Multi-level wavelet deep learning for soil property prediction based on multi-platform, multi-modal remote sensing data

In precision agriculture, accurately predicting soil properties at the field scale is crucial for maintaining soil functions and developing site-specific fertilization strategies. Although unmanned aerial vehicle (UAV) imagery offers high-resolution spatial information, existing multi-source soil mapping approaches still face a structural limitation: high-resolution UAV data and medium-resolution satellite observations differ substantially in spatial support, feature representation, and information density, making it difficult to jointly capture within-field heterogeneity while maintaining regional transferability. In addition, multimodal inputs often contain heterogeneous and partially redundant signals, which weakens stable cross resolutions information coupling. To address these issues, this study proposes an integrated framework that couples a multi-level wavelet convolutional neural network with a Lasso-guided channel re-weighting mechanism (MW-CNN-L). This framework effectively learns from and integrates information derived from UAV and satellite imagery, highlighting spectral, textural, and structural features that are critical for soil property prediction under multimodal inputs. Specifically, spectral indices and textural features were derived from UAV and Sentinel-2A imagery and combined with Sentinel-1 data and multiple terrain characteristics as model inputs. Wavelet downsampling was employed to expand the effective receptive field while preserving sensitive features, thereby enabling scale-dependent soil responses to be decomposed and aligned across input sources. Lasso-guided channel re-weighting was applied to assign sparse and interpretable weights across variables, enhancing model robustness and physical interpretability. Based on this framework, we developed Strategy (i) which used only satellite imagery, and Strategy (ii), which integrated UAV and satellite imagery, to evaluate the performance of MW-CNN-L with multiple algorithms, including CNN and MW-CNN. Additionally, we assessed the model’s predictive capability at a larger scale by transitioning from the field-level to the farm-level. The results demonstrate that: (1) under the MW-CNN-L model, Strategy (ii) achieved an average improvement of 0.05 in the coefficient of determination (R2) and 0.26 in the ratio of performance to interquartile range (RPIQ) compared with Strategy (i) using the MW-CNN-L model, while reducing the root mean square error (RMSE) by 0.98 g kg−1. This highlights the necessity of incorporating UAV spatial information for farm-scale predictions; (2) MW-CNN-L effectively improved the predictive accuracy of both CNN and MW-CNN for soil prediction; and (3) in regional-scale validation, the model achieved an R2 of 0.71 and an RMSE of 6.99 g kg−1. After a simple post-hoc calibration, the external performance improved to R2 of 0.73 and an RMSE of 6.71 g kg−1, indicating reasonable transferability under cross-domain conditions rather than strict same-condition generalization. Overall, MW-CNN-L exhibits advantages for soil mapping through the integration of multi-platform, multimodal information and provides a structured framework for cross-scale mapping.