Improving Land Cover Mapping in Riverine Environments with Machine Learning and Spectral Indices

Abstract. Accurate land cover classification in riverine environments is essential for understanding hydrological dynamics, ecological health, and resource management. This study utilizes high-resolution LISS-4 imagery from the IRS-R2 satellite to address the complex classification challenges in the dynamic confluence of the Mahanadi and Shivnath Rivers. To enhance classification accuracy, two distinct feature sets were developed: the first feature set (FS-1) contained the original three spectral bands of the imagery, and the second feature set (FS-2) added eight derived spectral indices to improve class separation. Three machine learning classifiers, Random Forest (RF), Support Vector Machine (SVM), and Gradient Tree Boosting (GTB), were applied to assess the effectiveness of each feature set. The results demonstrated that the combination of FS-2 and RF yielded the most accurate and interpretable classification, achieving an overall accuracy of 84.48% and a Cohen’s Kappa coefficient of 0.81. Visual results demonstrated precise delineation of narrow water channels and sandbars, when using the enhanced feature set FS-2 with the Random Forest (RF) classifier, the model achieved an F1-score of 86.96% for dense vegetation and 83.78% for water. FS-2 consistently improved F1-scores across all classifiers, aiding in distinguishing visually similar classes like wet sand, dry sand, and sparse vegetation. These findings highlight the value of combining spectral indices with high-resolution imagery to achieve accurate land cover classification in complex landscapes, supporting applications in ecological monitoring, flood risk assessment, and resource management for riverine ecosystems.