Automated Plot-Level Paddy Land Mapping in Smallholder Agriculture Using Deep Learning and Multi-Temporal Satellite Imagery - A Sri Lankan Case Study Rajanayake R.M.A.B.1*, Fernando W.J.G.S.T.V.2, and Dharmawansha. B.A.D.K.H.3
GeoEDGE (Pvt) Ltd
Abstract
ABSTRACT
Accurate and efficient mapping of paddy lands in Sri Lanka remains a critical challenge due to the heterogeneous and homogeneous nature of plot-level cultivation, coupled with the limitations of manual surveys and conventional remote sensing techniques. Existing methods struggle with spectral similarities between paddy fields and other vegetation, particularly in smallholder-dominated landscapes where field boundaries are irregular and fragmented. This study develops an automated deep learning model to precisely identify and measure paddy land extent using multi-temporal Sentinel-2 satellite imagery at 10-metre resolution. The proposed approach combines a U-Net convolutional neural network with spectral-temporal feature extraction, integrating vegetation indices (NDVI, NDWI) and phenological characteristics to improve discrimination of paddy fields from other land covers. Model training and validation use a ground-truth dataset of 8,000 plot boundaries from key paddy-growing areas in Sri Lanka. Comparative evaluation demonstrates the model^s superior performance, achieving 91.4% classification accuracy and an F1-score of 0.88 for paddy identification, significantly outperforming conventional machine learning approaches such as random forest (78%) and support vector machines (73%) in plot-level delineation. The system successfully addresses key challenges in smallholder paddy mapping, including mixed cropping patterns and seasonal variations in field conditions. This AI framework replaces labour-intensive surveys with accurate paddy mapping, yield estimation, and decision support, and can be adapted for similar regions to support sustainable land and yield management.
Keywords: Deep learning, convolutional neural networks, remote sensing, paddy land mapping
Topic: Topic C: Emerging Technologies in Remote Sensing
