Application of Spectral Index-Based Logistic Regression to Detect Inland Water in the Caucasus

Abstract

The Caucasus is a diverse region with many climate zones that range from subtropical lowlands to mountainous alpine areas. The Caucasus region also contains irrigated agricultural lands, heavily vegetated wetlands, reservoirs, and canals in various states of repair. This study evaluates logistic regression models based on multiple optical water indices to establish the best-performing water detection method (MNDWI) for three Landsat path/row tiles and introduces the optimal probability cut-off method (OPC). The OPC logistic regression model is applied to the Caucasus region covering 19 Landsat tiles from May to October 2019. The global water product from the European Commission Joint Research Center (ECJRC) (1) was used to generate training and validation data points by stratified random sampling. We used 6745 manually classified points (3261 non-water, 3484 water) to validate the OPC and ECJRC max extent water products' performance using an estimated proportion of area error matrix to evaluate accuracy. This approach produced Max extent water maps with higher accuracy (89.2%) and detected 392 km2 more water than the ECJRC Max extent product (86.7%) for the entire year. Despite high levels of overall accuracy, the error matrix is evaluated using the proportion of area to create user’s and producer’s accuracy by class (2); we witness a significant loss in performance in detecting accuracy for the water class (15.6% OPC, 11.5% ECJRC). Evaluation of the accuracy of land cover products solely by overall accuracy can be misleading to the user, emphasizing the need for a more rigorous assessment of water detection performance metrics beyond the traditional methods of evaluating accuracy.