Tillage classification with random forest algorithm using high-resolution satellite data
To minimize soil disturbance and improve yields, transition from the measures of intensive tillage to reduced or zero tillage is becoming prevalent throughout the world. Zero-tillage (ZT) practices have become popular among farmers in different countries and regions. Many recent studies focused on the tillage detection utilizing various techniques, such as satellite remote sensing, yet most of them were not focusing on separating ZT and conventional tillage (CT) in a large spatial scale. In this study, we analyzed the classification performance of Sentinel-1, Sentinel-2, and Landsat satellites using random forest algorithm for the large-scale tillage practices at Guanajuato, Mexico in 2017. The research aims to 1) accurately classify agricultural fields with zero tillage measures in Guanajuato; 2) find out classification accuracies among different composite methods for different sensors individually, and for sensor combinations to evaluate the improvement of classification accuracies; 3) analyze the phenology features of the study area, and compare the performance across different composites and sensors; and 4) figure out important features for random forest classification. We had a classification accuracy up to 86.46% for the multi-month sowing plus peak season composite with full sensor combinations. Furthermore, we found that the classification accuracy (up to 82.27%) when using 30-day composite only in sowing season was still reliable if compared to a multi-month composite. The results illustrate that it is reliable to apply combined high-resolution satellite data with different compositing methods to classify zero-tillage agricultural fields in Guanajuato, Mexico.
Haoyu Wang, MS (CE, EI)