A Multi-Sensor Approach to Classify Tillage Practices in Mexico

Abstract

To minimize soil disturbance, there has been an increased adoption of reduced or zero tillage (ZT) technologies among farmers in different regions across the globe. Yet, to date, the scale of adoption remains unclear because it is difficult to collect adoption data on-the-ground at large spatial and temporal scales. Remote sensing can offer a way to map such technology adoption at large scales and at low cost. This study uses Sentinel-2, Landsat 7 & 8, and Sentinel-1 satellites to map tillage practices in Guanajuato, Mexico, a region where the use of zero-tillage has been promoted by national and international agencies over the last decade. We specifically compared accuracy scores of different sensors and sensor combinations, and different timing of imagery in a random forest classification. The results indicate that when considering the accuracy of a single sensor, Sentinel-2 has the highest classification accuracy. However, using a combination of all three sensors dramatically outperformed all single sensor analyses, with an overall classification accuracy of 85.96%. Considering image timing, we find that using imagery from only the sowing season performs almost as well as using imagery throughout the growing season. We conclude that using freely-available satellite images is effective in classifying tillage practices in Mexico at large spatio-temporal scales.

Keywords: zero tillage; Mexico; optical and SAR sensors; Google Earth Engine; random forest