Merging Random Forest Classification with an Object-oriented Approach for Analysis of Agricultural Lands

Machine learning algorithms recently have made major advances, with decision tree classifiers gaining wide acceptance. Boosting and bagging of decision trees have added to the predictive capabilities of these approaches. Object-oriented (O-O) analyses have been developed during this same period, offering important improvements in classification over pixel-based approaches under certain conditions. Classification algorithms for O-O approaches, however, have been fairly limited and generally have not incorporated new statistical approaches used for pixel-based classifications. One of the most promising new classification algorithms is Random Forest (Breiman-Cutler) classification (RF). We incorporated RF into an O-O classification of Landsat-based imagery for mapping agricultural lands in north-central Montana, USA. The Definiens multi-resolution segmentation algorithm was used to generate fieldbased objects. RF was used to classify land management (tillage, conservation reserve, crop/fallow) based on reference data from >400 field sites. Object-based attributes included factors such as average spectral response, spectral variability, texture, and shape characteristics. Accuracy was assessed using “out-of-bag” estimates in RF. This classification approach was able to efficiently and accurately merge RF with an object-oriented approach for improved classifications.