machine learning algorithms

Flood zoning maps provide valuable information about the nature of flood and its effect on floodplain lands. A set of effective factors must be defined to map flood susceptibility or, in general, to develop a model for assessing natural disaster risk. The factors affecting flood were used in eastern catchment area of lake Urmia which include altitude, slope, distance from the river, topographic moisture index, topographic position index, roughness index, curvature level, topographic curvature section, total curvature, NDVI index, land use, lithology and rainfall according to the experiences of experts and researchers reported in previous studies. After preparing the effective layers on the flood and the point layer of the flood points, as well as performing the linear test, five methods including Multiple Linear Regression, Partial Least Squares Model, Quantile Regression, Ridge Regression and Robust Regression were used for modeling and predictions. Then the ROC curve was used to validate the results. The results of this validation showed that the partial least squares (PLS) and multiple linear regression (MLR) models with the maximum area under the curve (AUC) (0.983 and 0.997, respectively) and the lowest standard deviation (0.015 and 0.018, respectively) ) have performed better. Among these two models, PLS has slightly better results than MLR. Finally, a random forest model was used to determine the importance of the input factors, and it was found that the factors of height, distance from the waterway and slope percentage are the most influential factors on floods in the study area.

Rice is an important crop and the main food of more than half of the world’s population, which needs water and heat to grow. Thus, mapping and monitoring rice fields with efficient means such as remote sensing technology is necessary for food security and the lack of water sources. The phenology extracted from the time series of vegetation indices is used for monitoring and mapping the area under rice cultivation. In addition to the phenological curve, the LST time series map, which is calculated from Landsat 8 images and is related to the phenomenon of evaporation and transpiration of irrigated crops, can cause the separation of rice cultivation from rainfed crops, summer crops, water, etc. Therefore, in this study, the effect of the LST time series map is investigated map for improving the accuracy of rice field identification.

Since the planting to harvest period of rice is from May to October, in this study, the time-series maps of LST and NDVI for the 3rd of April, 21st of May, 6th of John, 22nd of John, 8th of July, 24th of July, 9th of August, 12th of October, and 28th of October have been calculated after download the Landsat-8 time-series in 2020 The ground truth map of the study area has been obtained from the US Department of Agriculture. To identify rice fields and calculate the LST and NDVI using the Landsat-8 images, initial pre-processing including radiometric and geometric corrections has been applied to these images first. After initial corrections and the calculation of NDVI and LST maps, to identify rice fields in the study area, machine learning algorithms such as Support Vector Machine, K-Nearest Neighborhood, Multilayer Perspective, Logistic Regression, and Decision Tree, have been proposed.

The results of the proposed method at the state of California showed that using the time series map of  Land surface temperature (LST) with the time-series map of  Normalized Difference vegetation Index, improved the results of identifying rice fields (the average Overall Accuracy= + 3/572% and the average kappa coefficient= +7/112%). Visual results showed that some cultivation such as tomato, corn, cucumber, fallow, and water were removed from the rice final map when using the LST time-series map with the NDVI time-series map. According to the numerical results, the Support Vector Machine algorithm (Overall Accuracy 94/28 and Kappa Coefficient 88/29), the Multilayer Perceptron algorithm  (Overall Accuracy 94/26 and Kappa Coefficient 88/21), and the K-Nearest Neighborhood algorithm (Overall Accuracy 93/71 and Kappa Coefficient 87/08) showed the highest Overall Accuracy and Kappa Coefficient compared to the Logistic Regression algorithm (overall accuracy 91/96 and kappa coefficient 83/54) and the Decision Tree algorithm (Overall Accuracy 91/34 and Kappa Coefficient 81/97), respectively.

Although, many methods have been proposed to identify rice fields from satellite images. But, the similarity of rice class with other classes is one of the main challenges related to rice identification. In this research, the effect of LST time series maps to improve the identification accuracy of rice fields in Landsat-8 time-series images was investigated. In this study, the effect of the time series map of land surface temperature index extracted from Landsat-8 images on improving the accuracy of identifying rice fields from other rice fields due to the evapotranspiration process using machine learning algorithms was investigated. The results showed the effectiveness of the proposed index in improving the identification accuracy of rice fields. One of the reasons for improving the accuracy of identifying rice fields is to extract the characteristics of the thermal growing season from the Earth's surface temperature time series (LST) maps along with the rice phenology curve. The results showed that due to the flooding of rice fields when using the NDVI time series map, water class and fields summer crops were identified as rice class. But, water and summer crops classes were removed from the rice final map using a land surface temperature time-series map with the extraction of thermal growth season characteristics. Therefore, the results showed that there was a direct relationship between LST time-series maps and rice cultivation.

Twin Sentinel-2 MSI sensors are spatially like the Landsat-8 OLI super spectral instrument, aiming to additional data continuity for land surface monitoring were launched by European Space Agency. In this paper, the potential of these data was evaluated for discrimination of lithological units and alterations in the Esfordi phosphate deposit area and was compared with OLI and fused OLI data. Decorrelation stretch method was used for enhancing the lithological units of the study area, and all of the 3 datasets acceptably discriminated the rock units. Among these, MSI data could produce the lithological map with high resolution and highest level of reality owing to its high spatial resolution. For statistical comparison, Support Vector Machine and Random Forest methods were applied on datasets for classification of the lithological units and their accuracy was assessed using confusion matrices. Furthermore, the corresponding band ratios to which were defined for Landsat-5 TM, were applied on datasets for detecting the altered areas. Then the areas of each highlighted alteration zones were estimated for comparison. Furthermore, the scatterplots of band ratio images were prepared. MSI dataset revealed the highest overall accuracy and Kappa coefficient in Support Vector Machine and Random Forest classification. Also, the results of band ratioing showed that MSI and fused OLI data have the most correlation and similarities. This study demonstrated that MSI data are more optimal than OLI data for lithological and alteration mapping. Also, using fused OLI data in dates which there is no MSI data acquisition or for producing seamless geological maps in continental scale besides to MSI data, is efficient.

It is necessary to know about the quantity of urban tree canopy cover due to its role in air and noise pollution reduction, wind prevention, saving rain water, and runoff control. Being expensive and time consuming, the manual extraction of tree canopy has been replaced by remote sensing techniques conducted on the images, digitally. There are several parameters which must be optimized prior to use of object-oriented classification. One of these parameters is Scale affecting the segmentation results, significantly. Scale is usually set by trial and error which is an experimental approach. One of the aims of this study is to optimize Scale parameter, automatically. In addition, after segmentation process based on a proper Scale, it is required to classify the identified segments based on the attributes which are extracted from these segments. In this stage, the selection of suitable classification method fed by the proper attributes is critical. In this research, LiDAR data and aerial image acquired on Vaihingen, Germany, were utilized for segmenting the urban area. In order to identify suitable attributes, random forest feature selection was applied on the attributes derived from the identified segments. Machine learning methods including support vector machine, random forest, and decision tree were compared for classifying the segments based on their suitable attributes into two classes including tree canopy cover and others. The results indicated that Scale of 25 is the best one to segment this area. Also, the tree canopy cover map derived from support vector machine with quality index of 79.90 showed the best performance among different classifiers used in this study.