particularly suitable for analyzing the complex linear and nonlinear relationships between growth and the numerous factors influencing it24. Its formula is as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{k}=f\left(\sum\:_{j=1}^{n}{w}_{{k}_{j}{x}_{j}}+{b}_{k}\right)$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{k}$$\end{document} is the input of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}th neuron, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{j}$$\end{document} represents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}th input feature, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{{k}_{j}}$$\end{document} denotes the connection weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{k}$$\end{document} is the bias term, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f(\bullet\:)$$\end{document} stands for the activation function. Extreme gradient boosting (XGBoost) The XGBoost model has the capability to capture complex nonlinear relationships, gradually reduce prediction errors, and enhance the accuracy of wheat growth prediction25. Additionally, it effectively handles high-dimensional data by employing feature selection and data dimensionality reduction, which helps identify the most significant features and address the issue of multicollinearity resulting from multiple influencing factors. The selection of XGBoost as the core model is further justified by its proven advantages in handling nonlinear relationships, multicollinearity, and high-dimensional feature spaces. Unlike traditional linear models or standard neural networks, XGBoost incorporates built-in regularization terms to control overfitting and performs automatic feature selection through its tree-based structure learning. Moreover, its scalability and computational efficiency—enabled by parallel processing—make it highly suitable for limited but high-quality remote sensing datasets, such as those used in this study. The XGBoost model’s objective function is defined as shown in Eqs. (7) and (8). The final objective function expression is presented in Eq. (9).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{t}=\sum\:_{i=1}^{n}L({y}_{i},{f}_{t-1}({x}_{i})+{h}_{t}({\varvec{x}}_{\varvec{i}}))+{\Omega\:}\left({h}_{t}\left(\varvec{x}\right)\right)$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Omega\:}\left({h}_{t}\left(\varvec{x}\right)\right)=\gamma\:{J}_{t}+\frac{\lambda\:}{2}\sum\:_{j=1}^{{J}_{t}}{\omega\:}_{tj}^{2}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{t}=-\frac{1}{2}\sum\:_{j=1}^{J}\frac{{\left(\sum\:_{{x}_{i}\epsilon{R}_{tj}}{g}_{i}\right)}^{2}}{\sum\:_{{x}_{i}\epsilon{R}_{tj}}{h}_{i}+\lambda\:}+\gamma\:J$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{t}$$\end{document} is the XGBoost objective function containing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} decision trees, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L(\bullet\:)$$\end{document} is the loss function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Omega\:}\left(\bullet\:\right)$$\end{document} is a regular term based on the tree complexity, which helps to reduce the overfitting risk of the model. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}_{t}$$\end{document} is the number of leaf nodes of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t}\left(\varvec{x}\right)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{tj}^{2}$$\end{document} is the optimal weight of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}th leaf node. Model optimization algorithms The Particle Swarm Optimization Algorithm (PSO) is a heuristic algorithm inspired by the foraging behavior of birds26. To facilitate the evolution of the PSO algorithm, it is necessary to define a fitness function to assess the quality of each particle’s position. According to the value of the fitness function, the particle updates its position \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} by the following formula:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i+1}={v}_{i}+{c}_{1}\times\:r\times\:\left({pbest}_{i}-{x}_{i}\right)+{c}_{2}\times\:r\times\:\left(gbest-{x}_{i}\right){v}_{i+1}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i+1}={x}_{i}+{v}_{i+1}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i+1}$$\end{document} denotes the speed of particle updating, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{i}(i=\text{1,2})$$\end{document}is the learning factor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r$$\end{document} is a random number, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{pbest}_{i}$$\end{document} is the individual optimal position of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}th particle, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:gbest$$\end{document} is the group optimal position. After each round of iteration, the individual optimal position and the group optimal position are updated. After many rounds of iteration, the particle population gradually converges to the optimal solution of the problem from the scattered disorder, at this time, the iteration can be stopped, and the optimal position of the group gbest is the optimal solution of the problem. In this study, PSO algorithm was applied individually to the LR, BPNN, and XGBoost models to optimize their respective hyperparameters. For LR, the regularization strength and learning rate were optimized. For BPNN, the learning rate, the number of neurons in the hidden layer, and the number of training iterations were tuned. For XGBoost, the optimization targeted the number of estimators (trees), learning rate, maximum tree depth, minimum split gain, and minimum child weight. The coefficient of determination (R²) was selected as the fitness function to guide the particle updates. During each iteration, particles representing different hyperparameter combinations were evaluated using 10-fold cross-validation. The fitness values determined whether the particle’s current state should update its personal best (pbest) or the global best (gbest). Through successive iterations, the PSO algorithm effectively searched the hyperparameter space and identified parameter sets that minimized prediction error and enhanced model stability. Model evaluation metrics and methods During the phase of constructing the wheat growth monitoring model, it is necessary to evaluate and compare the performance of several candidate models. This study employs two commonly used evaluation metrics for regression problems to assess the models: the coefficient of determination (R2) and the mean squared error (MSE). R2 reflects the extent to which the independent variable explains the variation in the dependent variable and is calculated by formula (12). MSE reflects the degree of deviation between the model’s predicted values and the actual values and is calculated by formula (13).12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}=1-\frac{\sum\:_{i}{\left({{\widehat{y}}_{i}-y}_{i}\right)}^{2}}{\sum\:_{i}{\left(\stackrel{-}{y}-{y}_{i}\right)}^{2}}$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:MSE=\frac{\sum\:_{i=1}^{n}{\left({{\widehat{y}}_{i}-y}_{i}\right)}^{2}}{n}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}$$\end{document} is the true value of the target variable for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}th sample, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{i}$$\end{document} is the model’s predicted value, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{y}$$\end{document} is the mean of the target variable for all samples. To ensure that the dataset division does not affect the model evaluation, 10-fold cross-validation is used, and the dataset is divided into the training set and validation set in the ratio of 9:1. There are a total of 107 samples in the dataset, and 90% of the samples are used for each training, and the remaining 10% is used as the validation set, which is rotated to perform the training for 10 times and to take the indicator mean values. The wheat growth monitoring model was constructed using the eight vegetation indices derived from the screening and the measured LAI data. Results and discussions Analysis of model prediction results In this research, the PSO algorithm is integrated with LR, BPNN, and XGBoost machine learning models individually. The hyperparameters that need optimization for each model are treated as variables to be solved, while the coefficient of determination R2 of the model serves as the fitness function value. Through iterative training cycles that guide particle evolution, the goal is to find the optimal set of hyperparameters that minimizes the model’s prediction error. This study has identified the key hyperparameters to be optimized for each model, as detailed in Table 3. Table 3Hyperparameter information to be optimized for LR, BPNN and XGBoost model.ModelHyperparameter nameSymbolRangeDefault valueOptimal valueLRRegularization parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,100}\right]$$\end{document} 1.09.15Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{L}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 1.00.03BPNNNumber of hidden layer neurons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,100}\right]$$\end{document} 100.0150.0Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{B}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 0.0010.01Iteration number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{B}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,300}\right]$$\end{document} 200.0400.0XGBoostIteration number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{X}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,300}\right]$$\end{document} 100.0300.0Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{X}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 0.300.04Maximum tree depth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mtd$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{3,40}\right]$$\end{document} 6.008.00Minimum node splitting gain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:[0,\:+{\infty\:}]$$\end{document} 0.000.30Minimum leaf node weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mw$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:[1,+{\infty\:}]$$\end{document} 1.002.60 Table 3Hyperparameter information to be optimized for LR, BPNN and XGBoost model.ModelHyperparameter nameSymbolRangeDefault valueOptimal valueLRRegularization parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,100}\right]$$\end{document} 1.09.15Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{L}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 1.00.03BPNNNumber of hidden layer neurons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,100}\right]$$\end{document} 100.0150.0Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{B}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 0.0010.01Iteration number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{B}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,300}\right]$$\end{document} 200.0400.0XGBoostIteration number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{X}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{10,300}\right]$$\end{document} 100.0300.0Learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{X}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\text{0,1}\right]$$\end{document} 0.300.04Maximum tree depth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mtd$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\text{3,40}\right]$$\end{document} 6.008.00Minimum node splitting gain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:[0,\:+{\infty\:}]$$\end{document} 0.000.30Minimum leaf node weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mw$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:[1,+{\infty\:}]$$\end{document} 1.002.60 Hyperparameter information to be optimized for LR, BPNN and XGBoost model. For the LR model, the regularization parameter can control the model’s overfitting, typically using L1 or L2 regularization. The learning rate is a hyperparameter that controls the step size for updating the model’s weights, which can slow down the model’s convergence speed, making it easier for the model to approach the optimal solution; however, a value that is too small will lead to excessively slow convergence. For the BPNN model, the number of neurons in the hidden layer refers to the number of neurons in the hidden layer of the neural network, which controls the complexity and expressive ability of the model. The learning rate determines the step size of model parameter updates at each iteration. An appropriate learning rate can speed up the convergence of the model and enable the model to achieve optimal performance within a reasonable time. The number of iterations refers to the number of training rounds, which is used to control the training time and accuracy of the model. For the XGBoost model, the number of iterations also refers to the number of decision trees included in the model. Increasing this number helps the model’s error converge, but too many iterations will significantly increase the training time. The step size of model parameter update in each iteration is also determined by the learning rate. The maximum tree depth specifies the maximum depth for splitting each decision tree. The minimum node splitting gain refers to the minimum gain required for each node split; if the node split gain is less than this parameter, further splitting is abandoned. The minimum leaf node weight specifies the minimum number of samples each leaf node must contain during the construction of the decision tree. These last three parameters are all used to prevent model overfitting. The larger their values, the more conservative the model, but at the same time, the model’s performance may also decline. To further visually demonstrate the model’s fitting effect and generalization performance on the dataset, a scatter plot of the actual versus predicted values on the wheat LAI dataset was created, as shown in Fig. 5. Figure 5 was generated using Python 3.10.9 (https://www.python.org/downloads/). In the figure, the horizontal axis represents the model’s predicted values for LAI, and the vertical axis represents the measured values of LAI. The diagonal black dashed line indicates where the predicted values are equal to the measured values. After performing 10-fold cross-validation, the performance of the default LR model was evaluated. The results showed a relatively low coefficient of determination R2 (R2 = 0.4669) and a high mean squared error MSE (MSE = 0.7852). To enhance the model’s performance, the PSO algorithm was employed to optimize the regularization parameter and learning rate of the LR model, with the optimized parameters listed in Table 3. Following optimization, the LR model’s performance significantly improved, with R2 increasing to 0.6058, representing a 29.7% enhancement compared to the unoptimized model. Simultaneously, MSE decreased to 0.5807, reducing by 26.04%. However, as depicted in Figure (b), the optimized LR model still exhibits high dispersion and a notable number of outliers in its scatter plot, indicating a substantial performance gap. Due to the evident nonlinear relationship between wheat LAI and vegetation indices, the LR model based on linear relationships struggles to effectively predict wheat growth on this dataset. In terms of the BPNN model, after 10-fold cross-validation, it was found that the model with default parameters had a high coefficient of determination R2 (R2 = 0.6980) and a low mean square error MSE (MSE = 0.4447). The parameters of the BPNN model are optimized by PSO algorithm and the optimized parameters are shown in Table 3. The performance of the optimized BPNN model is significantly improved, with R2 reaching 0.8092, an improvement of 15.93%, and MSE decreasing to 0.2810, a reduction of 36.81%. The comparison of scatter distribution plots (c) and (d) showed that the optimized BPNN model more accurately fitted the relationship between wheat LAI and real data. Although the optimized BPNN model is biased towards overestimating the wheat LAI value to a certain extent, the overall prediction error of the model is substantially reduced compared with the optimized LR model, which is a better performance and has certain reference value. This tendency may stem from the model’s limited regularization capability and higher sensitivity to sample noise compared to the XGBoost model. On the other hand, the XGBoost model exhibited impressive performance with default parameters. After conducting 10-fold cross-validation, it achieved an R2 of 0.7822 and an MSE of 0.3207. Upon optimizing the parameters of the XGBoost model using the PSO algorithm, its performance was further enhanced, and the optimized parameters are presented in Table 3. The optimized XGBoost model achieved an R2 of 0.9485, representing a 21.26% improvement, and an MSE of only 0.0757, marking a substantial reduction of 76.39%. A comparison of scatter plots (e) and (f) clearly demonstrates that the optimized XGBoost model more accurately predicts wheat LAI values, and the model’s stability is enhanced. The scatter points for wheat LAI predictions by the optimized XGBoost model are evenly distributed around the diagonal line, indicating that the model avoids overestimation or underestimation for most samples. This suggests that the XGBoost model is more stable compared to the aforementioned two models, and it is less susceptible to performance fluctuations caused by dataset noise. As a result, the XGBoost model is better suited for accurate inversion of wheat LAI. Fig. 5Results of individual model predictions. (a, c, e): using default parameters. (b, d, f): using PSO optimization parameters. The figure presents the relationship between the measured and predicted wheat LAI values. The horizontal axis represents the model’s predicted values for LAI, and the vertical axis represents the measured values of LAI. The diagonal black dashed line indicates where the predicted values are equal to the measured values. The scatter plot visually demonstrates the model’s fitting effect and generalization performance, with points closer to the dashed line reflecting higher prediction accuracy. Figure 5 was generated using Python 3.10.9 (https://www.python.org/downloads/). Results of individual model predictions. (a, c, e): using default parameters. (b, d, f): using PSO optimization parameters. The figure presents the relationship between the measured and predicted wheat LAI values. The horizontal axis represents the model’s predicted values for LAI, and the vertical axis represents the measured values of LAI. The diagonal black dashed line indicates where the predicted values are equal to the measured values. The scatter plot visually demonstrates the model’s fitting effect and generalization performance, with points closer to the dashed line reflecting higher prediction accuracy. Figure 5 was generated using Python 3.10.9 (https://www.python.org/downloads/). Selection results of wheat growth monitoring model Summary of the coefficients of determination (R2) and mean squared errors (MSE) for LR, BPNN, and XGBoost models before and after optimization can be found in Table 4. As shown in Table 4, considering the predictive performance and generalization ability of the models on the wheat LAI dataset, the performance of the XGBoost model is the most prominent among the three models, with the best fitting and error metrics before and after optimization. Among them, the XGBoost model optimized by the PSO algorithm has an R2 as high as 0.94 and an MSE below 0.1, indicating that the PSO-optimized XGBoost model has a high fitting accuracy for the wheat LAI dataset and can excellently perform the inversion of wheat LAI. The performance of the other two models is slightly inferior. Among them, the unoptimized LR model has the poorest overall performance, possibly due to the significant nonlinear relationship between wheat LAI and vegetation indices, as well as the correlations among vegetation indices, making the LR model based on linear relationships unable to perform well in predicting wheat growth on the dataset. However, the performance of the LR model improved after PSO optimization. The performance of the BPNN model is inferior to that of the XGBoost model, but it still clearly surpasses the LR model. As shown in Fig. 5, it can also be observed that the performance of the optimized BPNN model is relatively excellent, but it has a tendency to overestimate the fitting of wheat LAI and lacks stability. In summary, the XGBoost model optimized by the PSO algorithm has high prediction accuracy of wheat LAI, high stability, small fluctuation by the data set, and is suitable as the optimal model for wheat growth monitoring27–36. Table 4Performance comparison of LR, BPNN and XGBoost models before and after optimization.Model nameDefault parametersPSO optimization parameters R 2 MSE R 2 MSELR0.46690.78520.60580.5807BPNN0.69800.44470.80920.2810XGBoost 0.7822 0.3207 0.9485 0.0757 Note: Bold is the optimal value for each column of data, where R2 is the maximum value and MSE is the minimum value. Table 4Performance comparison of LR, BPNN and XGBoost models before and after optimization.Model nameDefault parametersPSO optimization parameters R 2 MSE R 2 MSELR0.46690.78520.60580.5807BPNN0.69800.44470.80920.2810XGBoost 0.7822 0.3207 0.9485 0.0757 Note: Bold is the optimal value for each column of data, where R2 is the maximum value and MSE is the minimum value. Performance comparison of LR, BPNN and XGBoost models before and after optimization. Note: Bold is the optimal value for each column of data, where R2 is the maximum value and MSE is the minimum value. Conclusions To improve the accuracy and effectiveness of wheat growth monitoring, this study proposes a monitoring method based on a comprehensive vegetation index system. The following conclusions were drawn through vegetation index screening and model selection: (1) Compared with previous studies, this study systematically evaluated multiple vegetation indices and determined that most vegetation indices were correlated with leaf area index (LAI), among which the green normalized difference vegetation index (GNDVI) had the highest absolute correlation value of 0.8, ranking first. (2) In response to the problem of early studies ignoring the collinearity between vegetation indices, we strictly screened and finally selected 8 optimal indices (GNDVI, NDVI, RVI, NDRE, DVI, OSAVI, TVI and EVI) to build a more reliable monitoring system. (3) This study determined the optimal model for wheat growth monitoring based on data from specific regions and time frames, achieving good prediction results. However, to address the limitations in generalizability, future research will incorporate transfer learning techniques to adapt the model to new regions or crop varieties using small amounts of labeled data. Moreover, datasets covering multiple wheat growth stages across various geographic and climatic zones will be collected to enhance the robustness and applicability of the model. (4) The model’s current focus is limited to wheat. In future studies, transfer learning will also be explored to adapt the model to other crops, leveraging pre-trained models and fine-tuning them on smaller crop-specific datasets. Additionally, multi-source data—including high-resolution UAV images, additional satellite platforms (e.g., PlanetScope, Gaofen), and IoT-based field sensors—will be considered to enrich input features and further improve accuracy. (5) To support real-world deployment, future work will investigate lightweight model optimization for real-time prediction in resource-constrained environments. Data augmentation techniques such as image rotation, noise injection, and vegetation index perturbation will be explored to expand training data and improve model generalization. Integration with low-cost data acquisition platforms and edge computing devices will also be examined to facilitate scalable, efficient wheat growth monitoring systems.