Crop Phenology Retrieval from Polarimetric Decomposition and Random Forest Algorithm During Smapvex16-Mb Campaign

Abstract

The objective of this study is to investigate the retrieval of crop phenology using polarimetric decompositions and Random Forest (RF) algorithms. To realize this objective, we used multi-temporal RADARSAT-2 data and ground measured vegetation characteristics acquired during the SMAPVEX16-MB (Soil Moisture Active Passive Validation Experiment 2016 in Manitoba) campaign in Canada. Polarimetric parameters with the potential to quantify the volume scattering mechanism were extracted, and then analyzed with respect to ground identified phenology for different crop types. The RF algorithm was subsequently trained based on 60% of the data, and validated using the remaining data. Results show that the crop phenology can be monitored, through the combination of multiple polarimetric parameters to build different decision trees in the RF algorithm. By averaging the estimates from multiple decision trees, the complex relative patterns between the polarimetric parameters and crop phenology were recognized, leading to appropriate estimations on crop phenology. The obtained spearman correlation coefficients between the retrieved and ground identified crop phenology were 0.94, 0.91, 0.81 and 0.89 for canola, corn, soybean and wheat, respectively. This study also suggests suitable polarimetric parameters for a timely monitoring of crop phenology.