The winter wheat experiment took place in Chuzhou City, located in Anhui Province, China (32°48′52″N,117°46′7″E) from 2020 to 2021 ( Figure 1A ). This region, positioned in the mid-lower reaches of the Yangtze River, experiences a subtropical monsoon climate. It is characterized by a moist weather pattern, distinct four seasons, an average annual temperature of 15.4°C, and an annual rainfall between 1000-1100mm. The region records an average of 144 rainy days per year and approximately 210 frost-free days annually.

The field experiment consisted of 36 plots, each measuring 16m² (2×8 m). It involved three winter wheat varieties with stable high-yield potential (V1: Huaimai 44, V2: Yannong 999, and V3: Ningmai 13) and four levels of nitrogen fertilizer application (N0-N3: 0, 100, 200, 300 kg/ha), with each treatment replicated three times ( Figure 1C ). The fertilization scheme involved the application of phosphorus (P=90kg/ha) and potassium (K=135 kg/ha) as base fertilizers before sowing, with a 6:4 ratio of nitrogen fertilizer applied pre-sowing and at the jointing stage. The sowing date was November 7, 2020, with the winter wheat manually strip-sown at a row spacing of 30 cm. The harvest took place on June 3, 2021, and other field management measures were based on local high-yield cultivation practices. Throughout the entire growth stage of the winter wheat, there were no occurrences of diseases, pests, or weed hazards, and the field environment was good, with no instances of drought or waterlogging.

The experiment utilized a Phantom 4 Multispectral (P4M) RTK (DJI Technology Co., Shenzhen, China) UAV to collect multispectral (MS) data during the wheat heading stage on April 18, 2021, the flowering stage on April 29, 2021, and the late filling stage on May 24, 2021. The UAV is equipped with five monochrome sensors, each with 2.08 million pixels: Blue, Green, Red, Red edge, and NIR, with center wavelengths of 450 ± 16nm, 560 ± 16nm, 650 ± 16nm, 730 ± 16nm, and 840 ± 26nm, and bandwidths of 20nm, 20nm, 10nm, 10nm, 40nm, respectively. It is also equipped with a real-time kinematic (RTK) system, which allows for vertical positioning accuracy of ±1.5 cm and horizontal positioning accuracy of ±1cm, enabling the P4M to obtain high-precision spectral and texture information (Oktay and Coban, 2017). Flights were conducted between 11:00 and 13:00 on clear, windless days with stable solar radiation intensity. Four radiometric calibration panels with known reflectance were deployed on the ground for radiometric calibration during subsequent data processing ( Figure 1B ). The DJI GS PRO software (https://www.dji.com/cn/ground-station-pro/) was used to pre-plan the flight route, relying on the UAV’s autopilot system to execute the predefined flight plan (Şahin et al., 2022). Each flight lasted 20 minutes, with a flight altitude of 30 m, a flight speed of 2m/s, forward and side overlap of 90% and 85% respectively, and an image resolution of 1600×1300 pixels.

The UAV MS images collected at each growth stage were imported into the PIX4Dmapper software (version 4.4.12, Pix4D SA, Prilly, Switzerland) for image stitching. The initial step involved aligning the images using the feature point matching algorithm. Subsequently, a dense point cloud and texture mesh were generated by utilizing both the UAV image data and location information. To improve data quality, this study carried out radiometric correction on the MS images using the empirical line method (ELM) and the image information of ground radiometric calibration panels with known reflectance (Di Gennaro et al., 2022). A shapefile was generated using ArcGIS 10.2 (Environmental Systems Research Institute, Inc, RedLands, CA, USA), and plot areas were delineated based on the orthoimage of the experimental area.

Field measurements at each stage were conducted prior to the UAV flights on the same day, with chlorophyll values based on SPAD determined for winter wheat across the 36 plots. Winter wheat SPAD was measured using a SPAD-502 chlorophyll meter (Konica Minolta Optics Inc., Osaka, Japan) (Jiang et al., 2022). For each plot, three average-growth wheat plants were selected, and SPAD readings were taken at the 1/6, 3/6, and 5/6 length of the flag leaves. The average of the 9 SPAD values in the sample plot was used as the actual SPAD measurement in the sample plot. A total of 324 times were collected in each growth stage, and the measured SPAD data of 36 sample plots were calculated.

UAV multispectral remote sensing data was analyzed in order to identify the most effective approach to maximize improvements in SPAD estimates during later stages of growth ( Figure 2 ). The features selection strategy is used to select SF and TF that are sensitive to SPAD. SPAD prediction model is further constructed to check the difference between implementing and not implementing the features selection strategy, different types of feature subsets are fused in the features fusion strategy to construct SPAD estimation model together.

Based on R language version 4.1.3 (R Foundation, Vienna, Austria), the support vector machine regression (SVR) algorithm was used to predict winter wheat SPAD. The idea of SVR originated from the support vector machine. It transforms the input low-dimensional sample set into a high-dimensional space and finds a hyperplane that can be closest to all feature sample sets to implement regression, minimize structural risks, and improve sample discreteness (Smola and Schölkopf, 2004). In this study, the SPAD data set of winter wheat in each growth stage was divided into a calibration set and a validation set. The dataset was randomly sampled, 2/3 was used for model training (Calibration), and 1/3 was used to evaluate the performances (Validation) ( Table 3 ). Three statistical indicators are used to test the machine learning model: the coefficient of determination (R²), root mean square error (RMSE), and the ratio of performance to deviation (RPD) (Bruning et al., 2019). Their calculation formulas are presented as Equations 1–3.

In the formula, y i ^ and y i are the observed values and measured values of SPAD, y ¯ is the average of SPAD observed values, n is the sample size, and SD is the standard deviation of reference values.

Figure 3 illustrates the dynamic changes in SPAD for wheat under different nitrogen treatments, varieties, and growth stages. As observed in the Figure 3 , SPAD exhibits noticeable differences across various nitrogen gradient treatments, generally trending upward with increasing nitrogen application, peaking at the N3 level. The performance of SPAD varies across growth stages, initially showing a slight increase followed by a sharp decrease. Notably, the SPAD values are highest during the Flowering stage and lowest in the Late filling stage. Furthermore, no significant differences in SPAD were found among different wheat varieties, with the trends being relatively consistent across various periods and nitrogen treatments.

In this study, 30 spectral features and 160 texture features were extracted from UAV multispectral images. Not all features are helpful for wheat SPAD estimation, and redundant features may affect model accuracy. Optimizing remote sensing features sensitive to SPAD using features selection strategy ( Figure 4 ). As can be seen from Figures 5A–C , during the heading stage, three identical feature sets were selected, encompassing all five types of spectral features. In the flowering stage, spectral feature types IV and V performed the most effectively, with C-SF and CT-SF selecting all features from types IV and V, while R-SF predominantly chose type IV as the optimal variable set. In the late filling stage, types IV and V continued to excel, and the three feature sets showed a high degree of consistency in favoring spectral feature types IV and V. The commonality between spectral feature types IV and V is their composition, which includes participation from the near-infrared band.

The results of texture features band type selection for the three stages are shown in Figures 5D–F . During the heading stage, the R, RE, and NIR bands showed better texture performance, far exceeding the B and G bands. During the flowering stage, the trend of texture feature band types selected by the three methods was consistent, with the Red and NIR bands showing better performance. During the late filling stage, the texture feature band performances were consistent across the three selection methods, with no significant differences among the five texture feature band types. The texture performance of the R and NIR bands was relatively stable, and the number of retained features after selection was significantly higher than that of other bands, except during the late filling stage where there was no significant difference among the five bands.

Wheat exhibits significant SPAD variation during the heading, flowering, and late filling stages, with a close relationship between SPAD and nitrogen treatment. Specifically, wheat SPAD exhibits a gradual upward trend with increased field nitrogen gradient treatments. This is because nitrogen is one of the essential nutrients for plant growth and development (Wang W et al., 2021) and a crucial component of chloroplasts in leaves. SPAD, as a measure of relative chlorophyll content in leaves, is inevitably closely related to nitrogen. During the heading and flowering stages, wheat plants absorb nitrogen from the soil through their roots. This nitrogen is then translocated to various parts of the plant, including the leaves (El-Hendawy et al., 2017). In contrast to the heading and flowering stages, SPAD in wheat drops sharply during the late filling stage. This may be due to nitrogen translocation within the plant at this time. During the filling stage, nitrogen from the leaves gradually moves into the grains, a process accompanied by a decline in photosynthetic capacity and senescence of the leaves (Bowman et al., 2015). Leaves begin to turn yellow, and the chlorophyll content decreases sharply, leading to lower SPAD values.

Spectral features have been verified to be closely associated with crop agronomic traits and are considered important indicators for estimating crop phenotypic information (Duan et al., 2019). It is well known that the sensitivity of spectral data to vegetation decreases during the later growth stages of crops (Zhu et al., 2023). This can be attributed to the canopy closure phenomenon and the complex spectral mixing and canopy structure effects that affect the sensitivity of spectral data to vegetation (Li R et al., 2022). Additionally, the canopy information generated by the spike emergence further exacerbates this effect (Wang et al., 2022). In this study, it was found that the NIR spectral bands combined with other spectral bands during the later growth stages of crops were the most frequently selected spectral features after optimal treatment ( Figures 5 , 10 ). This finding is consistent with the conclusion reached by Yin et al. (2023), that the sensitive spectral features for estimating SPAD in winter wheat are combinations of near-infrared and other bands using three features selection methods. This may be related to the sensitivity of the NIR bands to the growth status of crops. As the growth stage progresses, nitrogen in winter wheat leaves gradually transfers to the spike organs, the leaves wilt, and the appearance of wheat plants changes, leading to a sharp change in reflectance in the near-infrared bands. Zheng et al. (2018) found that the near-infrared spectral information was most important when estimating leaf nitrogen content (LNC) using mean squared error (MSE) to evaluate variable importance. LNC is strongly correlated with SPAD, and the results of this study are similar to their conclusion, indicating the great potential of near-infrared spectral data for estimating SPAD during the later growth stages of winter wheat.

In this study, the texture features of the red and near-infrared bands maintained a high sensitivity to SPAD information during the reproductive growth stage of wheat ( Figures 5 , 11 ). This is consistent with the conclusion reached by Bai et al. (2022), that the texture features of the red band can better predict soybean yield compared to the blue and green bands, and with the conclusion reached by Lu et al. (2021), that extracted texture features from the near-infrared band obtained by multispectral sensors are more effective in predicting potassium content in rice plants than other bands. This may be because the texture features of the red and near-infrared bands can better describe the spatial distribution changes of the crop canopy, providing valuable information for estimating crop phenotypes. It also demonstrates the feasibility of using UAV equipped with RGB sensors (including the red band) or multispectral sensors (including the red and near-infrared bands) to monitor crop phenotypic information using texture features. Furthermore, previous studies that used texture features to construct crop phenotyping models mostly considered the relationship between the mean (MEAN) of texture feature indices and crop phenotypes (Fu et al., 2021; Liu et al., 2022a; Luo et al., 2022; Zhou et al., 2022), while rarely exploring other statistical indices. In this study, further analysis of statistical indices of texture features ( Figures 11 , 12 ) extracted four different statistical indices based on GLCM features for each sample area: MEAN, SD, MAX, and MIN. The commonly used MEAN feature consistently maintained a leading position among the four statistical indices ( Figure 12 ), but its importance gradually decreased as the growth stage progressed ( Figure 11 ). The changing trend of importance for other indices was opposite to MEAN. From the heading stage to the late filling stage, their importance gradually increased and reached a peak in the late filling stage, with SD showing the best trend. This indicates that other statistical indices also have the potential to estimate SPAD in wheat, especially SD. This is consistent with the conclusion reached by Bai et al. (2022), that using texture features incorporating SD information achieved good accuracy and precision in predicting early soybean yield. The results of this study indicate that the SD information of texture features also has the potential to estimate crop SPAD.

Features selection methods play a significant role in machine learning and can directly impact the predictive model (Kursa and Rudnicki, 2010). When estimating crop phenotypes using remote sensing features, selecting appropriate remote sensing features through features selection strategy can improve model performance and eliminate interference (Lin et al., 2021; Xu et al., 2023). In this study, two features selection methods, Boruta and RFE, were used to select optimal spectral and texture features and build models. Additionally, a comparison of model accuracy before and after features selection was conducted to further analyze the mechanism of features selection methods and select suitable methods for crop phenotype studies. From Figure 13 , it can be observed that the accuracy of the SPAD estimation model constructed using the subset of variables after features selection is generally better than that of the original variable set. However, there are also some exceptions, as the Boruta method shows a slight decrease in model accuracy ( Figure 13 , Late filling, TF and C-TF). This may be related to the principle of the Boruta method. Boruta cannot identify and explore features with low individual value but high combined value, leading to the removal of low-value features and reducing or eliminating the high value generated by feature combinations, thus resulting in decreased model accuracy. Furthermore, Boruta focuses on selecting all features related to SPAD (Kursa and Rudnicki, 2010), overlooking the relationship between features, which may lead to multicollinearity among features and subsequently reduce model performance.

Therefore, features selection techniques should not only consider the relationship between features and SPAD but also take into account the relationship between features. RFE does this well by continuously combining subsets of variables to build models in order to find the optimal variable set with the minimum RMSE. For example, Wang J et al. (2021) used two features selection methods, RF and r, to construct RF and SVR models for estimating winter wheat SPAD, and compared them with the RFE features selection method. They found that the RFE features selection method combined with RF and SVR models had higher predictive accuracy. Lin et al. (2021) compared three features selection methods, Boruta, Sequential Forward Selection (SFS), and Permutation Importance-Recursive Feature Elimination (PI-RFE), and found that the PI-RFE method had the best dimensionality reduction effect on multidimensional feature sets, greatly improving the accuracy of FVC inversion. In this study, whether it was at the heading stage or flowering stage, the RFE features selection method combined with the SVR algorithm provided the best SPAD prediction model accuracy. In the late filling stage, the RFE-SVR combination still showed high SPAD estimation ability ( Figures 7 , 8 ). Overall, compared with the Boruta method, the RFE combined with the SVR algorithm can achieve higher model accuracy and stronger model performance. RFE takes into account the interrelationships between features and between features and SPAD, thereby improving the performance of prediction model.

Features fusion strategy involve merging different types of features into a more optimal feature set, combining features with different features to construct models, and have the potential to mitigate the impact of spectral saturation effects, enhance model performance and increase the interpretability of prediction models (Zhou et al., 2022; Sun et al., 2023). This explains why the optimal SPAD estimation models in the late growth stage of winter wheat are constructed using features fusion strategy. This study comprehensively reports the essence of the impact of features fusion strategy on model performance through two features fusion approaches. When only implementing features fusion strategy ( Table 4 ), it was found that not all feature fusions lead to improved accuracy. Blindly combining features together to build models may have a counterproductive effect, causing data redundancy and reducing model performance (Son et al., 2015; Li et al., 2020). This conclusion is in line with the discoveries of Zhou L et al. (2023), who observed a decrease in model accuracy when agronomic practice information (API) was fused with spectral and texture features to construct a rice yield estimation model. The essence of the impact of features fusion strategy on model accuracy lies in the contribution of multi-characteristic features to the model and the interactions between features. This also explains why there are differences in predictive model results when different variables are fused ( Figure 8 ). Similar to the findings of Liu J et al. (2022), who estimated rice nitrogen use efficiency (NUE) using different numbers of spectral features and concluded that increasing the number of variables does not necessarily improve model accuracy, this study also discovered that there is no essential correlation between model accuracy and the quantity of features ( Table 4 ; Figures 4 , 6 – 8 , 12 , 13 ). Adding variables does not necessarily make the model more robust, and features selection strategy are powerful means to address these issues.

After combining features fusion strategy with features selection strategy, the stability, accuracy, and robustness of the SPAD estimation models for winter wheat are at the forefront ( Figure 9 ). As the growth stage progresses, the improvement in model accuracy gradually increases. The highest increase in model accuracy is observed in the late filling stage, where compared to models using only initial spectral or texture features, R² _Val increased by 0.092 to 0.202, RMSE_Val decreased by 0.076 to 4.916, and RPD_Val increased by 0.237 to 0.960. Sun et al. (2023) developed maize LAI estimation model using SVM, RF, BPNN, and PLSR algorithms by selecting five spectral features and three texture indices through correlation analysis. The results showed that the SVM algorithm with multi-variable fusion achieved the highest accuracy. Compared to SVM models using only spectral or texture features as input variables, R² _Val increased by 0.023 to 0.192, RMSE_Val decreased by 0.015 to 0.036, and RPD_Val increased by 0.074 to 0.746. Ma et al. (2022) used four features selection methods, including correlation coefficient, MIC, RF, and RFE, to select three spectral features and three texture features to construct a cotton yield estimation model. The results showed that the RFE_ELM model based on the fusion of spectral and texture features achieved the highest accuracy, with an increase in R² _Val of 0.073 to 0.187 compared to the non-fusion model. The findings of this study surpass those of prior research, indicating that the combination of features fusion strategy and features selection strategy is an effective method to mitigate the impact of spectral saturation effects and improve the accuracy of crop phenotype estimation, fully tapping into the potential of multi-characteristic features in estimating crop phenotypes.

This study conducted field experiments on winter wheat using three locally stable and high-yielding varieties (Huaimai 44, Yannong 999, and Ningmai 13) and four nitrogen fertilizer application levels (N0-N3: 0, 100, 200, 300 kg/ha). In the future, we will systematically supplement the dataset and conduct field experiments on more varieties and nitrogen fertilizer treatments. At the same time, because of the swift advancement in sensor technology, there are still limitations in building SPAD prediction models relying solely on single sensor data. In the future, we plan to incorporate RGB and hyperspectral data into the SPAD prediction model. The machine learning algorithms employed in this study are relatively single, and in the future, we will introduce multiple algorithms for comparison and select the most suitable algorithm for SPAD estimation. Furthermore, the use of multi-source remote sensing information and background effect removal can also improve the accuracy of crop phenotype estimation based on UAV. In the future, we will combine them to further improve the accuracy of SPAD estimation for winter wheat. Although the combination of features fusion strategy and features selection strategy performs well in SPAD estimation during the later growth stage of wheat, the results of this study only come from small-scale experiments. In the future, we will conduct large-scale integrated winter wheat SPAD remote sensing monitoring research using satellite and UAV platforms.