Sugarcane Crop Classification from Satellite Imagery using Deep Learning Approach

Abstract

Sugarcane holds great importance as a crop in Pakistan, being one of the top global producers. Precise mapping of sugarcane fields is vital for effectively monitoring their size, production, and evaluating their influence on society, the economy, and the environment. This study focuses on presenting a deep learning-based framework that employs pixel-based classification to identify sugarcane crops in Pakistan, along with other commonly grown crops, using Sentinel 2A multi-spectral imagery. The framework encompasses various stages, including the selection of Sentinel products, preprocessing, extraction of spectral indices, a compilation of spectral features, labeling through spectral unmixing, and crop classification. The selection of Sentinel products for each crop field is based on the NDVI values. Multiple spectral and biophysical indices are derived from these Sentinel products. Each pixel is compiled into an image, representing the temporal evolution of that pixel across spectral bands and indices. These synthetic images are resampled using bilinear sampling and fed into various deep-learning models to solve the classification problem. Additionally, linear spectral unmixing is employed to assign labels to each pixel, ensuring accurate identification of the predominant crop in that pixel. The dataset used in this study comprises samples from different districts in Pakistan. Two combinations of datasets are created to assess the robustness of the developed methodology: one for training and testing within the same district and another for training and testing across separate districts. For the first dataset combination, most of the tested classification models achieve a high accuracy of approximately 99%. In the second dataset combination, LSTM outperforms other models, achieving an accuracy of 90%. The classified pixels are then integrated into the classification map, accurately representing their geolocation within the corresponding field. The proposed framework demonstrates promising results compared to the convNext model and exhibits the potential to effectively classify ready for harvest sugarcane among other crops using a limited number of products. Furthermore, it proves capable of classifying sugarcane across different districts.