Cotton Crop Classification using Optical and Microwave Remote Sensing Datasets in Google Earth Engine

Abstract. Cotton maps for Telangana which is one of the cotton producing region in India were produced through supervised classification in Google Earth Engine. To make well-informed decisions, farmers, governments, scientists, and agricultural organizations need accurate information on crop prediction. However, automated crop type mapping remains challenging due to the limited availability of field-level crop labels required to train supervised classification models. Cotton mapping was made more accurate and efficient by using a two-step mapping approach, which consists of mapping the cropland and then extracting the cotton crop for areas with more heterogeneity, this framework increased the accuracy from 83% to 91%. For a more accurate estimate of the cotton crop, this study combined high resolution Sentinel-1 and Sentinel-2 data with several secondary data types in the SMILE Random Forest (RF) model at various stages of the crop growth season. For that First, cropland/non-cropland area were predicted to extract features from time series. Next, cotton crops through RF classifiers were applied on median composites of Sentinel-1 and Sentinel-2 data for each pixel in the region. Furthermore, spectral, structural and phenological feature time-series satellite data were merged and processed into a supervised random forest classifier. The classification of cotton, cropland and non-cropland model produced with producers accuracy of 98%, 88% and 90%. Through experiments, we also discovered that employing time-series imagery generates substantially higher classification results than single-period images. The inclusion of shortwave infrared bands, followed by the addition of red-edge bands, can increase crop classification accuracy more than using simply traditional bands like the visible and near-infrared bands. Incorporating common vegetation indices and Sentinel-2 data, combining with Sentinel-1 reflectance bands improved the overall crop classification accuracy by 0.2% and 0.6%, respectively. This study demonstrates how combining optical and microwave remote sensing data, the GEE platform, transfer learning, and cotton cropland mapping algorithms can enhance insights into precision agricultural systems.