Salman Goodarzdashti; Mohamad Seifi; Mahshid Kohandel; Davoud Ashourloo; Hossein Aghighi
Abstract
Potatoes are the fourth most cultivated crop worldwide. Regarding the strategic role of this crop in food security, accurate potato mapping provides essential information for national crop censuses and potato yield estimation /prediction at any scale. Although remote sensing (RS) approaches based on ...
Read More
Potatoes are the fourth most cultivated crop worldwide. Regarding the strategic role of this crop in food security, accurate potato mapping provides essential information for national crop censuses and potato yield estimation /prediction at any scale. Although remote sensing (RS) approaches based on optical and/or microwave sensors have been widely employed to monitor cultivated land (including crop area, type, condition, and yield forecasting), the identification of potato planting areas using RS data has not been much addressed. Hence this study addresses the literature gap by suggesting an effective potato mapping approach that uses the time series of the Sentinel-2 (S2) images, Google Earth Engine (GEE) platform and machine learning methods. Since most crops have specific spectral and temporal characteristics during the growing season, this research has presented a method to discriminate potato fields from other crops using time series images without explicit thresholding. We employed 1648 ground truth data to optimize, train, and evaluate the model at the study site, which includes potatoes and other fields. A handheld GPS receiver was used to collect these data. The performance of this approach is evaluated by conducting a set of experiments in Hamedan and Bahar cities, as the regions grow more potatoes than any other places in Iran. Accurate identification of potato fields was completed by extracting the required features, namely the potato phenology feature and NDVI medians, from the time series of the S2 satellite bands. After that, these features were utilized as the input parameters to Support Vector Machine (SVM) technique. In order to train the most optimal SVM model using RBF kernel, Gamma and C values were optimized with the help of the 5-fold cross-validation method. These values were then employed during the algorithm's implementation on GEE platform. The estimated overall accuracy and Kappa coefficient are 90.9% and 0.82 for Hamedan and 93.3% and 0.87 for Bahar, respectively. The results of this research indicate the efficiency of SVM technique in potato acreage mapping. Moreover, the selected features such as potato phenology feature can be considered as discriminating features for improved identifying of crop farms.
