Estimating maize LAI by exploring deep features of vegetation index map from UAV multispectral images

Visible and near-infrared imaging spectroscopy is an efficient method for non-destructive estimation of crop leaf area index (LAI), in which effective features extraction is a key factor affecting model accuracy. Vegetation index like normalized difference vegetation index (NDVI) is a typical parameter extracted from spectral images for LAI estimation. However, it always faces challenges of values saturation and sensitivity decreasing due to the canopy coverage changes and complex environment influences. Therefore, the study aims to propose a method to explore deep features based on deep learning model (ResNet50) and vegetation index (VI) map, by which contributes to improve the accuracy of maize LAI estimation. In two-years’ experiments of 2020 and 2021, the multi-spectral imaging sensor carried by the unmanned aerial vehicle (UAV) was used to collected remote sensing images of the maize canopy during multi-growth stages under different nitrogen fertilizer treatments, and a total of 792 LAI values were collected. In terms of feature extraction, 10 VIs were extracted based on the spectral images of maize canopy, and comprehensive features (texture and deep features) were extracted based on the VI map. In terms of model construction, partial least squares regression (PLS) models were constructed based on three feature types. The results showed that normalized difference red edge (NDRE) VI was the best indicator for monitoring maize LAI among all the VIs investigated. Meanwhile, LAI multivariate linear model constructed based on texture features extracted from NDRE images had better accuracy. Deep features of NDRE images were further extracted by ResNet50 model. And the maize LAI estimation models based on PLS were constructed with VI, texture features and deep features of two-year data, respectively. The results showed that compared with VI and texture features, the LAI estimation model based on deep features showed optimal accuracy in two-year data (R2 = 0.827, greater than 0.781 and 0.780; RMSEP = 0.405, less than 0.455 and 0.455; MAE = 0.320, less than 0.366 and 0.363). In addition, the LAI estimation model based on deep features was successfully evaluated with 2020 and 2021 validation data (2020: R2 = 0.872, RMSEP = 0.379, MAE = 0.294; 2021: R2 = 0.733, RMSEP = 0.433, MAE = 0.348). The experimental results demonstrate the effectiveness of the proposed deep feature extraction method based on deep learning model and VI images in LAI estimation, which provides a feasible method for monitoring crop growth information based on UAV platform.