Satellite aerosol optical depth prediction using data mining of climate parameters

Tropospheric aerosol particles play an important role in the Earth's radiative energy balance directly by scattering and absorbing solar radiation and indirectly by modulating the microphysical and radiative properties of clouds. Aerosol optical depth (AOD) based on satellite remote sensing data is a quantitative estimate of the amount of aerosol in the atmosphere and can be used as an indicator of aerosol particle concentration. In general, the review of previous studies indicates the high importance of aerosol products based on satellite remote sensing data in modeling the spatial-temporal patterns of dust storms and in particular the identification of dust sources. The advantages of using satellite AOD to identifying dust events are possible in arid areas with relatively little cloud cover. The presence of clouds in the sky also severely limits AOD terrestrial and satellite measurements. Thus, AOD datasets sometimes have a gap due to factors such as cloudiness. Since the possibility of monitoring and measuring aerosols in cloudy conditions is limited, the use of proxy datasets to fill the gap will be an advantage. In this regard, several studies based on the analysis of satellite data have emphasized the association between climatic parameters and dust events (specifically AOD) in different regions. Therefore, considering the relationship between climatic parameters and AOD, these parameters can be used as a proxy data set to estimate AOD values for areas without data or with cloud cover. Also, using the predicted values of climatic parameters, AOD values can be predicted. Accordingly, in order to achieve reliable AOD prediction results, it is necessary to use a generalizable approach that can model the complex relationships between large data sets and satisfactorily solve the mentioned problems. For this purpose, one of the efficient data mining algorithms called M5P was considered to analyze and extract the relationships between climatic parameters and AOD to obtain predictive models. The M5P algorithm is a combination of tree and regression models with capabilities such as high prediction accuracy and ease of interpreting results.

In this study, in order to derive AOD predictive models based on climatic parameters, M5P data mining algorithm based on tree structure and multivariate linear regression analysis were used. Accordingly, a spatial database of remote sensing time series data related to 4 climatic parameters (as independent variables) including surface air temperature (SAT), precipitation (P), surface relative humidity (SRH) and wind speed (WS), and AOD (as dependent variable) was generated. WEKA software was used to implement the M5P model. After analyzing the relationships between independent and dependent variables through the tree model structure and linear multivariate regression, AOD predictive rules were extracted. Statistical indicators were used to validate the linear predictive models.

After pre-processing the time series data of climatic parameters and AOD as training data set, the input independent and dependent variables of the M5P were defined. Implementation steps of the M5P algorithm, including homogenization of independent input data sets by forming decision-making trees based on a series of "if-then" rules, multivariate linear regression analysis in homogeneous classes, and finally validation of the model results was performed in WEKA software. Thus, a total of four linear models (LM) or predictive rules for estimating AOD based on the values of climatic parameters were extracted. Finally, by placing the values of climatic parameters in the obtained linear models, the AOD value can be estimated based on the thresholds defined by the M5P algorithm. The obtained linear models are able to predict AOD values in different conditions (based on climatic parameters). Validation of the results of the M5P algorithm based on correlation analysis between input variables and evaluation of prediction errors through MAE and RMSE statistics shows the acceptable performance and accuracy of linear models in relation to AOD prediction. Given the dynamics of aerosol particles (especially dust) and their ability to transportability by the wind even at very far distances from their source of emission, it is likely that the amount of measured AOD for a pixel by a satellite sensor, does not exactly belong to the same area on earth. Therefore, in relation to the prediction error of the models, it should be noted that this may be due to the ability of the aerosol particles to be carried by the wind. Due to the strong correlation between AOD and climatic parameters, possible discrepancies may be due to the mentioned reason. Because a dust storm arising from a source may have no relation with the values of the climatic parameters at the destination.

In general, in this study, the capability of M5P data mining algorithm in order to AOD prediction was evaluated. Using the M5P algorithm based on inductive learning and using remote sensing time series data, through the formation of decision trees based on the set of "if-then" rules, four linear predictive models based on climatic parameters were extracted. Predictive models were extracted and validated using a data set for Ahvaz city. AOD, as an indicator of the state of the atmospheric aerosol, has great importance for dust storms studies.Access to AOD data is restricted in some parts of the world and in some seasons due to some limitations such as cloud cover. On the other hand, it is important to be aware of future spatial-temporal patterns of dust storms in order to adopt crisis management measures. Using the obtained predictor linear models in this study, it is possible to make an acceptable estimation of AOD in some areas, there are restrictions on access to AOD. Also, by entering the predicted values of climatic parameters, it is possible to estimate the future spatial-temporal patterns of AOD.Dust storms generally occur as a function of a wide range of environmental conditions, including atmospheric properties, as well as surface parameters such as vegetation, soil moisture, and soil texture. With this background, only considering the atmospheric conditions and their impacts on the spatial-temporal patterns of AOD may sometimes not produce the desired results. Therefore, it is recommended in future studies in this field, in addition to climatic parameters, which are mostly indicators of the atmospheric condition; ground surface parameters should also be used in modeling. By doing so expected to increase the accuracy of linear models for predicting AOD.