Towards operational radar-only crop type classification: comparison of a traditional decision tree with a random forest classifier

The potential of a random forest (RF) classifier for radar-only crop classifications was evaluated for an eastern and western Canadian site. Overall classification accuracies were improved by approximately 4%-5% over traditional boosted decision trees with gains of up to 7% in the accuracies of specific classes. Accuracies above 85% were obtained for key crops including canola, soybeans, corn, and wheat. Variable importance measures generated by the RF classifier showed that the most important acquisitions occurred in late August to early September at peak biomass and after wheat harvest. The least important images were acquired in May and mid-July. The HV and VV polarizations had the most significant contributions, while the HH polarization contributed little throughout the season, except in late September when the HH response was largely driven by soil conditions. The sensitivity of three RF parameters (number of training pixels, number of trees, and number of variables to select from at each split) was evaluated and shown to have negligible influence on overall accuracy. The RF classifier provided large performance gains in terms of processing time relative to the decision tree classifier. The operational potential and implementation considerations for radar-only Canada-wide crop type mapping are discussed in the context of these results.