Performance of Different U-Net Architectures for Inventory of Coconut Plantations Using Cartosat-2 Multispectral Data

Over the last decade, the horticulture sector has become one of the important driving forces for the rapid development of agriculture in India. Geospatial technology is being operationally used in India for mapping and monitoring horticultural plantations. In general, per pixel or OBIA techniques have been used for mapping while use of deep learning techniques for analysis of high-resolution data for these plantation crops is limited. In the present study, seven deep learning models have been attempted for mapping of coconut plantations to overcome some of the challenges faced by conventional classifiers. U-Net and Siamese architectures are used to develop hybrid models for improved classification. These hybrid models have been implemented with batch normalization and the results indicated that the developed hybrid models showed reasonable IoU score in comparison to other DL models with better than 90.0 per cent classification accuracy for mapping coconut plantations. The trained DL model has been tested in different geographic regions for achieving reasonable accuracy for mapping of coconut plantations. The present study demonstrated the use of DL methods for mapping of coconut plantations using Cartosat-2 MX data within acceptable accuracy and trained DL model could be developed with larger no. of labelled training samples for operational applications.