یک روش نمایش spare برای تشخیص زمین‌های زراعی زعفران با استفاده از تصاویر ماهواره‌ای sentinel-ii

امروزه، مدیریت کشاورزی از طریق فناوری سنجش از دور جایگاه ویژه‌ای در بین مدیران و افرادی که عهده‌دار این صنعت هستند، پیدا کرده است. زعفران (طلای سرخ) یکی از محصولات کشاورزی خاص ایران با ارزش اقتصادی بالا است که در زمینه‌های مختلف صنایع غذایی و پزشکی استفاده می‌شود. با توجه به شرایط کشت زعفران، شرایط پایدار برای کشت در زمین‌های زراعی وجود ندارد و نمی‌توان توصیه کرد که زعفران به‌طور مداوم در یک زمین کاشته شود. بنابراین، مساحت کشت آن هر سال متفاوت است و پیش‌بینی عملکرد سالیانه می‌تواند برای اهداف مدیریتی مفید باشد. در این مقاله، با توجه به رفتار فنولوژیکی مزارع زعفران، شناسایی این مزارع با استفاده از یک الگوریتم شناسایی هدف نوین پیشنهاد شده است.

Nowadays, agricultural management via remote sensing technology has gained a special position among managers and the people who are in charge of this industry. Saffron (Red Gold) is one of specific Iran’s agricultural products with a high economic valance which is used in different fields of food and medical industries. Considering the cultivation conditions of the saffron, there has not a persistent condition to plant in farmland, and it could not be recommended to plant saffron on the same land continuously. So, their cultivation area varies every year and the prediction of annual yields could be useful for managing aims.

In this paper, considering the phenological behavior of the saffron farmlands, the detection of these farmlands using a novel target detection algorithm is proposed. To do so, a time series of the Normalized Difference Vegetation Index (NDVI) extracted from Sentinel-2 satellite images have been used as the indicator of the phenological of cultivation areas. In the proposed method, a sparse representation method is used as the target detector. In this procedure, each pixel of the NDVI time series is reconstructed through a dictionary consists of the spectra-temporal response of the saffron farmland and background samples.

The sub-dictionary of the background samples has randomly sampled from a clustered feature space spanned by time series pixels. A filtering step has also been designed to avoid the selection of the target-like samples in the sub-dictionary of the backgrounds. On average, the results achieved in three different datasets in the Neyshabour city have reached 93.1% accuracies. Also, the proposed method in comparison with the well-known target detectors CEM, ACE, MF, and the parallelepiped and SVM classifiers have been indicated, on average, the 4.8% accuracy improvements.