Monsoon paddy crop discrimination using machine learning algorithms to multi-temporal Sentinel-1A (C-band) data in Alathur block of Palakkad district of Kerala state, India

Crop identification and acreage estimation are always challenges for passive remote sensing in cloudy environments, especially during the rainy season. Microwave remote sensing data, with its availability and advantages of cloud penetration, can offer a solution to this problem. In the present study, a dual-pole polarization Sentinel-1 (C-band) SAR dataset was used for the assessment of the paddy temporal growth stage in the Alathur block of Kerala, India. The acreage estimation of coconut, fallow, mixed trees, paddy, rubber, plantation, paddy-coconut, and plantation fallow classes were also carried out using ArcGIS software. The multi-temporal data from June 2021 to October 2021 is openly accessible from the European Space Agency (ESA) via the Sentinel-1. The image was preprocessed for radiometric corrections, speckle filtering, terrain correction, and co-registration, and was finally converted from a linear to a decibel (dB) domain using ESA (SNAP software). The crop discrimination of paddy and other classes was also classified by using the decision tree rule (DTR), K-nearest neighbors, minimum distance (MD), and random forest (RF) classifiers. The results revealed that during the early stage of the paddy crop, specular reflection occurred, which enabled a low backscatter signature of - 18.45 to - 10.9 dB in VH (vertical transmit and horizontal receive) and in VV (vertical transmit and vertical receive) polarization, and with each growth stage of the crop, the backscatter varied. However, there was no significant variation observed for coconut (- 15.02 dB) in VH (vertical transmit and horizontal receive) and (- 8.67 dB) in VV (vertical transmit and vertical receive) and for Rubber (- 14.57 dB) in VH (vertical transmit and horizontal receive) and (- 7.78 dB) in VV (vertical transmit and vertical receive) as compared to paddy field. The results obtained for paddy and other classes from different classifiers revealed that random forest (RF) and K-nearest neighbor classifications gave 95.0 percent and 94.0 percent accuracy for the classification of paddy, respectively, as compared to decision tree rule (DTR) and minimum distance classification.