Determining the Best Classification Algorithm in order to Estimate the Area under Date Palm Cultivation using LANDSAT 8 Satellite Imagery

Introduction Date palm is one of the most valuable horticultural products in Iran, which includes 16% of non-oil exports to the world. Kerman province has the second rank for the cultivation area of date palm in Iran. Having information about the exact cultivated area has gained importance for further decision makings. To determine the cultivated area, organizations usually use census which has the disadvantages of high cost, wasting time and labor intensive. The aim of this research was to study the feasibility of using Landsat 8 OLI images to identify and classify the area under date palm cultivation. To accomplish this purpose, four supervised classification methods were evaluated. Materials and Methods The study area was in Bam region located at 200 km southeast of Kerman province. In this research, a total of 14 images of Landsat8 OLI satellite from the study area during fall and winter were downloaded from Landsat official web page. After preliminary inspections for interested classes (Date palm gardens, Lands covered with bare soil and forage crop fields), one of the images that was taken on Jan 14, 2017, was selected for further analysis. After initial corrections and processing, 32 images of alfalfa farms, 32 images of date palm gardens and 32 images of lands covered with bare soil, were selected using GPS data points collected in study area scouting. Shape files of all selected fields were created and utilized for supervised classification training set. The same process was also done for the unsupervised classification method.  To evaluate the classification methods confusion matrix and Kappa coefficient were used to determine the true and miss-classified area under date palm cultivation. It is worth mentioning that these factors alone cannot identify the most powerful method for classification and they just give us a general overview to choose acceptable methods among all available methods. To identify the most powerful method among selected methods, confusion matrix and investigating the pixel transfers between classes is the crucial method. Results and Discussion Results of classifications revealed that the overall classification accuracy by using NN, MLC, SVM, MDC, and K-Means were 99.10% (kappa 0.973), 98.77% (kappa 0.975), 98.66% (kappa 0.973), 98.52% (kappa 0.97), and 52.66% (kappa 0.31) respectively. Concerning the confusion matrix in the NN method, the percentage of producer accuracy error in date palm class was 0% and in user, accuracy error was 1.44%. In the review of other methods, the lowest producer accuracy error value in date palm class obtained by NN and SVM methods was 0% and the highest producer accuracy error belonged to MLC method which was 1.35%. Checking the recognition power of other classes showed that in the soil class, the highest producer accuracy error was 2.32% by MDC method and the least one was 0.64% by MLC. For forage class, the highest producer accuracy error was calculated 37.07% by SVM and the least accurate one was 4.92% by MDC. Although the K-Means method with Kappa Coefficient of 0.31 did not have a good classification quality, concerning classes and samples, it successfully could identify date palm according to selective samples with 100% accuracy. Results of calculated date palm area using supervised classification methods versus actual area measurements showed that NN and SVM methods with the coefficient of determination (R2) of 0.9995% and 0.9986% had the highest coefficients. K-Means method with R-square of 0.9228% had a good correlation. In general, all supervised classification methods obtained acceptable results for distinguishing between date palm classes and two other classes. NN and SVM methods could successfully recognize date palm class. K-Means method also could recognize date palm class but the recognition included some errors such as dark clay soil textures which were classified as the date palm. Conclusions In general, overall accuracy and kappa Coefficient alone cannot identify the most powerful method for classifying and these methods just give us a general overview to choose an acceptable percentage of accuracy coefficients among available methods. After the initial selection, to identify the most powerful method of classification the pixel transfer calculations in a confusion matrix would be an acceptable technique.