Optimization of the Results of the ML-Based GMDH Algorithm in order to Increase the Accuracy of Walnut Pollen Detection and Horizontal Optical Depth through the TLBO Algorithm

  As one of the most essential needs of living beings, clean air quality has been threatened by natural and human activities. In recent years, dust storms have been increasing spatially and temporally, causing numerous damages to social, economic, and environmental health for the residents of the southern and southwestern regions of Iran. In the present study, MODIS sensor data were used to investigate dust storms and detect horizontal optical depth.

The  advantages of MODIS sensor data include high spectral and temporal resolution. Additionally, meteorological station data were collected based on the study period. After preprocessing the data and preparing field observations, the necessary features for modeling were extracted using the differential method between selected bands of each MODIS sensor image, along with features extracted from ground-based meteorological station sensors. After further investigations and evaluations and using the viewpoints of meteorological experts, 36 differential features from various MODIS image bands and six features from ground-based meteorological station data, totaling 42 features, were extracted. Subsequently, using feature selection techniques, the best features were identified. A novel method named ML-Based GMDH, which improves the GMDH neural network by altering partial functions with machine learning models, was employed to detect dust concentration and horizontal optical depth. To achieve optimal accuracy, the hyper-parameters of this model were heuristically tuned using the TLBO optimization algorithm. Additionally, machine learning methods such as Basic GMDH, SVM, MLP, MLR, RF, and their ensemble models were implemented to compare with the main approach. According to the results, the TLBO-tuned ML-Based GMDH method provided superior accuracy in detecting dust concentration compared to the aforementioned machine-learning methods.

The SVM-PSO method was selected as the best method in the feature selection phase, the RF method was chosen as the best method among basic classification methods, and the Ensemble SVM and Ensemble RF methods were selected as the best methods in the ensemble and classification phase. It was also observed that using the ensemble approach led to a desirable improvement in horizontal optical depth classification. In the second approach, a method titled ML-Based GMDH, which improves the GMDH neural network by altering partial functions with machine learning algorithms, was used for estimating dust concentration. Additionally, to achieve suitable accuracy, the hyper-parameters of this model were finely tuned using the TLBO optimization algorithm. The results showed that this method provided appropriate accuracy in estimating dust concentration and horizontal optical depth, out performing the best-selected methods from the first approach