Ziheng Sun, Liping Di, Hui Fang

Remotely sensed signals from crop fields are full of variabilities due to the complex interactions among the environment, seeds, climate, market, and farmers. It is a common phenomenon that the crops in neighbouring fields are in different growing stages, e.g., the corns are in the V5 leaf stage in one field and V10 stage in another neighbouring field. The phenomenon results in remote sensing images that are non-ideal for producing crop maps as the crops are in different phenological stages across the fields. For instance, July is the best month for monitoring soybeans but bad for monitoring winter wheat which has been harvested before July. Most of the wheat fields are already harvested in July and appear like fallow fields on satellite images. This paper proposes a customized classification approach based on greenest pixels (GP) to enhance the quality of the satellite images for mapping. The greenest pixels are obtained by calculating the highest NDVI (Normalized Difference Vegetation Index) values of every pixel from all the captured images each year. In the process, we filtered the bad quality pixels like clouds, ice, snow, shadow, etc. A filtering step is added to distinguish the non-vegetation and vegetation pixels first. The overall workflow uses state-of-art remote sensing classification techniques. Machine learning (ML) algorithms like KNN, Gaussian Naïve Bayes, Decision Tree, AdaBoost, Random Forest, SVM, and Neural Networks were used simultaneously to evaluate the approach. The study area is located in the state of Nebraska and the satellite imagery used includes Landsat 8 surface reflectance products. The ground truth data comes from field surveys, roadside samples, and USDA (United States Department of Agriculture) crop maps. Google Earth Engine was used to accelerate the data pre-processing. We tested all the ML models on two sets of experiments: GP and non-GP. In each set, the training has only one-year data (2013) and the testing uses the rest years (2014-2018). The results show that the proposed GP-based approach can significantly improve the classification precision by ~15% (from ~70% to ~85%) on average. This research proves that greenest pixels have large potential and should be considered as the major input data in the crop mappings in the future.