Cloud detection in MODIS satellite images using ensemble methods

Cloud detection is one of the essential pre-processing steps in optical remote sensing. In this article two kinds of ensemble learning methods, boosting and random forest were applied for cloud detection and distinguishing cloudy pixels from spectrally similar classes, snow and ice. MODIS L1-B data were used as input and MODIS cloud mask (MOD35) as a reference for accuracy assessment. Bands 1, 2, 8 and 26 were selected as non-thermal bands, also normalized difference vegetation index (NDVI) and normalized difference snow index (NDSI) indexes were computed. Bands 20, 22, 31, 32 and 35 and also some of the brightness temperature differences applied in MOD35, like, BT11-BT12 and BT13.3-BT11 were considered as thermal bands. Spectral (reflectance and temperature) and gray level co-occurrence matrix (GLCM) textural features were extracted from non-thermal (1, 2, 8 and 26) and thermal (20, 22, 31, 32 and 35) bands. 4 kinds of boosting methods including, Adaboost.M1,AdaboostSVM, Logit boost and Total boost and random forest were used for cloud detection. Nonthermal (visible and infrared) and thermal classifies were fused using majority vote in all previously mentioned methods except Logit boost. The performance of AdaboostSVM was compared to other methods based on cloud, snow/ice and cirrus detection using McNemar statistical test. Also, the agreement of cloud, snow/ice and cirrus detection of ensemble learning methods with three kinds of tests applied in MOD35, brightness temperature (BT), brightness temperature difference (BTD) and visible near infrared (VNIR) reflectance, was computed. All ensemble learning methods showedhigher agreement with non-thermal classifiers to MOD35 tests than thermal ones. Only fusion of nonthermal and thermal classifiers in random forest (RF) method was efficient and after fusion, the agreement with all three kinds of MOD35 tests mentioned before increased to above 88%.

کلیدواژه ها:

Cloud ، MODIS ، ensemble learning methods ، VNIR and thermal classifiers