Field Crop 2025, Vol.8, No.4, 204-212 http://cropscipublisher.com/index.php/fc 207 al., 2020; Qiu et al., 2021). SVM is mostly used for classification, such as geophysical classification and growth stage division, and sometimes can achieve the highest accuracy (Ramadhani et al., 2020; Guo et al., 2021; Fatchurrachman et al., 2023). Extreme gradient boosting (XGBoost) and decision trees also perform well in predicting yield and biomass, especially when dealing with large-scale and multi-source data (Singha and Swain, 2023). Deep learning methods, such as convolutional Neural Networks (CNNS), EfficientNet, ResNet and YOLO, are more suitable for image-based tasks. They are often used to identify growth stages and detect rice panicles, with an accuracy rate of over 95%, and are more reliable than traditional models in complex field environments (Zheng et al., 2025). 4.2 Feature selection, model training, and validation approaches For a model to perform well, choosing the right input features is crucial. The commonly used features in the research include vegetation index, texture features, structural parameters (such as plant height, canopy coverage), and multi-temporal data (Sheng et al., 2022; Wang et al., 2023). The methods of feature selection include stepwise regression, random forest, and embedding methods in deep learning, all of which can help pick out the most useful variables (Ge et al., 2024). When training the model, it is usually necessary to divide the data into the training set and the validation set, and then use cross-validation and independent test sets to test the generalization ability of the model to avoid overfitting (Lyu et al., 2023). Common evaluation metrics include accuracy, precision, recall, F1 score, as well as R² and RMSE in the regression model (Sheng et al., 2022; Guo et al., 2023). 4.3 Challenges of model generalization and cross-regional transferability Although these models performed well in experiments, there are still problems when they are extended to different regions, varieties and environmental conditions. For example, sensor Angle, image resolution, plant overlap and field differences can all affect model stability (Zha et al., 2020; Tan et al., 2022; Tseng et al., 2022). In addition, different management methods, soil types and climatic conditions also limit the migration effect of the model. To address these issues, researchers have attempted to incorporate soil, weather and management information using methods such as transfer learning, multi-site training and data fusion (Iatrou et al., 2021). The latest research emphasizes that to enhance the generalization of models, more abundant multi-temporal data and standardized processes are needed, so that the models can be more stable when applied on a large scale. 5 Integration Framework of UAV Remote Sensing and Machine Learning 5.1 Data preprocessing and feature engineering (illumination correction, noise reduction, variable construction) The first step in integrating unmanned aerial vehicle (UAV) remote sensing with machine learning is to do a good job in data preprocessing and feature engineering. Common operations include: light correction (adapting to different lighting conditions), noise reduction (removing sensor and environmental interference), and geometric and radiative calibration (ensuring data consistency) (Zha et al., 2020; Shen et al., 2024). Afterwards, some variables need to be constructed, such as vegetation index, texture features and structural parameters (plant height, canopy coverage, etc.), which can help capture the relationship between the spectrum and the growth indicators of rice. Feature selection methods (such as recursive feature elimination or the built-in selection mechanism of the model) can identify the variable with the largest amount of information as input. 5.2 Multi-source data fusion methods for model input Combining drone images (multispectral or hyperspectral), satellite data, meteorological information and soil data can make the model more robust and also improve the prediction accuracy. The fusion methods include: aligning unmanned aerial vehicle (UAV) images with satellite images using scale transformation, or integrating spectral, texture and auxiliary data together to form a comprehensive feature set (Jin et al., 2024). The research found that after combining the data from unmanned aerial vehicles (UAVs) and satellites, the inversion accuracy of indicators such as leaf area index (LAI) and yield was significantly improved. If meteorological and field management data are added, the model effect will be better (Islam et al., 2023; Chen et al., 2024). At present, multi-source feature fusion (such as combining average spectra, vegetation indices and textures) generally performs better than that of a single data source.
RkJQdWJsaXNoZXIy MjQ4ODYzNA==