The overall methodology is summarized in Fig. 4. UAV-derived LiDAR, multispectral (MS), and thermal infrared (TIR) data were processed to generate complementary structural and spectral features for aboveground biomass (AGB) estimation. To ensure a consistent comparison among sensor types, all datasets were georeferenced, resampled, and aligned to a common spatial resolution before extracting zonal statistics corresponding to the destructive biomass sampling plots. These standardized feature sets were subsequently evaluated within a common artificial neural network (ANN) framework.

LiDAR point clouds were used to derive CH, multi-layer GF, and INT, with INT representing a wavelength-dependent return signal (903 nm) influenced by both canopy structure and vegetation surface properties, while MS and TIR imagery provided spectral reflectance and thermal radiance information (Fig. 5). Ground points were identified using the Cloth Simulation Filter (CSF) (Zhang et al., 2016) with a 4 cm grid resolution corresponding to the ranging precision of the LiDAR system. These ground returns served as the reference surface for both gap fraction calculations and intensity normalization. All raster products were converted to tabular feature sets with geographic metadata removed and normalized to a 0–1 range prior to ANN modelling. A tiered evaluation strategy was adopted to isolate the contribution of progressively more complex sensor information. The analysis began with LiDAR gap fraction using multiple vertical segmentation schemes to identify the optimal representation of canopy density (Sect. 5.1). The best-performing GF configuration was subsequently combined with crop height and normalized intensity to evaluate complementary LiDAR information (Sect. 5.2). Multispectral reflectance (Sect. 5.3) and thermal observations (Sect. 5.4) were then incrementally incorporated to quantify the additional value of multi-sensor fusion while maintaining an identical modelling framework. Following model training and validation, the selected ANN models were applied to co-registered and stacked predictor rasters using the compute() function to generate spatially continuous AGB predictions. Geographic information was subsequently reattached and the resulting biomass maps were exported as GeoTIFFs (UTM Zone 32N) for visualization and further spatial analysis.

Although multispectral reflectance and normalized LiDAR intensity are both influenced by canopy structure, they were grouped as signal-based features because their recorded response depends on the interaction of electromagnetic radiation at specific wavelengths with canopy optical and surface properties. In contrast, crop height and gap fraction are derived primarily from the geometric distribution of LiDAR returns and therefore represent predominantly structural canopy attributes.

CH was derived by subtracting the digital terrain model (DTM) from the digital surface model (DSM), with the DTM generated from UAV LiDAR data collected under bare-soil conditions prior to planting. This approach minimizes interpolation errors and improves accuracy (Cao et al., 2019). Several CH-derived metrics (e.g., mean, maximum, standard deviation) were tested (see Fig. B1) but the maximum CH per pixel was selected because it consistently showed stronger correlations with AGB in vertically structured, dense crops such as wheat (Bazrafkan et al., 2023). By focusing on a single robust height metric, feature redundancy and model complexity were reduced, helping to limit overfitting while maintaining a consistent comparison among sensor feature combinations (Ruwanpathirana et al., 2024). Correlations of alternative CH metrics with AGB are provided in Fig. B1.

Traditional GF metrics estimate canopy density by comparing ground returns to total LiDAR returns, often within Beer–Lambert formulations for LAI estimation (Heiskanen et al., 2015; Richardson et al., 2009). However, these approaches reduce the three-dimensional canopy to a single integrated metric and can become less reliable in dense wheat canopies where laser penetration to the ground is limited (Sabol et al., 2014). To preserve vertical canopy information, we extend the GF concept by calculating normalized GF values independently for multiple canopy layers rather than relying on a single ground-based estimate. A related approach, the three-dimensional point index (3DPI), similarly partitions canopy returns into vertical layers but combines them using an empirically calibrated extinction coefficient to estimate LAI or AGB (Jimenez-Berni et al., 2018). In contrast, the normalized GF layers developed here (see Fig. 5) were incorporated directly into the ANN as independent predictor variables, allowing the model to learn the relative contribution of each canopy stratum throughout the growing season without requiring additional parameterization or extinction coefficient calibration.

Four vertical discretization schemes were evaluated, consisting of a conventional ground-only GF, three 30 cm layers, five 20 cm layers, and ten 10 cm layers. Each configuration was further assessed using horizontal grid resolutions of 10, 20, and 30 cm, selected based on the LiDAR ranging precision (∼4 cm), point density, winter wheat crop height (approximately 1–1.2 m), and subplot dimensions. For comparison, additional LiDAR canopy profile representations, including the Vertical Distribution Index (VDI), Height-Weighted Profile Index (HWPI) and 3DPI, were also evaluated in the Results. Mathematical formulations for these metrics are provided in Appendix A. The objective was not to identify a universally optimal parameterization but to systematically investigate how different representations of canopy density influence AGB estimation under the sensor characteristics, acquisition settings, and crop conditions of this study.

LiDAR intensity represents the strength of the returned laser signal at 903 nm (NIR) and, unlike crop height and gap fraction, provides a wavelength-dependent response influenced by both canopy structure and vegetation surface properties (Eitel et al., 2014). To minimize footprint variation, light dispersion, and incidence angle effects, LiDAR data were acquired using a maximum scan angle of ±25°, providing a relatively consistent viewing geometry across the experimental field (Li et al., 2016).

The temporal use of LiDAR intensity is challenging because the recorded signal is influenced not only by vegetation properties but also by environmental conditions, including soil moisture, illumination, and acquisition geometry. In the absence of dedicated reflectance targets, ground returns provide a practical reference for improving temporal consistency. Previous airborne LiDAR studies have employed ground-return normalization or related indices, such as the Light Penetration Index (LPI), to improve the estimation of vegetation structural properties including LAI (Solberg et al., 2006; You et al., 2017; Sasaki et al., 2016), while Luo et al. (2018) demonstrated the importance of accounting for ground intensity when relating LiDAR intensity to vegetation characteristics.

Several intensity representations were initially evaluated for their relationship with AGB and canopy properties (Fig. B1). Ground-return normalization produced the most consistent relationships across the spatial and temporal variability represented within the experiment and was therefore selected for subsequent analyses. The mathematical formulation of the LiDAR intensity metric used is provided in Appendix A. Although this approach does not constitute a full radiometric calibration, it provides a practical relative normalization that improves the comparability of canopy intensity measurements among acquisition dates while maintaining a simple processing workflow.

MS data were collected in the red (R), red-edge (RE), and near-infrared (NIR) bands, which capture complementary information related to canopy reflectance and vegetation condition and form the basis of many commonly used vegetation indices (VIs) (Han et al., 2019). The RE band is particularly valuable because it is generally less susceptible to saturation at high biomass than the R band, providing complementary information when combined with NIR reflectance (Gitelson et al., 2003). Rather than deriving predefined VIs, the raw reflectance bands were used directly as ANN inputs. This approach avoids introducing redundancy among highly correlated indices while allowing the ANN to evaluate the relative contribution and nonlinear interactions of the individual spectral bands within a consistent modelling framework. Using the original reflectance values also preserves band-specific information that may otherwise be reduced through index formulation and facilitates a more direct comparison with LiDAR-derived structural and intensity features.

TIR data were used to derive both land surface temperature (LST) and the crop water stress index (CWSI). LST provides an absolute measure of canopy thermal status but is strongly influenced by environmental conditions, particularly air temperature, making direct comparisons across dates and growth stages challenging. To reduce this temporal dependence, LST was additionally converted to CWSI by normalizing canopy temperature relative to seasonal “wet” and “dry” reference limits and air temperature (Katimbo et al., 2022). The mathematical formulation of the CWSI used in this study is provided in Appendix A. Both LST and CWSI were evaluated as ANN inputs to compare the contribution of absolute canopy temperature and a normalized thermal stress indicator for AGB estimation. Including both metrics enabled assessment of whether normalization improved the information content of TIR observations and provided a direct comparison between absolute and relative thermal representations within the common sensor fusion framework.

A multilayer perceptron ANN with backpropagation was implemented using the neuralnet package in RStudio. Model complexity (number of hidden layers and neurons) was optimized through iterative sensitivity testing, where one to two hidden layers with 2–14 neurons were evaluated. A logistic activation function was applied to the hidden neurons, while a linear activation was used for the output neuron to support continuous regression. Training employed resilient backpropagation, the default optimizer in neuralnet, which performs full-batch weight updates each iteration. Accordingly, batch size is not applicable in this framework. Training was capped using stepmax = 1×106 (maximum number of weight updates) and threshold = 0.01 (convergence criterion). These settings effectively provide a high maximum number of training iterations while terminating optimization once the convergence criterion is satisfied.

A total of 86 destructive AGB samples were available for model development. This sample size reflects the constraints of destructive biomass collection, which is highly labor-intensive and logistically limited when covering an entire growing season in parallel with other field experiments (Han et al., 2019). The number and quality of destructive samples collected are comparable to those reported in UAV studies (Bendig et al., 2014; Morgan et al., 2025; Smith et al., 2024). The samples spanned seven campaigns representing major phenological stages and a wide range of canopy conditions, resulting in substantial structural and spectral variability across the growing season. Given the modest sample size, the ANN architecture was intentionally kept shallow and all predictor variables were normalized to reduce the risk of overfitting.

Model performance was evaluated using two complementary validation strategies. An initial 70/30 training-testing split was used as the primary independent holdout evaluation for comparing sensor features and feature combinations (Bendig et al., 2014; Han et al., 2019; Morgan et al., 2025; Smith et al., 2024). The training and testing subsets contained observations from multiple nitrogen treatments, wheat genotypes, and measurement campaigns, with testing samples distributed across the experimental field. Model performance was assessed using RMSE and R2 by comparing predictions for both the training and testing subsets, where small differences between training and testing performance indicate good generalization, whereas larger discrepancies may indicate overfitting.

A stratified 10-fold cross-validation analysis was additionally performed for selected feature combinations as a supplementary assessment of sensitivity to data partitioning. Within each fold, approximately 90 % of the observations were used for training and 10 % for validation. To account for stochastic ANN initialization, 20 networks were trained using identical architecture and training settings within each fold. Because ANN performance can vary between runs due to random initial weights, repeated initializations were used to reduce this variability and improve the consistency of comparisons among sensor combinations. The network with the lowest validation RMSE was then selected from these repeated initializations. Because the validation subset was used to select the best-performing initialization, the reported cross-validation performance may be slightly optimistic and should therefore be interpreted primarily as a comparative assessment of sensor feature combinations rather than as an estimate of model generalizability.

The cross-validation folds were stratified by measurement date but were not spatially blocked. Consequently, spatially adjacent subplots could occur in both training and validation subsets, and similar spatial dependence may also have been present within the initial 70/30 training-testing split. However, the dataset incorporated substantial variability through repeated measurement campaigns spanning the growing season, multiple wheat genotypes, and contrasting nitrogen fertilization treatments distributed across the experimental field. These sources of temporal, structural, and management variability reduced the likelihood that model performance was driven solely by neighboring observations, although some influence of spatial autocorrelation could not be avoided within the relatively small experimental site. In addition, retaining the best-performing ANN initialization within each fold may have produced slightly optimistic cross-validation estimates. Consequently, the cross-validation results are interpreted primarily as a comparative assessment of feature combinations rather than as fully independent estimates of model generalizability.

The influence of horizontal grid resolution (GSD) and vertical layer segmentation on GF performance was systematically evaluated (Fig. 7). The tested parameter ranges were selected based on winter wheat crop height (∼1 m) and the point density achievable with the UAV LiDAR system, balancing canopy representation with sufficient point counts within each voxel.

Overall, the configuration consisting of five 20 cm vertical layers and a 30 cm GSD produced the best performance (training RMSE = 2.04 t ha⁻¹, R2=0.81; testing RMSE = 2.04 t ha⁻¹, R2=0.83). Ground-only GF and coarser vertical segmentations produced moderate performance but provided limited discrimination of dense canopy structure during peak growth. Conversely, increasingly fine segmentations (10×10 cm) substantially reduced performance (testing RMSE = 7.34 t ha⁻¹; R2=0.12), likely reflecting insufficient point density within individual layers and increased model complexity.

Following the identification of the optimal 5 layer GF parameterization, additional LiDAR-derived features were incorporated to evaluate whether complementary structural and intensity information could further improve AGB estimation. Crop height (CH) and normalized LiDAR intensity (INT) were assessed individually and in combination with the 5 layer GF model using both the 70/30 holdout evaluation and the repeated stratified 10-fold cross-validation framework (Fig. 9). Detailed 70/30 training and testing scatter plots are provided in Fig. B7.

Within the 70/30 holdout evaluation, CH alone provided strong predictive performance (RMSE = 1.99 t ha⁻¹; R2=0.83), substantially outperforming INT alone (RMSE = 2.56 t ha⁻¹; R2=0.73). Combining CH and INT resulted in only a marginal improvement, suggesting that crop height already captures much of the structural variability related to biomass. In contrast, incorporating either CH or INT with the 5 layer GF representation consistently improved prediction accuracy. The combination of GF + CH achieved RMSE = 1.99 t ha⁻¹ (R2=0.84), while adding INT further reduced the error to RMSE = 1.73 t ha⁻¹ with R2=0.87, representing the best-performing LiDAR-only model under the holdout evaluation.

The repeated stratified 10-fold cross-validation produced a slightly different ranking of feature combinations. Under this evaluation framework, the combination of 5 layer GF and normalized INT achieved the best overall performance (mean RMSE = 1.23 t ha⁻¹; mean R2=0.95), followed closely by the integration of CH, GF, and INT (mean RMSE = 1.41 t ha⁻¹; mean R2=0.93). In contrast, GF + INT showed comparatively weaker performance under the 70/30 holdout evaluation. This difference may partly reflect the larger amount of training data available within each cross-validation fold, allowing the ANN to better exploit the complementary information contained within canopy density and intensity features. However, because the cross-validation folds were stratified by measurement date rather than spatially blocked, some influence of spatial autocorrelation between training and validation subsets may also have contributed to the higher cross-validation performance. Additional analyses showed negligible relationships between scan angle, point density, and the features used in this study (Fig. B4; R2<0.04).

Multispectral (MS) reflectance was subsequently evaluated both independently and in combination with the LiDAR-derived features to assess whether spectral information provided complementary value for AGB estimation (Fig. 10). The MS model consisted of the red (R), red-edge (RE), and near-infrared (NIR) reflectance bands and was compared with combinations of crop height (CH), 5 layer gap fraction (GF), and normalized LiDAR intensity (INT).

Within the 70/30 holdout evaluation, MS reflectance alone produced strong predictive performance (RMSE ≈2.0 t ha⁻¹; R2≈0.83), comparable to several individual LiDAR-derived models. Integrating MS with LiDAR consistently improved biomass estimation, with the largest improvement observed for the combination of MS and normalized LiDAR intensity (RMSE ≈1.5 t ha⁻¹; R2≈0.91). Similar improvements were observed for MS combined with CH or 5 layer GF, demonstrating that passive spectral reflectance provides complementary information to LiDAR-derived canopy descriptors.

The repeated stratified 10-fold cross-validation showed a similar overall pattern while further reducing the differences among the evaluated feature combinations. The combination of MS and 5 layer GF achieved the lowest mean RMSE (1.06 t ha⁻¹) and a mean R2 of 0.92, followed closely by MS + CH + 5 layer GF (RMSE = 1.09 t ha⁻¹; R2=0.94) and MS + CH (RMSE = 1.21 t ha⁻¹; R2=0.94). Although these combinations differed slightly in ranking, all LiDAR–MS models exhibited overlapping performance variability and substantially outperformed MS alone (RMSE = 1.62 t ha⁻¹; R2=0.89). The relatively small differences among the combined models suggest that the primary improvement arises from integrating complementary structural and spectral information, whereas the addition of further LiDAR-derived features provides comparatively modest gains.

Overall, these findings indicate that multispectral reflectance provides complementary information for AGB estimation, with the largest improvement achieved through its initial integration with LiDAR-derived structural features. Once complementary structural and spectral information is incorporated, additional LiDAR features provide only relatively modest improvements, suggesting diminishing returns as feature dimensionality increases under the conditions of this study.

Thermal infrared (TIR) information was subsequently evaluated independently and in combination with the LiDAR and multispectral (MS) features to assess whether canopy thermal status provided additional information for AGB estimation (Fig. 11). Both land surface temperature (LST) and the normalized crop water stress index (CWSI) were investigated individually and as complementary inputs within the sensor fusion framework.

Using TIR alone produced the weakest biomass predictions of all evaluated sensor types, with LST (RMSE = 4.27 t ha⁻¹; R2=0.25) and CWSI (RMSE = 4.26 t ha⁻¹; R2=0.26) exhibiting nearly identical performance. Because normalization to CWSI provided little additional predictive value, subsequent sensor fusion analyses focused on LST, which offers a simpler and more operationally practical metric by avoiding the need for air temperature measurements and wet/dry reference (“anchor”) pixels.

Integrating LST with LiDAR-derived features substantially improved biomass estimation. The combination of LST with 5 layer GF and normalized intensity reduced the testing RMSE to approximately 1.5 t ha⁻¹ while increasing R2 to approximately 0.90, demonstrating that thermal observations provide complementary information when combined with structural canopy descriptors. Additional incorporation of multispectral reflectance or further LiDAR features produced only modest improvements, suggesting diminishing returns as feature dimensionality increased.

The repeated stratified 10-fold cross-validation showed a similar pattern, with all TIR-containing fusion models achieving consistently high predictive performance and relatively small differences among feature combinations. The combination of LST + CH + 5 layer GF produced the lowest mean RMSE (1.03 t ha⁻¹) and highest mean R2 (0.95), followed closely by LST + 5 layer GF (RMSE = 1.09 t ha⁻¹; R2=0.95) and the comparatively simple LST + MS model (RMSE = 1.15 t ha⁻¹; R2=0.95). The full multi-sensor combination (LST + MS + CH + INT + 5 layer GF) achieved a mean RMSE of 1.11 t ha⁻¹ and R2 of 0.92, indicating that once complementary thermal, structural, and spectral information is incorporated, additional sensor features provide only modest improvements.

Overall, these results suggest that TIR radiance serves as a complementary rather than primary predictor of AGB. While LST alone exhibits limited predictive capability, its integration with LiDAR-derived structural features and multispectral reflectance improves biomass estimation, and the minimal difference between LST and CWSI indicates that directly incorporating LST may provide a simpler and equally effective alternative for ANN-based sensor fusion under the conditions of this study.

To provide an overall comparison of the evaluated feature combinations, the top-performing single-sensor and multi-sensor models from Sect. 5.1–5.4 were ranked according to RMSE and R2 for both the 70/30 holdout evaluation and the repeated stratified 10-fold cross-validation (Fig. 12).

Across both evaluation strategies, LiDAR-derived features consistently dominated the single-sensor rankings. Models combining CH, 5 layer GF, and INT achieved the highest predictive performance, while MS reflectance also ranked among the strongest individual sensor approaches. In contrast, TIR features alone consistently produced the weakest biomass predictions, indicating that canopy temperature provides limited direct information on AGB when used independently.

The multi-sensor rankings further demonstrate the complementary nature of structural, spectral, and thermal information. Although the exact ordering of the top-performing models differed slightly between the 70/30 holdout and cross-validation evaluations, all leading feature combinations incorporated information from multiple sensor types and exhibited relatively similar performance. The 70/30 holdout evaluation favored combinations including MS, normalized LiDAR intensity, and LST, whereas the repeated cross-validation highlighted several LiDAR–MS feature combinations with comparable performance. Overall, LiDAR-derived information formed the foundation of every top-ranked model, with multispectral reflectance consistently providing complementary improvements and thermal observations contributing more modest gains.

To further investigate temporal sensor behavior, the best-performing single-sensor and multi-sensor feature combinations were evaluated separately for the early (19 April–31 May), mid (14 June–5 July), and late (27 July–5 August) portions of the growing season (Fig. 13). These periods correspond to distinct phases of canopy development observed in Fig. 2, including rapid vegetative growth, the divergence of LAI and GLAI during reproductive development, and complete senescence where GLAI approached zero. Additional growth-stage comparisons for all evaluated feature combinations and absolute error distributions are provided in Fig. B10.

Among the single-sensor models, LiDAR-derived structural features generally exhibited the lowest prediction errors during the early portion of the growing season, with multi-layer GF and its combinations achieving RMSE values below 0.5 t ha⁻¹. As crop development progressed and crop height approached a plateau while biomass continued to accumulate, prediction errors increased for all individual sensor approaches. The incorporation of normalized LiDAR intensity consistently improved the performance and stability of the multi-layer GF representation during the mid and late portions of the season.

Similar temporal trends were observed for the multi-sensor models. During the early portion of the season, combinations of LiDAR structural features and multispectral reflectance generally ranked among the best-performing configurations. During the mid and late portions of the season, feature combinations including multispectral reflectance together with crop height or multi-layer GF consistently exhibited low prediction errors, while combinations incorporating LST also became increasingly competitive. The larger error bars observed during the late season indicate greater variability among cross-validation folds, reflecting the increased variability present during the later stages of crop development.

To further investigate the biophysical information represented by the UAV-derived features, their relationships with destructively measured canopy properties were evaluated across the major growth phases (Fig. 14). Rather than training additional ANN models, feature values extracted from the destructive sampling zones were directly compared with plant area index (PAI), leaf area index (LAI), and green fraction of LAI (fg = GLAI / LAI) using the coefficient of determination (R2). This analysis provides insight into whether individual features primarily respond to total canopy structure or to the green and photosynthetically active component of the canopy.

During the jointing stage, when the canopy consisted almost entirely of green vegetation (GLAI ≈ LAI and fg≈1), the structural LiDAR metrics exhibited the strongest relationships with both PAI and LAI (Fig. 14). Crop height (CH), VDI, and particularly HWPI showed high correspondence with the destructive measurements (R2≈0.82–0.91), whereas the conventional ground-based GF performed substantially worse. These results highlight the advantage of preserving vertical canopy information during early crop development. During the flowering–ripening stage, the relationships changed markedly as LAI diverged from GLAI while PAI remained relatively stable (Fig. 2). Purely structural LiDAR metrics exhibited little correspondence with LAI or fg, whereas normalized LiDAR intensity showed the strongest relationship with fg and multispectral NDVI displayed a similar response. This suggests that the 903 nm LiDAR return captures complementary information related to canopy greenness and physiological condition beyond geometric structure alone. During the ripening–senescence stage, when GLAI approached zero and fg was no longer meaningful, structural LiDAR metrics again exhibited modest relationships with PAI, particularly HWPI, indicating that changes in canopy architecture remained detectable even as green vegetation declined. Spectral relationships were generally weaker during this period.

Overall, the correspondence between UAV-derived features and destructive canopy measurements varied with crop development. Structural LiDAR metrics were most informative during jointing and ripening, whereas normalized LiDAR intensity and multispectral NDVI exhibited the strongest relationships with green canopy fraction during flowering–ripening. These complementary temporal responses help explain the sensor fusion results, where structural features dominated biomass estimation early in the season while intensity and multispectral observations provided the greatest additional value as the canopy matured and senesced.

Among the evaluated UAV features, LiDAR-derived structural metrics consistently ranked among the strongest predictors of AGB throughout the growing season. In particular, representing canopy density using multiple vertical GF layers substantially improved biomass estimation compared with the conventional ground-return GF approach commonly used for LAI estimation (Bates et al., 2021; Dreier et al., 2024) and more recently for AGB estimation (Montzka et al., 2023; Hütt et al., 2022). These findings demonstrate that preserving within-canopy structural information provides a more complete representation of crop architecture than a single integrated canopy density metric.

The parameter sensitivity analysis further showed that increasing structural detail improves biomass estimation only while sufficient LiDAR sampling density is maintained. Under the acquisition conditions of this study, the five-layer (5×20 cm) GF representation combined with a 30 cm ground sampling distance provided the best balance between canopy characterization and point density, whereas finer 10×10 cm discretization likely contained insufficient returns within individual layers for robust density estimation. Rather than representing a universally optimal configuration, the appropriate layer depth and number should be selected according to sensor performance, flight parameters, resulting point density, and crop canopy structure.

The complementary nature of the LiDAR features was further demonstrated when CH, multi-layer GF, and normalized INT were evaluated together. CH describes overall canopy development, while multi-layer GF characterizes the vertical distribution of canopy density and remained sensitive to within-canopy structural variation after crop height had largely stabilized. Temporal comparisons with destructive measurements further showed that structural LiDAR features exhibited the strongest relationships with PAI and LAI during early development, providing a physical explanation for their consistently strong biomass prediction performance.

Overall, these results highlight the importance of feature design for UAV-based biomass estimation. While most UAV LiDAR biomass studies have relied primarily on crop height as the dominant structural predictor (Bendig et al., 2014; Madec et al., 2017; Pan et al., 2022), our results demonstrate that incorporating vertically resolved canopy density and normalized LiDAR intensity provides complementary information that better captures canopy heterogeneity and seasonal dynamics. Rather than relying solely on increasingly complex machine learning architectures, improving the physical representation of canopy structure through optimized LiDAR features produced substantial gains in predictive performance while maintaining direct links to measurable crop properties.

Compared with crop height and canopy density metrics, UAV LiDAR intensity has received relatively limited attention for crop biomass estimation despite being acquired simultaneously with the three-dimensional point cloud. Previous studies have reported relationships between LiDAR intensity and canopy properties including LAI (Luo et al., 2018), green area index (Liu et al., 2017), nitrogen treatments (Hütt et al., 2022), and AGB (Montzka et al., 2023). Consistent with these findings, normalized LiDAR intensity in the present study consistently improved biomass estimation when combined with crop height and multi-layer gap fraction. Because intensity is inherently recorded during LiDAR acquisition and the simple ground-based normalization adopted here improved temporal consistency without requiring full radiometric calibration, it represents a practical complementary feature that can be extracted without additional sensors or acquisition costs.

The temporal analyses provide further insight into this complementary behavior. During the flowering–ripening period, when crop height had largely stabilized and PAI diverged from LAI and GLAI, normalized LiDAR intensity maintained stronger relationships with LAI and, particularly, the green fraction of LAI than the purely structural LiDAR metrics. This complementary response may reflect the influence of both canopy density (Hütt et al., 2022) and leaf orientation and surface optical properties on the 903 nm return signal (Tian et al., 2021), helping explain why intensity was retained in many of the highest-performing LiDAR-only and multi-sensor feature combinations.

Multi-sensor fusion consistently improved AGB estimation compared with individual sensor approaches, supporting previous studies that demonstrate the value of combining complementary structural, spectral, and thermal information for crop monitoring (Li et al., 2024; Maimaitijiang et al., 2017; Wang et al., 2024). However, the magnitude of these improvements depended on both the selected feature combination and crop phenology.

Temporal analyses showed that LiDAR-derived structural features provided the strongest predictions during early vegetative development, whereas multispectral reflectance and normalized LiDAR intensity became increasingly valuable during the flowering–ripening period when crop height had largely stabilized and PAI diverged from LAI and GLAI. Thermal information produced comparatively modest improvements but consistently enhanced several of the highest-performing fusion models, indicating that physiological information provides complementary value when combined with structural and spectral observations.

Despite these benefits, many of the highest-performing fusion models exhibited very similar predictive performance despite differing substantially in sensor complexity. These results suggest diminishing returns as additional sensors or derived features are incorporated and emphasize the flexibility of a feature-based fusion framework. Rather than identifying a universally optimal sensor configuration, the findings indicate that sensor selection should be guided by the objectives and logistical constraints of the monitoring campaign, balancing prediction accuracy against acquisition cost, payload requirements, processing complexity, and operational practicality.

The results demonstrate the practical potential of UAV LiDAR for high-resolution biomass monitoring. A single LiDAR acquisition simultaneously provides complementary information describing overall canopy development (CH), the vertical distribution of canopy density (multi-layer GF), and normalized return intensity (INT), producing some of the highest-performing biomass models while requiring only a single sensor and acquisition workflow.

Figure 15 further illustrates this advantage by comparing UAV-derived AGB, destructive biomass measurements, and grain yield across nitrogen fertilization treatments throughout the growing season. While both UAV and destructive measurements captured similar temporal dynamics and treatment responses, the UAV estimates exhibited smoother transitions and lower variability because they represent entire treatment plots rather than a limited number of sampling quadrats. This spatial completeness provides a more representative characterization of treatment effects and within-field variability while reducing the influence of localized sampling heterogeneity; spatial AGB maps across all UAV acquisition dates are provided in Fig. B11.

From an operational perspective, the highest-performing multi-sensor models differed only marginally in prediction accuracy, suggesting that sensor selection should be guided by the broader objectives of the monitoring campaign rather than biomass estimation alone. For AGB-focused applications, LiDAR-derived structural features and normalized intensity may provide sufficient information, whereas multispectral or thermal sensors can be incorporated when additional products such as nutrient status, pigment content, evapotranspiration, or water stress are also required. Rather than replacing other sensing technologies, the results suggest that LiDAR forms a strong foundation onto which complementary sensors can be added according to the agronomic objectives and logistical constraints of individual monitoring campaigns.

The primary objective of this study was to systematically compare UAV-derived feature representations within a consistent ANN framework rather than to identify a universally optimal biomass prediction model. Consequently, the findings should be interpreted within the context of the winter wheat canopy structure, management practices, sensor configuration, and environmental conditions investigated here. Future studies should evaluate these feature combinations across additional crops, sites, growing seasons, and machine learning frameworks, as sensor performance is likely to vary with canopy architecture (e.g., grasses vs. broadleaf crops) and modelling approach (Han et al., 2019; Sharma et al., 2022; Bazrafkan et al., 2023).

The validation framework also has limitations. Although the dataset incorporated substantial temporal and management variability, the relatively small experimental field meant that some influence of spatial autocorrelation between training and validation samples could not be entirely avoided. Because the primary objective of this study was to compare the relative performance of individual sensors and their fusion rather than develop a universally transferable biomass model, these results provide a valuable starting point for systematic sensor comparison. Future studies should further evaluate the proposed sensor metrics using spatially blocked cross-validation and independent sites or growing seasons to better assess model transferability and feature generalizability.

The results also highlight opportunities for further LiDAR feature refinement. While the 5 layer GF representation consistently outperformed conventional single-layer approaches, future studies should investigate adaptive layer parameterization based on canopy structure, sensor characteristics, and acquisition settings rather than fixed layer depths. Similarly, radiometrically calibrated LiDAR intensity, which was not evaluated here, may further improve temporal consistency and strengthen the physical interpretation of intensity measurements. The weak relationships observed between scan angle, point density, GF, and AGB (see Figs. B3 and B4) suggest that the low-altitude, ±25° acquisition strategy effectively minimized these influences, although they may become increasingly important under wider scan angles, lower point densities, or more complex terrain.

Finally, extending vertically segmented features beyond canopy density represents a promising research direction. Multi-layer LiDAR intensity profiles could help separate structural and radiometric effects within the canopy, improving the estimation of green canopy fraction and providing complementary information for biomass estimation and other applications such as evapotranspiration modelling. Future multispectral or multi-wavelength LiDAR systems (Takhtkeshha et al., 2024) may further capitalize on these concepts by enabling vertically resolved vegetation indices that characterize within-canopy surface conditions in addition to structure. Combined with appropriate radiometric calibration and methods to disentangle the effects of canopy density, leaf angle, and viewing geometry, such systems could provide new physically meaningful features for precision agriculture and crop monitoring.