Comparison of Detailed Land Cover Mapping Based on SNI Classification Scheme using Conventional and Machine Learning Multispectral Classification

Abstract

Accurate and efficient land cover mapping remains a critical challenge in remote sensing, particularly when adhering to standardized classification schemes. In Indonesia, the Standar Nasional Indonesia (SNI) 7645-1:2014 classification scheme serves as the national standard for land cover mapping but is inherently complex, having been designed for visual interpretation, rather a digital classification approach. This study attempted to applied the SNI classification scheme for digital image classification by comparing the result of two commonly used algorithm in digital image classification, namely Maximum Likelihood (ML) and Random Decision Forest (RDF). Based on preliminary land cover map and fieldwork conducted, we successfully identified 18 land cover classes based on SNI classification scheme with the scale of 1:250.000. Because single date imagery and spectral information were only used, modification to the classification is conducted by grouping the classes into a new class based on its basis land cover type. Due to various spectral response of classes, the imagery is classified into 29 classes which later will be combine in the post processing process. Aside from the percent correct, the quantity disagreement and allocation disagreement were also used for assessing the result of both algorithms. Accuracy assessment of the land cover maps shows 81.17% overall accuracy for ML algorithm and 82.92% for the RDF algorithm. The ML algorithm shows higher allocation disagreement, compared to RDF algorithm, indicating higher miss-classification were detected in the ML algorithm. On the other hand, RDF algorithm shows higher quantity disagreement, indicating that this algorithm overestimated the land cover produced. This approach served as an alternative approach for mapping land cover using the SNI classification scheme, aside from the established approached.