Crop-type mapping and acreage estimation in smallholding plots using Sentinel-2 images and machine learning algorithms: Some comparisons

Abstract

Crop acreage analysis and yield estimation are of prime importance in field-level agricultural monitoring and management. This enables prudent decision making during any crop failure event and for ensuing crop insurance. The free availability of the high resolution Sentinel-2 satellite datasets has created new possibilities for mapping and monitoring agricultural lands in this regard. In the present study conducted on the Tamluk Subdivision of the Purba Medinipur District of West Bengal, the heterogeneous crop area was mapped according to the respective crop type, using Sentinel-2 multi-spectral images and two machine learning algorithms- K Nearest Neighbour (KNN) and Random Forest (RF). Plot-level field information was collected from different cropland types to frame the training and validation datasets (comprising 70% and 30% of the total dataset, respectively) for cropland classification and accuracy assessment. Through this, the major summer crop acreage was identified (Boro rice, vegetables and betel vine- the three main crops in the study area). The extracted maps had an overall accuracy of 97.16% and 97.22%, respectively, in the KNN and RF classifications, with respective Kappa index values of 95.99% and 96.08%, and the RF method proved to be more accurate. This study was particularly useful in mapping the betel leaf acreage herein since scant information exists for this crop and it is cultivated by many smallholder farmers in the region. The methods used in this paper can be readily applied elsewhere for accurately enumerating the respective crop acreages.

Keywords

• Machine learning algorithms
• Smallholder agriculture
• Crop acreage
• Betel vine cultivation