Comparative Assessment of Object-based and Pixel-based Approaches for Crop Cover Classification

Introduction/Importance of Study:Accurate crop identification and classification are crucial for effective agro-based planning and ensuring food availability. Reliable classification helps optimize agricultural productivity and resource management.Novelty Statement:This study innovatively compares pixel-based and object-based approaches for machine learning-oriented classification methods to develop crop-type maps in Rahim Yar Khan, Pakistan.Material and Method:Utilizing the Google Earth Engine (GEE) cloud computing platform, pre-processing steps were applied to Synthetic Aperture Radar Sentinel-1 and Sentinel-2 data. Integration of Sentinel-1 (VV, VH) and Sentinel-2 satellite bands enabled the computation of various indices and the production of composite images for subsequent analysis. The primary objective was to evaluate the effectiveness of these approaches in classifying major crops: cotton, rice, and sugarcane. Time-specific images were employed to leverage crop seasonality; for instance, an August composite image was prioritized for cotton, while September composites were used for rice and sugarcane classification. The study utilized two object-based segmentation approaches: Simple Non-Iterative Clustering (SNIC) on the GEE platform and Object-Based Image Analysis (OBIA) using Multi-Resolution Segmentation in E-Cognition software. The Random Forest (RF) machine learning algorithm was applied to both pixel-based and object-based approaches. Field sample data, including cotton, rice, sugarcane, orchards, and other crops, were used for classification, validation, and accuracy assessment. A comparative analysis was conducted to evaluate the performance of pixel-based and object-based methods.Result and Discussion:The RF algorithm applied to pixel-based approaches using Sentinel-1 and Sentinel-2 imagery bands with composite indices demonstrated superior results. The pixel-based RF classification achieved 98% accuracy with a kappa coefficient of 92%. In comparison, RF applied to SNIC in GEE achieved 96% accuracy with a kappa coefficient of 95%, while OBIA in E-Cognition attained an accuracy of 89%.Concluding Remarks:The study concludes that tuning the segmentation parameters in both E-Cognition and SNIC algorithms can enhance the accuracy of object-based classification.