Sugarcane Identification using Linear Time Weighted Cosine regression and One-class Support Vector Machine

Abstract

In-season crop classification through remote sensing time series data has proven effective in classifying in-season crop types during their specific growing periods. However, confusing patterns and spectral data result when confronted with crops such as sugarcane, exhibiting varying growth seasons. This issue was further compounded by the Cropland Data Layer (CDL) created by the United States Department of Agriculture (USDA), which only provided post-season crop maps that are often published too late to be practically useful. To address these challenges, this work aimed to construct a transfer learning workflow, utilizing Florida sugarcane samples to identify other growing sugarcane such as Louisiana. The training dataset used incorporated burning sugarcane samples from Palm Beach County, Florida, and Sentinel-2 satellite time series NDVI data. A Linear Time Weighted Cosine (LTWC) regression model was then applied in order to transfer the NDVI series into regression curves, where the coefficients represented sugarcane phenological features. Subsequently, a One-class Support Vector Machine (SVM) was trained using the coefficient dataset to classify sugarcane in Florida and Louisiana. The results demonstrated accuracies comparable to those of the CDL but with the significant advantage of being published at least six months earlier. Furthermore, with the fine-tuning of SVM kernel parameters, the model’s application can then be applied to other countries such as Brazil. Validation of the results will be performed using ground truth samples obtained from both photographs and historical images from Google Earth. Overall, this study presents an option to address the challenges posed by variable growth seasons in crop classification through remote sensing and machine learning, with potential applications in other sugarcane regions worldwide.