Automatic silage maize detection based on phenological rules using Sentinel-2 time-series dataset

Abstract

ABSTRACT Recently, availability of valuable high spatial and temporal resolution optical satellite imagery such as Sentinel-2 has provided a great opportunity for a wide variety of crop-related studies including crop phenology detection, crop biophysical and biochemical parameter estimation, and crop-type mapping. Most current crop mapping methods focus on supervised classification, demanding a large number of field sampling points to feed the classifier. In this paper, we propose an automatic method for silage maize mapping based on Sentinel-2 time-series data in three agricultural areas in Iran (namely Marvdasht, Abyek, and Mashhad) and an individual area in USA (Tulare County, California). We considered Marvdasht as the training site and other three sites as the test sites. After the preprocessing step, temporal profiles of the Normalized Difference Vegetation Index (NDVI) were computed from Sentinel-2 time-series images in 2017 and 2018. Then, the Savitzky–Golay filter was applied to the NDVI profile for noise reduction. Linear interpolation was also used for the construction of continuous NDVI time-series during the maize growing season. Then, phenological parameters were extracted in Marvdasht, for maize and other crops. After comprehensive analyses, the ratio of the slope from the peak of the greenness to the harvest (SPGH) to the length of the growing season (LOS) was suggested as an appropriate variable for automatic maize detection. Evaluation of this variable in the study sites confirmed its feasibility for automatic maize mapping, with the kappa coefficient values of 0.89, 0.8, 0.9, and 0.8 in Abyek, Marvdasht, Mashhad, and Tulare in 2017, respectively. The performance of the suggested method for the second year (2018) is close to those of 2017 which means that maize detection algorithm results are consistent and robust across years. Moreover, to ascertain high potential of the proposed variable, the results were compared with those of the maximum likelihood (ML) and support vector machine (SVM) classifiers in the test sites. The performance of SVM (with a large number of training samples) was analogous to that of the suggested automatic method in terms of accuracy; however, inferior performance of ML was evident. The results of this research demonstrated the great potential of our proposed phenological method in automatic silage maize mapping.