LMPF: lightweight multi-scale prediction framework in lettuce agriculture

• LMPF achieves superior accuracy (avg R 2 =93.77%) in predicting five lettuce growth traits with only 1.02M params. • LMPF outperforms heavyweight models (e.g., ResNet101) while using over 40 × fewer params, enabling efficient edge deployment. • An adaptive pyramid module (GPAR-Pyramid) robustly handles plant scale variation from seedling to maturity. • An intelligent attention mechanism (MTASCA) effectively fuses RGB and depth data, boosting multi-task prediction performance. Accurate and lightweight lettuce phenotyping throughout growth stages is crucial for field monitoring and precision cultivation. However, it remains a challenge due to device and computational limitations. To tackle this issue, we propose a novel compact network for lettuce biomass prediction, named the Lightweight Multi-Scale Prediction Framework (LMPF). By integrating RGB and depth image modalities, LMPF improves the generalization of phenotypic representation across different varieties and growth stages. It effectively exploits their complementary features to capture fine-grained biological traits vital for precision agriculture.The specific contributions are as follows: First, the Depthwise Separable Dense Fusion (DSDF) module, which combines depthwise separable convolution with dense connections, substantially reduces the number of parameters while preserving rich feature representations. Second, the Growth-stage Progressive Adaptive Receptive field Pyramid (GPAR-Pyramid) module builds a hierarchical pyramid structure based on dilated convolution. It dynamically adjusts the receptive field and multi-scale responses via learnable branch weights to more efficiently capture scale variations and structural details from seedlings to mature plants. Finally, the Multi-task Adaptive Spatial-Channel Attention (MTASCA) module uses learnable attention weights to intelligently fuse RGB and depth channel information, effectively alleviating the underutilization of cross-modal information that might result from simple concatenation.The evaluation of LMPF on a public lettuce dataset showed an average coefficient of determination (R 2 ) of 93.77% for five growth traits and an average NRMSE of 16.23%. With only 1.02M parameters, the model demonstrates an excellent balance between lightweight design and performance.