Classification of seed maize using deep learning and transfer learning based on times series spectral feature reconstruction of remote sensing

Accurately mapping the spatial distribution of seed maize fields is critical to securing seed supply, yet seed maize classification remains challenging due to similarities in interspecific crop cultivation. This study aims to construct a more suitable classification system for seed maize. A feature construction method based on time series spectral reconstruction was proposed, which explicitly enhanced the temporal-spectral correlation by reflecting the spectral reflectance information along with the temporal constraints onto the pixel-level grayscale image simultaneously. To clarify the benefits of classification strategies on the task of fine interspecific classification of maize, we compared two classification strategies, end-to-end direct classification and hierarchical classification, from the perspectives of both time cost and accuracy. The results showed that under the optimal time series range (March-early September) obtained by incorporating agricultural knowledge, ResNet-101 achieved an average accuracy of 91.34 %, which was better than other classification models. The input feature importance analysis revealed the classification mechanism of the model throughout the growth period. In order to improve the generalization ability of the model, we constructed two transfer learning frameworks for comparison. The accuracy of the method of constructing a joint dataset to train the model improved faster when the proportion of target domain samples introduced was small; the accuracy of the source domain pre-training-target domain fine-tuning method was higher when the number of samples introduced was larger. This study may provide a reference for the interspecific classification problem of other crops.