Lightweighting of kiwifruit root soil water content inversion model based on novel vegetation indices

• The established composite visual convolutional regression network can accurately estimate kiwifruit root soil water content • Convolution-derived vegetation indices correlate strongly with soil water content • The CVI-RSWC model is lightweight with a remarkable effect • A new method for real-time monitoring model deployment of orchard information embedded platform was provided With the increasing demand for real-time monitoring in orchard remote sensing, model lightweight has become a critical research focus. Aiming at the insufficiency of model lightweight in kiwifruit root soil water content (RSWC) inversion, which makes the model difficult to deploy in application scenarios with limited computational resources of embedded platforms, this study proposes a lightweight method of kiwifruit RSWC inversion model with novel vegetation indices. This research introduces two novel vegetation indices, the Convolutionally Derived Near-Infrared-Red Vegetation Index (CNRVI) and the Convolutionally Derived Near-Infrared-Green Vegetation Index (CNGVI), developed based on the Composite Visual Convolutional Regression Network (CVCRNet). Collectively termed Convolutionally Derived Vegetation Indices (CVIs), these indices serve as features for constructing a multivariate regression model, CVI-RSWC, to generalize the entire CVCRNet model inversion process. The results showed that the R 2 of the CVCRNet prediction result was 0.743, and the RMSE was 0.887; the R 2 of the CVI-RSWC prediction result was 0.636, and the RMSE was 1.219. Although the model accuracy was slightly reduced, the model efficiency was greatly improved, in which the number of parameters was reduced by 99.74%, the model size was reduced by 86.15%, and the inference speed was up to 96265.87 FPS. The model improvement in this study effectively balances inversion accuracy and model lightweight. It provides a new approach for deploying the Kiwifruit RSWC inversion model on mobile and airborne platforms with low computing power.