Journal of Spatial Analysis

---

Comparison of the effectiveness of four artificial intelligence methods in predicting drought
Abstract
Problem statement:
Drought is a temporary disorder whose characteristics vary from region to region, therefore, it is not possible to define a complete and absolute definition of drought. Drought is one of the most important natural disasters that can occur in any climate regime. Since drought is unavoidable, it is important to know it in order to optimally manage water resources. Drought prediction can play an important role in managing this phenomenon. In other words, recognizing and predicting this phenomenon is one of the topics of interest for scientists who are interested in solving the problem of water shortage. More than 80% of Iran's area is covered by arid and semi-arid regions and lack of rain is a predominant phenomenon in this region. So far, several methods have been proposed to predict drought. Each method offers different results in specific conditions.  Therefore, identifying the best method for predicting drought in the climatic conditions of central Iran is essential.
 
Material and methods:
In this research, in order to introduce a suitable method for predicting drought for the next month, four methods of artificial intelligence including Deeplearning (using the Alexnet network, one of the convoluted networks), K nearest neighbor algorithm (KNN), multi-class Support vector machines (SVM-MultiClass) and decision tree have been used. Monthly rainfall data from 11 syntactic stations of Yazd province during the 29-year statistical period (1988 to 2017) were used as experimental data. Standardized precipitation index (SPI) was calculated to indicate drought status in terms of severity and duration on 1, 3, 6, 9, 12 and 24 month time scales. Precipitation data was used as neural network input and SPI classification as network output and 80 percent of the data was used for training and 20 percent for testing the networks.
In this study, the Recurrence Plot method was used to interpret the time series to convert these series into images and RG and B pages were created. To convert rainfall data into photos at 1, 3, 6, 9, 12 and 24 month time scales, photo layers were created using a standardized rainfall formula, and by merging these three output layers into colored photos and black and white photos were obtained. Using three pages created in MATLAB software and merging them, the output was in the form of a photo, which was placed as the input of the Alexnet network. Combination of Recurrence Plot to create images and deep learning network for classification of drought data has been used for the first time in this research. To evaluate the effectiveness of the classification strategy, standard criteria of accuracy, micro-F1 and macro-F1 were used.
 
Results Description and interpretation:
 The results showed that all networks were able to predict drought. However, on short time scales such as 3 and 9 months, the accuracy assessment criteria for some classes are zero and the methods of learning from these classes are practically ignored due to their lack of data. But on a larger time scale, this issue has been addressed and the data of those classes are well categorized. Deep learning network with image input could not predict well in the short term due to lack of data, but in the long term due to increased data has improved its performance and had the best performance. The SVM method at different time scales has shown unreliable and variable behaviors that can not be said to be a suitable method for predicting drought at different time scales. Decision Tree and KNN methods have been able to predict drought better in the short term than in the long term. The two methods have been closely related. .Based on the Deeplearning network macro-f1 evaluation criterion, the one-month time scale with 22.71% was the most inefficient method and the Decision Tree with 64.65% was the most efficient method, But with the increase in time scale, the Deeplearning network improved its performance, so that at the 24-month time scale with 65.35%, the best performance for the Deeplearning network followed by the SVM-MultiClass network with 57.40%. For future research, it is suggested that Decision Tree and KNN methods be used to predict short-term drought. In this study, with increasing the time scale and increasing the data used, these two methods have lost their effectiveness compared to the short term.
 
key words: Drought, Standardized Precipitation Index, Artificial Intelligence, Deep Learning, Alexent, Recarence Plot
 

---

Nowadays, deep learning as a branch of artificial intelligence acts as an alternative for human with hopeful outcomes. Open Street Map as the biggest open source data is used as a complementary data sources for spatial projects. It is notable that is some advanced counties the accuracy of VGI data is higher than governmental official data. This research aims to use artificial intelligence to produce and subsequently promote completeness of OSM data. Res_UNet architecture was utilized to train landuse categories to the network. The result shows that IoU metric is about 83 percent that implies a high accuracy paradigm. Then, united-based method was used to calculated completeness of OSM data. The unit-based results show that completeness of building blocks, forest, fruits garden and agriculture land are: 3.6, 9.7, 90.4 and 81.88 respectively. It shows the low volunteer  participation rate to produce OSM data. On the other side the high accuracy achieved by deep learning leads us to complete OSM data by artificial intelligence instead of human prepared data. The advantage of using machine rather than human may be utilized in undeveloped countries or low density population regions as well as inaccessible areas.