Investigating the relation between agricultural and meteorology drought using MLP neural network in northwest Iran

Introduction
Drought can be considered as a dry period that lasts long enough to cause an imbalance in the hydrological situation. In the calculation of drought parameters, four characteristics of drought intensity, duration, frequency and extent are studied. In general, drought is divided into 4 main categories. The most basic type of drought is defined as meteorological drought, which is caused by a lack of precipitation, an increase in air temperature, and evaporation. In the long run, this phenomenon will lead to hydrological drought and lack of surface and underground water resources, and as a result, agricultural drought and decrease in soil moisture and loss of vegetation. Agricultural drought begins when the amount of moisture in the root of the plant decreases to such an extent that it causes wilting and ultimately the reduction of agricultural products. the severity, volume of damages, the boosting trend of drought and its negative economic, social and environmental effects can be predicted and controlled. So, damages can be Minimize the consequences. Also, remote sensing technology has made it possible to evaluate variable surface phenomena namely, drought. In recent decades, due to the nonlinear nature of the phenomena, artificial neural networks have shown the best ability in modeling and forecasting time series in hydrology and water resources engineering. On the other hand, artificial neural networks are able to identify the nonlinear relationship between input and output variables from the data structure. Drought monitoring in Iran, done through methods based on weather stations, is not accurate due to the lack of a scattered network and lack of access to timely data. Remote sensing technology, along with geographic information system, by creating appropriate spatial and temporal capabilities, has made it possible to evaluate and monitor variable surface phenomena such as drought, so that in the last two decades, the use of methods based on satellite data for Drought monitoring has become one of the first priorities of research and specialized organizations. Drought prediction in water resources systems plays an important role in reducing drought damages. In the last few decades, mathematical models have been widely used to predict drought. These models take time series into account and model processes linearly. In recent decades, due to the nonlinear nature of the phenomena, artificial neural networks have shown the most ability in modeling and forecasting time series in hydrology and water resources engineering.
 
Materials and Methods
In this research, a standardized precipitation index (SPI) and the normalized difference vegetation index were used to analyze the correlation in the mountainous climate of Iran. Firstly, monthly rainfall data of 88 meteorological stations from 2000 to 2018 were gathered. After performing the necessary statistical tests, the SPI values were calculated in time scales of 1-, 3-, 6-, and 12-month). Then, OLI sensor images of Landsat 8 satellite with a resolution of 30 m were used to extract the NDVI. These images were obtained from USGS on a monthly basis between 2013 and 2018. In total, 72 months were studied in the entire statistical period. After performing radiometric and atmospheric corrections, an average image was prepared every month for NDVI calculation. Then, a multi-layer perceptron (MLP) neural network was used to predict NDVI data for the next month. Last month's NDVI data and one-month SPI were used as input data to predict the next month's NDVI data during the growing season.
 
Results and Discussion
According to NDVI, between 2013 and 2018, May is responsible for the highest amount of vegetation density. In addition, SPI-1 shows the amounts of droughts with more intensity and accuracy than other time steps. Hence, in the mountain region of Iran SPI of the dry season takes a larger amount during the first & last months of the year while during summer, especially in October, drought is much more visible. According to SPI, the return period of droughts is 5-6 years. There is a significant correlation between monthly SPI data and NDVI in the growing season. The highest Pearson correlation coefficient between SPI & NDVI is related to SPI with a 1-month time series and the value of this correlation is much higher in April and May. So, the lack of rain in these months will cause a reduction in growing agricultural products in the spring.
 
Conclusion
Artificial neural networks are able to identify the nonlinear relationship between input and output. In this type of simulation, even when the set has disturbance and measurement error, the neural network will be able to provide good results. If there is a change in environmental conditions over time, the neural network will be able to provide new results by adjusting new parameters. NDVI has the highest sensitivity to changes in vegetation and is more useful against atmospheric and soil effects, except in cases where there is little vegetation. In conclusion, for predicting vegetation changes during growing seasons in the pastures of the mountainous climate of Iran, using NDVI data and the monthly SPI data is an efficient process. Therefore, it can be concluded that the neural network is a capable model in relation to agricultural drought prediction.