Pixel-based crop classification in Peru from Landsat 7 ETM+ images using a random forest model

Crop classification within large agricultural regions is challenging owing to the presence of crops with similar phenological variation and intra-class variability. The development of efficient and simple classification methods is needed for more accurate mapping, monitoring, and analysis of land-use categories. Multi-seasonal aggregated statistical variables of Tasseled-Cap (TC) bands (brightness (B), greenness (G), and wetness (W)) obtained from the Landsat 7 Enhanced Thematic Mapper Plus satellite (Landsat 7 ETM+) covering cropped areas in the catchments of the Ica and Grande Rivers of Peru were evaluated to assess the performance of random forest (RF) classifiers in identifying crop type. The effects of various TC band combinations on the classification results were also examined. Seventeen crops (asparagus, cotton, grape, maize, mango, and so on) were included. Overall accuracy and kappa coefficient analyses showed that the three-band combination of B-G-W, using multi-seasonal data, led to more accurate classification than did other combinations, yielding values of 86% and 0.81, respectively. The results indicate that employing aggregated statistical variables of TC bands in conjunction with RF classification techniques by using freely available multi-temporal satellite image data is not only a useful but also more economical and computationally efficient method for crop classification than the current one.