Future of Farming

Recharch Papers Published from our Project:

1) Land Use/Land Cover Classification with Spectral Indices and Otsu Thresholding

Authors: Subhash Mukherjee, Sourav Das, Somnath Mukhopadhyay, Sunita Sarkar, Wangjam Niranjan Singh and Ajoy Kumar Khan.

Accurate land use/land cover (LULC) classification is crucial for understanding environmental dynamics, monitoring natural resources, managing urban expansion, and promoting sustainable land management practices. The availability of labeled datasets is a significant obstacle to accurate land use/land cover (LULC) classification in isolated and underrepresented areas like the Barak River Basin. This study presents an unsupervised classification on Landsat 8 satellite imagery, implementing several spectral indices to overcome the insufficiency of the label data set. For vegetation identification, the Normalized Difference Vegetation Index (NDVI), Modified NDVI (MNDVI), Green NDVI (GNDVI), and Ratio Vegetation Index (RVI) were calculated. Water body detection utilized the Normalized Difference Water Index (NDWI), Modified NDWI (MNDWI), Water Ratio Index (WRI), and Automated Extraction of Water Index (AEWI). For built-up area mapping the Normalized Difference Built-up Index (NDBI), Urban Index (UI), and Built-up Index (BI) were evaluated. Amid these, it came to light that NDVI, WRI, and BI performed best for their respective categories. Otsu's thresholding technique was applied to further process these determined indices in order to classify the binary imagery of the Barak River Basin. Notwithstanding the lack of labeled training data, the thereby generated classification output was evaluated through ground truth verification and accuracy assessment, suggesting excellent performance. Utilizing the highest-performing indices, we were able to generate the label Landsat 8 imagery using an unsupervised method. In areas with inadequate information, this technique makes it possible to develop spatiotemporal datasets for long-term environmental monitoring and land management, and it determines the prerequisites for scalable LULC mapping.

In our study NDVI, WRI, and BI proved to be the most effective indices in their respective categories. Utilizing those indices we are able to generate the final unsupervised label data shown in Figure 7. Where the red color shows the built-up area, blue color resembles all types of water bodies, green shows the vegetation area and lastly black incorporates the background.In conclusion, by implementing NDVI, WRI and BI we are able to label the data using this unsupervised proposed methodology.

Full Recharch Paper Link: Documentদলিলदस्तावेज़

2) An Unsupervised-to-Supervised Framework for Vegetation Mapping Using Spectral Indices

Authors: Subhash Mukherjee, Bishal Banik, Somnath Mukhopadhyay, Sunita Sarkar and Wangjam Niranjan Singh.

Vegetation detection from satellite imagery is essential for ecological monitoring, agricultural planning, and land use analysis. Traditional supervised methods often demand extensive labeled datasets, which are time-consuming to prepare. This study presents a hybrid unsupervised-to-supervised framework for vegetation detection using unlabelled multispectral satellite imagery. Initially, vegetation pixels are labelled using NDVI thresholding, while water regions are excluded via NDWI-based masking. These selected pixels undergo dimensionality reduction using Principal Component Analysis (PCA) and are clustered using k-means++. The clustered pixels with the highest NDVI mean are labelled as vegetation, forming labels used to train a Random Forest classifier. The trained model then classifies the entire image, producing a refined vegetation map by reapplying NDWI to exclude riverine areas. Validation through Precision, Recall, F1 Score, Accuracy, AUC-ROC, Confusion matrix, Davies–Bouldin Score, and Dunn Index confirms effective clustering, and visual outputs highlight the framework’s reliability. This method offers a scalable, label-free approach that provides an interpretable and efficient solution for vegetation mapping in complex landscapes.

This study proposes an efficient hybrid approach for vegetation detection by combining unsupervised k-means++ clustering with Random Forest classification. NDVI and NDWI indices are used to distinguish between vegetation and water regions in order to enhance discrimination near river bodies. PCA is used to reduce the dimensionality of the multispectral imagery, and k-means++ is applied for clustering. Also, the clustering quality is validated by validation metrics DB and Dunn. The results are used to train the RF model, and the classification results accuracy is validated by several validation metrics such as accuracy, precision, F1 score, recall, kappa score, ROC-AUC, and confusion matrix. The proposed framework suggests a robust and label-free method for vegetation detection. In the future, we will add different machine learning and deep learning models for better clustering and classification.

Full Recharch Paper Link: Documentদলিলदस्तावेज़

Back to Home

Recharch Papers Published from our Project: আমাদের প্রকল্প থেকে প্রকাশিত রিচার্জ পেপারস: हमारी परियोजना से प्रकाशित शोध पत्र:

Recharch Papers Published from our Project: