|
Crop type information is crucial for several agricultural applications and even useful for different agencies to take important decision. Therefore, accurate crop mapping is required which is a simple yet critical issue in agriculture. Remote sensing has played an important role in acquiring necessary crop data. The free available Planet Scope dataset with a spatial resolution of 3m has generated new opportunities for mapping small land holding. The objective of the study is to map a Mustard field in Haridwar district, as it is considered as one of the important catch crop in the study area. Random Forest (RF) and Classification and Regression Tree (CART) algorithms of machine learning have been in this study. Further, spectral indices images of NDVI, EVI2, NDRE1 and BBI were generated from the original data set. On the temporal PlanetScope dataset, separability analysis is first carried out using the transformed divergence approach. This gives us the optimal three band combination and best time stamp for mapping mustard, which is then used in the study together with spectral indices for mapping mustard crop. Hyperparameter tuning is done to achieve high accuracy, and utilizing the optimized value of the parameters, the classification is carried out using the aforementioned powerful algorithms. The classification findings demonstrate that RF (85.78%) offers a more accurate result than CART (77.75%) in terms of total accuracy. However, both classifiers offer approximately same result in the field of agricultural. For example, RF classifies mustard with an accuracy of 93.33% while CART achieves 90.69%, while for mapping other crops, RF achieves 91% accuracy and CART achieves 84.67%. However, RF provides more precise mapping for Mustard than CART does. According to the results of the study, Random Forest produces the best outcomes when original data and spectral indices are combined.
|
