Hyperspectral image segmentation for maize stubble in no-till field

Abstract. Conservation agriculture (CA), taking zero tillage and organic soil cover as the fundamental principles, is gaining increasing worldwide attention for its pronounced contribution in mitigating soil degradation and boosting soil fertility. However, in North China Plain where annual maize-wheat rotation is dominantly applied, the standing maize stubble, especially its well developed root system remains in field, can hardly decompose after a considerably limited duration between maize harvesting and wheat sowing, which has caused great difficulties such as blocking for subsequent no-till sowing. Vision-based guidance is of great help for no till planter to avoid the aboveground standing stubble and consequently the belowground root system during wheat planting. Yet the similarity between colors of the stubble and the background makes it difficult to achieve accurate and prompt target detection by traditional vision identification. Therefore the hyperspectral imaging method was used in our study for better image segmentation. Hyperspectral images of standing maize stubble left in field were randomly obtained in a spectral region ranges from visible region to near infrared region (VIR-NIR, 347.4-952.8 nm) using a hyperspectral imaging system. The pre-processed data were subjected to the principle component analysis (PCA) to select the optimal wavebands for image segmentation; afterwards the PCA was carried out again based on these selected wavebands. The ultimate principle component images and images corresponding to each optimal waveband were enhanced via image filtering algorithms. Band fusion operation was also performed to eliminating random noise and interference regions in images. As indicated by the results of image enhancement and band fusion, hyperspectral imaging method is a promising way to achieve better discrimination of maize stubble under no-till conditions.