6 / 2023-02-27 11:24:59

Fine classification of typical crops based on UHD185 hyperspectral data

UHD185; hyperspectral image stitching; machine learning; ENVINet5; winter wheat classification

ABSTRACT: Winter wheat is the main crop, and there are many varieties of winter wheat in China. Monitoring its growth and providing it with the best growth conditions punctually and accurately are the main topics of smart agriculture research. Different varieties of winter wheat have different requirements for water, soil fertility, climate, etc. in each stage of their growth. Therefore, to better understand winter wheat future production and realize its smart production, independent identification of winter wheat varieties and distinguishing different winter wheat categories are necessary. The spectral characteristics of different varieties of winter wheat are very similar, which is a typical problem of weak inter-class characteristics, thus it is quite challenging to classify them with remote sensing methods. UAV hyperspectral remote sensing could combine high spatial resolution and high spectral resolution to truly realize ‘space-spectrum integration’, which may provide the possibility to solve this problem. To this end, this study aims to refine the classification of different varieties of winter wheat by integrating UAV and UHD185 hyperspectral camera into a hyperspectral remote sensing system. The study takes Henan Wuzhi Winter Wheat Research Base as the experimental area--firstly, obtaining the hyperspectral data of the experimental area; secondly, splicing and reducing dimension of the original data; finally, introducing the machine learning and deep learning theories to evaluate the refined classification methods of different varieties of winter wheat. The main work and conclusions of this paper are as follows:

(1) Integrating DJI M600 UAV and Cubert UHD185 hyperspectral camera to form a UAV hyperspectral remote sensing system. The UHD185 camera has a small frame, only 50×50 pixels, which can be used effectively after stitching. In this research, through multiple loop experiments of drone flight photography and hyperspectral image stitching, it is finally determined that the route parameters that can meet the hyperspectral image stitching of the weak feature area of winter wheat and the optimal heading overlap and side overlap are 90%, 85%.

(2) Aiming at the stitching problem of UHD185 images, firstly this study uses Pansharpen algorithm to perform hyperspectral data fusion, and secondly extracts and matches image feature points based on the improved SIFT algorithm PhotoScan, and then uses IDL programming to realize the extraction and merging of hyperspectral sub-band images. Finally, the geocoding of hyperspectral images is completed based on ArcGIS to realize the stitching of UHD185 images. The correlation coefficients of the spectral curves of typical ground objects before and after splicing all reached 0.970, and the image after splicing effectively retained the original spectral characteristic information.

(3) Principal component analysis transform (PCA), minimum noise separation transform (MNF) and independent component transform (ICA) are used to reduce the dimensionality of UHD185 data. The results are compared with Support Vector Machines (SVM), Maximum Likelihood Classifier (MLC), Independent Component Analysis (ICA), and the improved ENVINet5 model based on the U-net neural network, respectively. The results show that the MNF-ENVINet5 dimensionality reduction classification result is the best, and its overall accuracy OA and Kappa coefficients are 78.98% and 0.7555, respectively.