Multi-Scale Citrus Chlorophyll Transfer Learning Based on Convolutional Neural Network

Leaf chlorophyll content (LCC) is a key indicator for assessing crop health. However, cross-scale inversion faces challenges such as spectral mixing effects, atmospheric interference, and data scarcity. The traditional vegetation index (VI) method exhibits insufficient stability in cross-scale applications, while deep transfer learning provides a novel approach for multiscale transfer. This study systematically explores the transfer-ability of spectral features across scales by constructing a trans-fer learning framework integrating vegetation indices and convolutional neural networks (CNN) using spectral data from leaf, canopy, and UAV platforms. The results demonstrated that transfer learning significantly improved inversion accuracy at the canopy and UAV scales: the canopy-scale model achieved an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 7 9 9}$</tex>; the direct transfer strategy (leaf-UAV) at the UAV scale yielded an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{2}$</tex> of 0.784, outperforming the stepwise transfer strategy (leaf-can-opy-UAV) with an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 7 4 5}$</tex>. This method provides technical support for multi-source remote sensing data fusion and precision agriculture management.