Abstract:
Crop classification plays a vital role in crop status monitoring, crop area estimation, and food production. Remote sensing
data is widely accepted for crop classification at remote locations. However, crop classification is challenging due to spectral and spatial
similarities, complex land structures, temporal inconsistencies, and environmental parameters. Machine learning models must be robust,
particularly when dealing with a variety of crop types and changing environmental factors. This study examines the extent to which
various algorithms generalize, emphasizing the importance of adaptability to various farming systems and geographical conditions. It
explores transfer learning and ensemble approaches as possible ways to improve the resilience and flexibility of the model. In the present
study, an effort has been made to identify, classify, and map multiple crops from the complex environment using the Sentinel-2 dataset
and advanced machine learning methods such as random forest, Spectral Angle Mapper (SAM), Maximum Likelihood Classifier (MLC),
K-means clustering and Iterative Self-Organizing Data Analysis Technique (ISODATA). The crop spectral features were identified using
the Normalized Difference Vegetation Index (NDVI). The NDVI outcomes ranged between -0.91 and 0.54, which were then used to
identify crop areas. Ground reference data, Google map, and Google Earth data were used to determine the crop classes, train the data,
and validate the results. The five major crops viz. Cotton, Paddy, Orchard, Yellow split Pigeon peas, Chickpeas, and Other crops were
identified and classified efficiently. According to the experimental results, the random forest approach had the best overall accuracy,
87.71%, and a kappa value of 0.86 than other methods. Alternatively, the ISODATA method provided an overall accuracy of 85.01%
with a kappa value of 0.82. The agricultural decision-makers can use the results of this study for decision-making and management.
