Determination of Rainfall Rates from GOES Satellite Images by a Pattern Recognition Technique

Abstract

Radiances from clouds observed in visible and infrared images obtained from the SMS-2, GOES-2, and GOES-4 satellites have been used to estimate rainfall by means of a pattern recognition algorithm that was applied to single images. The algorithm classified rain into three classes: 0—no rain (0 ≤ R −1); 1—light rain (0.5 ≤ R −1); and 2—heavy rain (5.0 mm h⁻¹≤ R). The rainfall rates used in the training set and those used to test the algorithm were derived from a set of twenty-nine Plan Position Indicator (PPI) displays obtained from NOAA operational radars. Data were derived from summer storms, tropical storms and cyclones. Rainfall from precipitating clouds was classified by a pattern recognition technique that used textural and radiance features in a hierarchic decision tree. The analysis was applied to regions 20 × 20 km in area that were measured in the visible spectral region with 1 × 1 km and 2 × 2 km resolution and in the infrared with 4 × 8 km resolution. The radiance features used in this analysis were the radiance maxima, minima, and the means. The textural features that were used included the edge strengths per unit area and the maxima and means of the mean, contrast, angular second moment, and entropy in four directions. Of the arm sampled in this study, approximately one-third were in classes 0, one-half were in class 1 and one-sixth were in class 2. Case studies that employed data from both the visible and infrared sensors correctly identified rainfall classes 0 and (1 + 2) in about 96% of the cases and identification into classes 1 and 2 was correct in about 70% of the cases studied. The corresponding skill scores were ∼80 and 60% respectively. Data derived only from infrared images yielded correct identification of 0 and (1 + 2) classes in 85% of the cases and identification of classes 0, 1 and 2 was correct in 65% of the cases. The corresponding skill scores were ∼65% and 40% respectively.