Abstract:Crop classification studies generally focus on different types of crops, while there are fewer studies on the fine classification of different cropping patterns of the same crop. Research on the spatial distribution of seed production is essential to control the maize market, as private seed production and concealment of acreage are occurring in the maize seed market. Seed maize and common maize are the two modes of maize cultivation. Accurate identification of both and remote sensing mapping of the spatial distribution of seed maize are essential for maize seed industry and food security.Traditional crop remote sensing classification methods require a high number and distribution of samples, while visual interpretation of crop samples is difficult. How to improve the utilization of collected samples and at the same time reduce the dependence on samples in fine classification is a pressing issue nowadays. Based on this, combining migration learning with unsupervised classification methods, firstly, using the idea of transfer learning, Linze and Wuwei were used as the source domain, and the feature engineering in the source domain was constructed, including 8 original spectral bands BLUE, GREEN, RED, EDGE1, EDGE2, EDGE3, NIR, SWIR, and 18 vegetation indices NDVI, EVI, RVI, GNDVI, TVI, DVI, MSAVI, GCVI, RNDVI, NDRE, RRI1, RRI2, MSRRE, CLRE, IRECI, LSWI, GCI, SIPI. Then, the features that best characterized the differences in canopy spectra between seed maize and common maize and that differed least between seed maize in different source domains were extracted. Finally, it was used as prior knowledge in an unsupervised classification task in the target domain. The results showed that the near-infrared primordial band exhibited the strongest advantage among the many features. By comparing the classification accuracies of three time-series ranges, namely, before the removal of male ears from the seed maize females, after the removal of male ears from the seed maize females, and during the full-life span of maize growth, the NIR bands were best characterized after the removal of male ears from the seed maize females, i.e., when the DOY was 125~210.In order to further extract the NIR primary band information and enhance the performance of the features, the slope of the linear regression equation in the NIR band of the seed maize female after removal of male ears was used as a feature, and the K-means unsupervised method was used to classify the seed maize in the target domain of Shihezi and Kuitun.After comparative experiments on two target domains in 2019 and 2020, this method mostly showed different degrees of improvement over the K-means classification accuracies characterized by the near-infrared primitive bands in the same time period.In the two target domains Shihezi and Kuitun seed maize had F1 value of 74.35% and 64.97% in 2019 and 72.50% and 75.69% in 2020, respectively.This method effectively improved the utilization of samples by extracting prior knowledge and introducing an unsupervised classifier. It provided ideas for feature enhancement of the primary band by extracting the slope of the band regression as a feature, and provided a method for fine classification mapping of crops in sample-free scenarios.
