Implementation of random forest algorithm for crop mapping across an aridic to ustic area of Indian states

Abstract

The purpose of this study is to effectively implement random forest algorithm for crop classification of large areas and to check the classification capability of different variables. To incorporate dependency of crops in different variables namely, texture, phenological, parent material and soil, soil moisture, topographic, vegetation, and climate, 35 digital layers are prepared using different satellite data (ALOS DEM, Landsat-8, MODIS NDVI, RISAT-1, and Sentinel-1A) and climatic data (precipitation and temperature). The importance of variables is also calculated based on mean decrease in accuracy and mean decrease in Gini score. Importance and capabilities of variables for crop mapping have been discussed. Variables associated with spectral responses have shown greater importance in comparison to topographic and climate variables. The spectral range (0.85 to 0.88 μ m ) of the near-infrared band is the most useful variable with the highest scores. The topographic variable and elevation have secured the second place rank in the both scores. This indicates the importance of spectral responses as well as of topography in model development. Climate variables have not shown as much importance as others, but in association with others, they cause a decrease in the out of bag (OOB) error rate. In addition to the OOB data, a 20% independent dataset of training samples is used to evaluate RF model. Results show that RF has good capability for crop classification.