Extended Abstract
Introduction
Vegetation is considered to be one of the most important elements in all major ecosystems on the Earth. Thus, a proper understanding of vegetation and its growth trends and other environmental factors has always been of particular importance for environmental research. Estimating vegetation phenology parameters (VPPs) requires continuous NDVI data collection over a specific period of time. However, soil moisture, cloud cover, and particulate matter may affect the energy reflected from the vegetation cover and result in noisy images or erroneous data. Vegetation phenology parameters cannot be extracted from raw data due to the presence of random errors. These errors do not follow the phenological process and thus, overestimate or underestimate NDVI and fail to produce accurate results. Smoothing functions and especially the TIMESAT model are used to resolve this issue and eliminate errors in the NDVI time series. There is still no general consensus on which function acts more efficient and accurate in the TIMESAT model especially regarding the highlands. Naturally, each method yields different results in different regions, and thus it is necessary to compare and evaluate different functions used in the TIMESAT model and determine their accuracy in producing a continuous time series. The present study aimed to evaluate the performance of various functions such as asymmetric Gaussian (AG), double-logistic (DL), and Savitzky–Golay (SG) used to extract VPPs especially in mountainous regions.
 
Materials and Methods
TIMESAT model is a time-series analysis model based on remote sensing (RS) vegetation indices. It includes three functions: Savitzky–Golay, asymmetric Gaussian, and double-logistic, which are used to smooth collected data and identify outliers. Savitzky–Golay is an adaptive-degree polynomial filter (ADPF). The other two functions fit the information using nonlinear functions. These functions use unmodified NDVI data as input to produce modified and smoothed NDVI output. Four wheat farms in cold and warm regions of Khorramabad were used in the present study to investigate plant phenological behaviors and extract VPPs. The northern and eastern parts of Khorramabad have a cold climate, while the southern and western parts have a warm climate. One-year time series (2020) data of MODIS sensor was used in the present study. Using the infrared and near-infrared spectral reflectance values, NDVI was calculated in the Google Earth Engine environment. Errors of the NDVI time series were first corrected and a phenology curve was produced for wheat in both warm and cold farms. Asymmetric Gaussian, double-logistic, and Savitzky–Golay filter functions were also used to adapt the NDVI data. Following the reconstruction of growth curves in the time series of vegetation indices and smoothing the curve, various VPPs such as start of the season (SOS), end of the season (EOS), middle of the season (MOS), length of the growing season (LOS), base limit and value, maximum NDVI, vegetation growth season range, large seasonal integral, and small seasonal integral were extracted.
 
Results and Discussion
The model indicated that on average, beginning of the wheat growing season (SOS) in the warm regions of Khoramabad coincided with the 31.5th day of the year in the Gregorian calendar, whereas it happened on the 90th day of the year in the cold regions, thus indicating a 1.5-2 month difference between the beginning of the wheat growing season in cold and warm regions. The wheat growing season ended (EOS) on the 163rd day of year in the warm regions and on the 193rd day in the cold regions. In addition, in order to analyze the effect of climate on VPPs such as SOS and EOS, a comparison was made between the parameters obtained from farms in warm and cold regions. On average, the peak of vegetation growth has occurred in late March (Mar. 28, 2020) in farms of warm regions while cold regions experienced the peak of growth on May 20, 2020. In other words, warm regions have experienced peak growth approximately two months earlier than cold regions. Finally, the models were assessed and obtained values were compared with ground-based data collected in field surveys. Validation results showed that with an average RMSE of 2, Savitzky–Golay smoothing model reconstruct data more accurately as compared to asymmetric Gaussian, and double logistic function with an RMSE of 4 and 11, respectively. In other words, Savitzky–Golay estimates SOS and EOS with a higher accuracy and lower errors.
 
Conclusion
Findings indicate that Savitzky–Golay filter outperformed asymmetric Gaussian and double logistic functions in extracting VPPs in mountainous areas. Accordingly, it is suggested to use Savitzky–Golay in future studies aiming to investigate the phenological behavior of different vegetation covers in other Iranian highlands. The study has also showed that different climatic conditions within the study area affect plant phenological behaviors, which can lead to differences in SOS, peak of growing season, and EOS in different cold and warm regions of the province. Growing season of plants in cold regions of the province has occurred with an approximately two-month delay compared to the warm regions of the province.