Crop classification with UAV multispectral remote sensing, employing an enhanced ResNet50 residual network

Abstract

The ResNet50 residual network model is characterized by its rapid training speed and high classification accuracy, demonstrating significant advantages in extracting and classifying crop features. However, the correlation and redundant information among different spectral bands in multispectral data can negatively affect classification accuracy. Additionally, the ResNet50 model does not meet the requirements for training with multi-band data, and its classification accuracy in small-scale, high-precision scenarios needs further improvement. This study focuses on flue-cured tobacco and maize to address these challenges, identifying the optimal classification band combination. Enhancements to the ResNet50 model were made by incorporating Batch Normalization (BN) layers, pyramid pooling layers, and hidden layers, experimenting with seven different combinations. The experimental results indicate that the RGB + NIR+Edge combination is the most effective for classifying flue-cured tobacco and maize, achieving an accuracy of 94.48%, precision of 94.66%, recall of 94.48%, kappa coefficient of 91.72%, and an F1 score of 94.49%. Among the seven improvement strategies, solely introducing BN layers yielded the most substantial improvements, increasing accuracy by 0.42%, precision by 0.51%, recall by 0.42%, kappa coefficient by 0.63%, and the F1 score by 0.44%.