Pixel-Based Classification of Major Rice Varieties using Sentinel-2 L2A Data via CNN in Spectral and Time Domain

Abstract

Crop monitoring and management, efficient utilization of natural resources, and development of sustainable systems have an important role in keeping agriculture efficacy of crops with high productivity. However, in the area of crop monitoring, it remains a challenge to monitor crop varieties due to the lack of accurate data available on time. In Pakistan, rice crop is important both agriculturally and economically. Two major rice varieties including Basmati and IRRI were selected to classify in the present study. Satellite remote sensing offered a suitable approach as the bases for coverage, accuracy, costeffectiveness, well revisit time, and type of data which hold great potential for routine applications in crop monitoring and management. Sentinel-2A/B on board Multispectral Instrument (MSI) delivers data with a superior spectral, spatial, and temporal resolution, makes them suitable for crop monitoring. Deep learning techniques such as convolutional neural networks (CNN) in agriculture outperforms the existing approaches by extract the distinguishable and representative features of different land cover from remote sensing images in a hierarchical way to classify. In the present study, a pixel-based deep CNN is devised in both spectral and time domains to map rice varieties in Punjab province, Pakistan. We proposed a spectral unmixing based feature extraction strategy, which provides sub-pixel level information and determines the pixel’s abundance map before supervised classification. Abundance representation allows identifying refined structures that are not available in the ground truth map. Vegetation and water indices are also adopted over the full rice-growing season to increase the classification accuracy. Experimental results exhibited an excellent overall accuracy of 98.6 % with the proposed approach.