LUF-net: A physically informed color calibration method for UAV RGB images based on exposure and irradiance information Authors: Chufeng Wang, Bei Liu, Jingyang Zhang, Shichao Jin, Fengjie Yuan Published: 2026 ABSTRACT AbstractAccurate color representation is critical for UAV-based crop phenotyping, yet UAV images are often distorted by variable illumination and camera exposure settings. Here, we propose Lite U-net FiLM (LUF-net), a lightweight U-net framework integrated with feature-wise linear modulation (FiLM) layers, which incorporates both multispectral-derived irradiance and camera exposure parameters as auxiliary inputs. This metadata-aware design enables dynamic modulation of intermediate image features, allowing the network to disentangle unified canopy color from environmental artifacts. Experiments conducted across diverse crop types (soybean, rice, maize), flight altitudes (6m, 12m, 25m, 40m), and illumination conditions (overcast skies, cloudy, sunny, early morning) demonstrate that the LUF-net model substantially reduces the mean absolute percentage error to below 5.5%, outperforming conventional Gray-world, U-net, AlexNet-MLP, and SIDBlock-MLP methods. Ablation experiments further show that both irradiance and exposure metadata provide complementary information, while FiLM-based conditioning effectively integrates these acquisition parameters into feature learning, jointly contributing to improved color reconstruction performance. Moreover, correlation analysis indicates that the model's color reconstruction errors are weakly dependent on irradiance and exposure settings, confirming that LUF-net reduces sensitivity to external imaging conditions while maintaining physically meaningful and robust corrections.Keywords: Unmanned aerial vehicle, Color calibration, LUF-Net, Irradiance, Exposure parameters Introduction Unmanned aerial vehicles (UAVs) equipped with consumer-grade cameras and multispectral sensors are increasingly deployed in precision agriculture [[1], [2], [3], [4], [5]]. Compared with traditional ground-based surveys [6,7] and satellite remote sensing [[8], [9], [10], [11]], UAV-based platforms offer unique advantages in terms of temporal flexibility, centimeter-level spatial resolution, and suitability for monitoring critical indicators for crop growth assessment [[12], [13], [14]], stress detection [[15], [16], [17]], and yield prediction [[18], [19], [20]]. Increasingly, canopy color information has been recognized as a pivotal feature in these estimations, offering a perceptually intuitive yet quantitatively rich descriptor of plant physiological and biochemical dynamics [[21], [22], [23], [24]]. Accurate and consistent color information is therefore a fundamental prerequisite for reliable phenotypic trait extraction, as even small deviations in color representation can propagate errors in vegetation indices, biophysical parameter retrieval, and subsequent agronomic decision-making. Despite the rapid progress in UAV-based crop phenotyping, one persistent challenge is the inconsistency of color representation under highly dynamic field conditions. The color information recorded by digital cameras is influenced by multiple external and internal factors, including solar elevation angle, cloud-induced irradiance fluctuations, as well as sensor-related adjustments such as ISO, shutter speed, and aperture settings [[25], [26], [27]]. These influences introduce nonlinear distortions in image color values, leading to discrepancies between images acquired at different times, altitudes, or weather scenarios [28]. Such inconsistencies hinder temporal comparability, compromise vegetation index stability, and reduce the transferability of phenotyping models across sites and seasons. Without appropriate correction, UAV imagery may fail to reflect actual canopy conditions, thereby weakening its value for precision agricultural monitoring and large-scale crop modeling [29]. To mitigate these challenges, various color calibration strategies have been developed. Traditional methods often rely on reference color targets or regression-based transformations to standardize UAV imagery across acquisition conditions [[30], [31], [32]]. While effective under controlled experiments, these approaches present practical limitations in field-scale applications. Furthermore, many of these methods implicitly assume stable illumination conditions, an assumption rarely satisfied in real-world agricultural environments where clouds, shadows, and solar angle variations frequently occur. In addition to physical calibration targets, several statistical color constancy algorithms—such as the Gray-world and white-patch assumptions—have been applied to UAV imagery to achieve illumination-invariant color balance [[33], [34], [35]]. However, these methods rely on simplified assumptions about scene reflectance and lighting uniformity, which often fail under heterogeneous field conditions. More recently, deep learning has emerged as a promising alternative, capable of learning complex nonlinear transformations between distorted and reference images [[36], [37], [38], [39], [40], [41]]. Image-to-image regression frameworks, particularly U-net variants, have demonstrated strong performance in pixel-wise color correction without requiring explicit use of calibration charts [42]. These data-driven approaches offer automation and scalability, representing a significant step toward operational UAV-based phenotyping pipelines. However, most existing models rely exclusively on image content for correction and largely ignore auxiliary metadata that directly govern image appearance. Camera exposure parameters (e.g., ISO, shutter speed, aperture) and irradiance information strongly influence pixel brightness, contrast, and spectral fidelity, yet they are rarely integrated into learning frameworks [43,44]. Neglecting these environmental and imaging factors reduces robustness under variable conditions and limits the applicability of existing models in multi-temporal, cross-site UAV monitoring. To address the above limitations, this study aims to establish an irradiance- and exposure-aware deep learning framework, termed Lite U-net feature-wise linear modulation (LUF-net), for UAV image color calibration. The proposed framework integrates exposure parameters and multispectral irradiance values as auxiliary inputs, which are embedded into the feature extraction process through FiLM layers. Furthermore, this study will systematically evaluate the proposed LUF-net model against both traditional color constancy methods (e.g., Gray-world) and learning-based baselines, focusing on calibration accuracy, computational efficiency, and robustness. Additional ablation experiments will be conducted to investigate the individual and combined contributions of different metadata sources and to assess the effectiveness of FiLM-based feature modulation in integrating acquisition information into feature learning. The analysis will also examine the model's sensitivity to variations in exposure and irradiance and validate its applicability across different crop types, flight altitudes, and illumination conditions. Through these investigations, the study seeks to establish a lightweight and physically informed calibration paradigm that supports consistent and reliable color correction for UAV-based phenotyping and precision agriculture. Materials and methods Experimental design The study was conducted at field experimental stations located in Hangzhou and Jiaxing, Zhejiang Province, China (Fig. 1a). Although the fields are primarily used for breeding and cultivating summer crops such as maize (Fig. 1b), soybean (Fig. 1c), and rice (Fig. 1d), the present study focused on evaluating color correction methodologies rather than crop performance. Experiments were conducted under typical field conditions to collect images under diverse illumination and background scenarios, enabling robust evaluation of the proposed color correction approach across representative lighting and surface reflectance conditions. Study area and field previews. Spatial distribution of the study area (a); representative maize field (b); representative soybean field (c); representative rice field (d). UAV data acquisition and pre-processing A DJI M3M UAV (SZ DJI Technology Co., Shenzhen, China) served as the imaging platform, capable of simultaneously acquiring high-resolution RGB images (5472 × 3648 pixels) and four multispectral bands. As a widely adopted consumer-grade UAV platform in agricultural monitoring and crop phenotyping applications, the DJI M3M integrates a downwelling light sensor that provides synchronized irradiance measurements during image acquisition. In this study, these irradiance measurements were used as auxiliary metadata inputs to support illumination-aware RGB color calibration within the proposed LUF-net framework. The multispectral configuration corresponded to green (G, 560 ± 16 nm), red (R, 650 ± 16 nm), red edge (RE, 730 ± 16 nm), and near-infrared (NIR, 860 ± 26 nm) wavelengths, thereby covering both visible and near-infrared regions. Image acquisition was conducted under overcast, cloudy, sunny, and early-morning lighting conditions. (Fig. 2). Instead of pre-defined flight routes, flights were performed at multiple altitudes (6, 12, 25, and 40 m above ground level) to capture scale-dependent image characteristics. Because all UAV acquisitions were conducted at relatively low altitudes, atmospheric absorption and scattering effects were considered negligible compared with those encountered in conventional airborne or satellite remote sensing observations. Previous studies have shown that the extremely short atmospheric optical path in low-altitude UAV remote sensing substantially reduces atmospheric interference and often eliminates the need for explicit atmospheric correction [45,46]. All images were obtained using automatic exposure settings with manual triggering to ensure adaptability under changing illumination environments, and the associated camera parameters, including ISO, aperture, and shutter speed, were stored in the metadata of each image. RGB images were stored in both DNG and 8-bit JPEG formats, while multispectral data were saved in 16-bit TIFF formats, ensuring high spatial and spectral fidelity. UAV imagery at different flight altitudes and over different crop types. The proposed method was developed using individual UAV images rather than mosaicked orthophotos. A total of 230 original UAV images were acquired. To ensure data quality, a small number of images affected by motion blur or acquisition artifacts were removed. The remaining dataset maintained a relatively balanced distribution across crop species, illumination conditions, and flight altitudes. Each original image was cropped into non-overlapping patches of 256 × 256 pixels using a sliding-window approach with a stride of 256 pixels, resulting in a total of 58,292 image patches. Associated metadata, including camera parameters (aperture, exposure time, and ISO) and multispectral irradiance measurements, were retained as auxiliary inputs for model training. To avoid potential data leakage, dataset partitioning was performed at the original image level rather than directly at the patch level. Specifically, original images were first assigned to training, validation, and testing subsets, after which image patches were generated independently within each subset. This strategy ensured that adjacent samples originating from the same flight mission, experimental plot, or acquisition period did not appear simultaneously across different subsets, thereby improving the reliability of model evaluation. Ground truth generation It should be noted that the color correction method was designed to recover colors toward a unified absolute reference. The ground-truth labels were generated using the standard color values of a 24-color calibration chart (Fig. 3), ensuring that all corrected images were mapped into a consistent color space. The true color values of each patch, denoted as T, were measured using a colorimeter (WR-10QC, Wave Optoelectronics Technology Co., Ltd., Shenzhen, China), while the observed colors extracted from UAV images were recorded as O. A multiple linear regression was applied to model the relationship between the observed and true colors, expressed as where H represents the calibration coefficient matrix and δ is the residual error. The optimal matrix H was obtained by minimizing the sum of squared residuals, using the equation: Workflow schematic of UAV imagery color calibration, comparing different correction methods for improved color consistency. After estimating H, the estimated matrix is then applied to each pixel I in the input image: where G is the corrected RGB image. All color values were represented in the CIE Lab color space, which provides a perceptually uniform structure with three axes: L∗ for lightness, a∗ from green to red, and b∗ from blue to yellow. The Lab space was chosen over device-dependent spaces such as RGB because it better aligns with human visual perception. Multiple linear regression was selected due to its ability to account for systematic biases and scaling introduced by the imaging sensor and varying illumination. This combined Lab and regression framework ensures that the corrected UAV imagery accurately reflects true colors, enabling subsequent analyses such as color-based phenotyping. Calibration methods for color Gray-world method The Gray-world assumption was applied as a baseline color correction method, which assumes that the average reflectance in a scene is achromatic. For an input image I, the mean of each RGB channel was computed, and a scaling factor was applied to match the global gray level, yielding the corrected image Ic: where μR,G,B is the mean of each RGB channel and μgray is their overall average. To align with the reference 24-patch color chart, the Gray-world corrected image was linearly stretched to the chart's dynamic range. Corrected images were then evaluated against target images using MSE, MAE, and MAPE, providing a quantitative baseline for comparison under varying illumination conditions. AlexNet-MLP method To perform color correction, a deep learning framework combining CNN feature extraction and a fully connected multilayer perceptron (MLP) was employed. RGB images were first converted to the CIE Lab color space. Simultaneously, metadata vectors containing camera parameters (ISO, exposure, aperture) and spectral band information (irradiance of band R, G, RE, and NIR) were incorporated as additional inputs. Pretrained AlexNet was used to extract deep features F from the Lab images [47]: Extracted features were concatenated with the corresponding metadata vectors V to form the final input X=[F,V]. The MLP then predicted the corrected Lab image Ypred: where θ represents the learnable parameters of the network. The model was trained to minimize the mean squared error (MSE) between the predicted Ypred and target Lab images Ytrue: SIDBlock-MLP method A deep learning model that combines a convolutional encoder–decoder (SIDBlock) with a MLP to perform metadata-guided color correction. The encoder consists of three convolutional layers with progressively increasing channels (32, 64, 128) and down-sampling operations, while the decoder employs two transposed convolutions to restore spatial resolution, followed by a convolutional layer to reconstruct the corrected Lab image. The metadata branch uses an MLP with two fully connected layers to project the 7-dimensional metadata vector into a 3-dimensional embedding, which is spatially expanded and fused with image features through element-wise addition. RGB inputs were first converted into the CIE Lab color space and processed by the SIDBlock encoder to extract image features F: The metadata vectors V (aperture, exposure time, ISO, G, R, NIR, RE) were projected by an MLP into embedding E: The multimodal representation X was obtained by fusing F and E: where α is a scaling coefficient. The decoder then generated the predicted Lab image Yˆ: LUF-net method As illustrated in Appendix 1, the proposed LUF-net adopts a metadata-guided lightweight encoder–decoder architecture for RGB color correction in Lab color space. The framework consists of two main branches: an image feature extraction branch based on a Lite U-net and a metadata modulation branch based on FiLM conditioning. The input RGB image is first transformed into Lab color space and then fed into the encoder. The encoder contains three convolutional stages with feature dimensions of 32, 64, and 128 channels, respectively. Each stage consists of a convolution block (3 × 3 convolution + batch normalization + ReLU activation) followed by 2 × 2 max-pooling for spatial downsampling. A 256-channel bottleneck layer is subsequently used to extract high-level semantic representations. To incorporate illumination and camera acquisition information, the 7-dimensional metadata vector is simultaneously introduced into the network. The metadata vector is processed by a MLP and injected into the encoder through FiLM layers after each encoder block. These FiLM layers perform channel-wise affine modulation on intermediate feature maps, enabling adaptive feature adjustment under varying illumination conditions. Specifically, for an input feature map F and metadata vector V, the FiLM operation computes: where γ and β are learnable linear transformations. This allows the network to adapt image features according to camera settings and spectral information. The performance evaluation of the proposed method All models were implemented in Python 3.10.18 using PyTorch as the deep learning framework. The experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 2080 Ti GPU (22 GB VRAM). To evaluate the performance of the proposed learning-based color correction method, the dataset was randomly divided at the image level into training, validation, and test subsets with proportions of 60%, 20%, and 20%, respectively. The splitting was performed using randomized index sampling without overlap among subsets to ensure that test samples remained completely unseen during model training and hyperparameter tuning. For computational efficiency and balanced evaluation, a maximum of 50,000 image pairs were randomly sampled when the dataset size exceeded this limit. During inference, only the independent test subset was used for quantitative evaluation. The AdamW optimizer with a learning rate of 1 × 10−3 and weight decay of 1 × 10−4 was used. A ReduceLROnPlateau scheduler adjusted the learning rate according to validation loss. Batch size was set to 128, and training was conducted for 20 epochs. The corrected images were quantitatively assessed using the color error in the CIE Lab color space, defined as where L∗, a∗, b∗ represent the predicted and reference values of the Lab channels. In addition, relative errors such as mean absolute percentage error (MAPE) were calculated to evaluate the deviation of the predicted values from the ground truth: To quantitatively validate the effectiveness of each core component of the proposed LUF-net, comprehensive ablation experiments were designed and implemented. We set four comparative variant models: the complete LUF-net, LUF-net with only camera exposure setting embedded, LUF-net driven merely by irradiance information, and LUF-net without the FiLM feature modulation module. In addition, a classic image-only U-Net was adopted as the baseline for further performance comparison. All variant models were trained and tested under identical experimental configurations, evaluation metrics were unified, and quantitative error indicators were recorded to verify the contribution of metadata input and FiLM modulation to color calibration performance. Results Ground truth accuracy Ground-truth color calibration accuracy was assessed using 24 color patches. The mean R2 between estimated and measured values was 0.99 for L∗, 0.98 for a∗, and 0.99 for b∗ (Fig. 4a), demonstrating excellent consistency across all three color channels. In Fig. 4 b, the mean color errors were ΔL = 1.43, Δa = 2.63, and Δb = 1.95. Among them, the L∗ channel exhibited the smallest deviation, while the a∗ channel showed the largest, though still within an acceptable range. These high R2 values combined with the relatively low color errors confirm that the generated ground truth is highly reliable and can be confidently used for subsequent calibration and analysis. Ground truth accuracy of color calibration. (a) R2 between estimated and measured values of L∗/a∗/b∗ across 24 color patches; (b) Color errors between estimated and measured L∗/a∗/b∗ values across 24 color patches. Calibration accuracy without metadata vector Table 1 presents the accuracy and efficiency of color calibration by the Gray-world method and the U-net model without incorporating metadata vectors. The U-net substantially reduced the mean color error from 26.1 (Gray-world) to 16.7, and the MAPE from 14.6% to 9.31%, indicating a much higher calibration accuracy. In terms of computational efficiency, U-net also showed a short inference time with 1.18 s per hundred patches. Table 1Accuracy of color calibration by different methods.Table 1MetricGray-worldU-netAlexNet-MLPSIDBlock-MLPLUF-netColor error26.116.714.113.66.96MAPE14.6%9.31%7.89%7.6%3.89%Time (s/hundred patches)/1.181.01.171.27 Accuracy of color calibration by different methods. Calibration accuracy by different methods with metadata vector When metadata vectors were incorporated, the performance of different color calibration methods was evaluated, and the results are summarized in Table 1. Compared with AlexNet-MLP (color error = 14.1, MAPE = 7.89%) and SIDBlock-MLP (color error = 13.6, MAPE = 7.6%), the LUF-net model achieved the best accuracy, reducing the color error to 6.96 and the MAPE to 3.89%. Although the LUF-net required slightly longer inference time (1.27 s per hundred patches) relative to AlexNet-MLP (1.0 s) and SIDBlock-MLP (1.17 s), its significant advantage in calibration accuracy highlights the effectiveness of combining a lightweight U-net structure with feature-wise modulation for integrating metadata vectors. The proposed LUF-net contains 0.946 M trainable parameters and required 60.5 MB peak GPU memory during inference on an NVIDIA GeForce RTX 2080 Ti using 256 × 256 image patches (batch size = 1). During training, the model consistently minimized errors on both the training and validation sets, with a scheduled learning rate decay facilitating stable convergence and enhanced generalization (Fig. 5). Training and learning rate dynamics of the LUF-net model. Quantitative ablation results further intuitively validate the superiority of the proposed LUF-net (Fig. 6). As illustrated in the quantitative indicators, the complete LUF-net achieves theoptimal calibration performance with an error value of 6.96 and an error rate of 3.89%. In comparison, the variant models exhibit degraded performance, with error metrics rising to 8.86 (4.95%) for the exposure-only LUF-net, 7.34 (4.10%) for the irradiance-only LUF-net, and 8.09 (4.52%) for the model without FiLM modulation. Moreover, the image-only U-net presents a much higher error of 16.7 and an error rate of 9.31%, which fully demonstrates that the synergistic combination of exposure metadata, irradiance information, and FiLM feature modulation improves the color calibration accuracy and effectively reduces calibration errors. Ablation study of the proposed LUF-net under different configuration settings, evaluated using color error and MAPE metrics. Calibration accuracy across altitude, crop, and illumination conditions To further evaluate the robustness of the proposed model, its color calibration performance was examined under different crop types, flight altitudes, and illumination conditions, as summarized in Table 2. The model achieved consistently high accuracy across all scenarios. Among the three crops, soybean showed the lowest color error (5.37) and MAPE (2.77%), followed by rice (8.64; 5.26%) and maize (10.54; 5.07%). With respect to flight altitude, the optimal performance occurred at 25 m (color error = 5.39; MAPE = 2.72%), while slightly higher errors were observed at 6 m, 12 m, and 40 m. Under different illumination conditions, calibration accuracy remained high, with the best results under overcast skies (6.96; 3.53%) and early morning (7.23; 3.63%), and relatively higher errors under sunny conditions (9.39; 5.30%). Overall, the LUF-net model demonstrated robust and reliable color correction performance under diverse operational conditions. Table 2Accuracy of color calibration by LUF-net under different conditions.Table 2MetricCropmaizericesoybeanColor error10.548.645.37MAPE5.07%5.26%2.77%MetricFlight altitude6m12m25m40mColor error8.168.385.399.72MAPE4.33%4.41%2.72%4.98%MetricLight conditionovercast skiescloudysunnyearly morningColor error6.968.39.397.23MAPE3.53%4.19%5.3%3.63% Accuracy of color calibration by LUF-net under different conditions. Comparison of environmental and exposure sensitivity of color error between models The color reproduction performance of the standard U-net and the proposed LUF-net under varying imaging parameters is compared in Fig. 7. The scatter plots show the relationship between color error and acquisition parameters, including exposure time and irradiance in the G, R, NIR, and RE bands. Overall, the standard LUF-net exhibits weaker correlations between imaging parameters and color error than the U-net model. The Pearson correlation coefficients (r) demonstrate that, while the U-net shows noticeable correlations (e.g., r = 0.20 for exposure time; r = 0.27 and 0.28 for irradiance in the R and RE bands), these values are substantially reduced with LUF-net (r ≈ 0.07-0.08). Relationship between color error and imaging parameters for the LUF-net (a) and U-net (b) models. Discussion Effects of irradiance and exposure parameters on model performance UAV-based color information acquisition is inevitably affected by ambient illumination and camera exposure, which jointly modulate the recorded radiometric response of the crop canopy. Variations in solar irradiance, atmospheric transmittance, and exposure parameters such as shutter speed, ISO, and aperture often introduce inconsistent brightness and chromaticity, leading to color distortions that deviate from the crop's intrinsic reflectance characteristics ([11,43]; Zhang et al., 2020b). Recent efforts have explored hardware-dependent solutions, such as cooperative dual-UAV in-flight color correction, where one UAV continuously carries a ColorChecker to provide real-time reference targets for the imaging UAV [48]. Although this approach substantially improves color fidelity—achieving up to 66.1% reduction in color error and enhancing cross-sensor consistency—it requires synchronized flight control, additional equipment, and strict operational coordination, making it difficult to scale to routine phenotyping or large-area monitoring. The influence of acquisition conditions on color reconstruction extends beyond simple brightness variation, as identical canopy surfaces may exhibit substantially different image appearances under varying illumination and sensor settings despite unchanged intrinsic reflectance. Previous studies have shown that image-based deep learning frameworks such as U-net can achieve favorable color correction performance under relatively controlled conditions ([36,37]. However, relying solely on image appearance may lead models to learn acquisition-specific color distributions and become more sensitive to illumination-induced variations under field environments [[49], [50], [51]]. This suggests that image content alone may not sufficiently represent the physical factors governing color variability in UAV imagery. The ablation analysis further suggests that different metadata sources contribute complementary information regarding image formation. Exposure parameters and irradiance measurements do not represent redundant descriptors but instead provide distinct constraints on environmental and sensor-induced variability. Their combined use offers a more complete representation of acquisition conditions, potentially reducing the entanglement between illumination effects and intrinsic canopy properties. The reduced sensitivity of LUF-net to acquisition conditions shown in Fig. 7 further suggests that the metadata are effectively utilized during feature learning, enabling the model to capture more stable color representations that are less influenced by external imaging variability. Comparative and robustness analysis of the LUF-net model Quantitative results show that LUF-net consistently achieves lower color reconstruction errors than AlexNet-MLP and SIDBlock-MLP under diverse imaging conditions, indicating improved handling of acquisition variability in UAV color correction tasks. The performance gain is primarily attributed to the FiLM-based conditioning strategy. By modulating intermediate feature representations using irradiance and exposure vectors, FiLM enables dynamic adaptation of feature activations according to acquisition conditions [52]. This design integrates metadata into the transformation process rather than treating it as a simple auxiliary input [53]. In contrast, AlexNet-MLP and SIDBlock-MLP rely on image-driven feature extraction without explicit conditioning, resulting in weaker adaptability to acquisition changes [54,55]. Meanwhile, the lightweight architecture of LUF-net, containing fewer than 1 M trainable parameters and modest memory requirements, suggests that these performance improvements are achieved without substantially increasing model complexity. The ablation study further compares two fusion strategies using identical metadata inputs: direct concatenation and FiLM-based modulation. Results indicate that concatenation-based fusion provides limited improvement, whereas FiLM achieves more effective utilization of the same information by embedding it into feature-wise transformations. This demonstrates that performance gains are not solely due to the presence of metadata, but are strongly influenced by the way it is incorporated into the network. The robustness analysis shows that performance varies across environmental scenarios. Lower errors are observed under overcast conditions due to more uniform illumination, which reduces shadow and directional effects. In contrast, performance degradation under strong sunlight is mainly associated with specular reflections and non-Lambertian canopy effects, which cannot be fully described by global irradiance. Across different crop types and flight altitudes, the model maintains consistently stable performance, demonstrating strong robustness under diverse UAV imaging conditions. Contributions and perspectives The proposed LUF-net framework introduces a lightweight FiLM-based architecture that integrates acquisition metadata into feature learning for UAV-based color calibration, enabling explicit conditioning of visual representations on imaging conditions. A key contribution of this work lies in the design of a feature modulation mechanism that incorporates exposure and irradiance information into intermediate representations, allowing the network to account for acquisition-dependent variations in a structured manner. This provides a more direct pathway for embedding physical imaging cues into deep color correction models compared with purely image-driven approaches. Future research can further extend this framework toward more autonomous and scalable radiometric modeling. One direction is to infer illumination-related cues directly from image data, reducing reliance on explicit sensor measurements and enabling more flexible deployment in resource-constrained scenarios. Another promising extension is the adaptation of the current architecture to broader radiometric tasks, such as reflectance normalization, vegetation index stabilization, and spectral reconstruction. Beyond single-platform UAV applications, this modeling paradigm also opens opportunities for multi-source integration. Coupling physics-guided feature modulation with cross-platform sensing systems may facilitate more consistent agricultural monitoring across UAV observations under varying environmental conditions. Conclusion This study proposes LUF-net, a lightweight FiLM-conditioned U-Net for UAV agricultural image color calibration. The framework incorporates multispectral irradiance and camera exposure metadata into feature learning to address complex field interference. Experimental results verify that the proposed LUF-net delivers superior color calibration accuracy and environmental robustness over image-only U-Net and other baseline models under variable crop types, flight altitudes and illumination conditions. Such performance improvements are further explained by ablation analyses, which demonstrate that the synergistic integration of acquisition metadata and FiLM-based feature modulation yields better results than conventional feature concatenation, pure image-based learning, and separate modeling that only utilizes exposure or illumination information. This work demonstrates the efficacy of integrating acquisition-aware metadata into feature learning, providing a reliable solution for UAV image color calibration in complex agricultural environments. Future work will extend this framework for autonomous, scalable radiometric modeling via image-based illumination estimation and generalization to diverse agricultural radiometric tasks. Authorship contribution statement Chufeng Wang: conceptualization, software, writing; original draft, visualization, validation, writing; review & editing. Fengjie Yuan: conceptualization, methodology, data curation, writing; original draft, supervision, writing; review & editing. Shichao Jin: conceptualization, methodology, data curation, writing; original draft, supervision, writing; review & editing. Bei Liu and Jingyang Zhang: writing; review & editing. Data and code availability The data and source code for model training, evaluation, and inference, together with pretrained model weights, example data, and detailed usage instructions, is publicly available at: https://github.com/wangchufeng3652/color-correction. The repository also includes installation instructions, software dependencies, dataset organization, and example commands to facilitate reproduction of the reported results. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This research was supported by the Xianghu Laboratory Special Scientific Research Project (2023C2S02001). Contributor Information Shichao Jin, Email: jschaon@njau.edu.cn. Fengjie Yuan, Email: yuanfengjie@xhlab.ac.cn. model structure of LUF-net References Associated Data Data Availability Statement The data and source code for model training, evaluation, and inference, together with pretrained model weights, example data, and detailed usage instructions, is publicly available at: https://github.com/wangchufeng3652/color-correction. The repository also includes installation instructions, software dependencies, dataset organization, and example commands to facilitate reproduction of the reported results.