Spectral Indices based Land Cover Classification using Deep Learning

In this paper, we use Landsat 8 and 9 satellite data to predict the land area that is suitable for agriculture and farming. Early identification and deriving insights from areas and their land properties will give us the scope for better utilization of the area. To achieve this, We used 2 manually created datasets using google earth engine. Even though the main motive is to predict the productive cropland using the created dataset. classification task we intended is to identify the type of region in the given area of land as Water, Barren land, cropland, forests, and urban areas. Deep feed forward neural network and 1D CNN models are used for this classification. The DFNN model consistently outperformed the 1D CNN across all datasets, showing superior classification accuracy and overall performance. On both the KAPLCU, MLCU, and Hybrid datasets, DFNN demonstrated better precision, recall, and F1 scores, confirming its effectiveness in classifying land regions based on satellite imagery. Future work could involve exploring land cover classification using government datasets and developing labeled repositories from unlabeled satellite images to further expand research potential in this domain.