Abstract:
Monitoring crop cultivation is vital for food security, land-use planning, and sustainable agri
culture, especially in regions dependent on rainfall. Accurate mapping and measurement of cultivated and
non-cultivated paddy lands are essential for making successful agricultural plans and climate-resilient farm
ing in Sri Lanka, particularly in a dry zone like the Northern Province, which relies heavily on rainfall for
agrarian practices. This study aims to develop an automated, deep learning-based segmentation model us
ing high-resolution satellite images from Google Earth Pro to estimate paddy field area. A comprehensive
methodology was followed, involving satellite image collection, preprocessing, manual annotation of masks
using Photopea, model training, and performance evaluation. We employed four segmentation strategies:
Type 1 distinguished cultivated fields from background, Type 2 separated non-cultivated fields from back
ground, Type 3 combined cultivated and non-cultivated fields against background, and Type 4 implemented
a multi-class model to segment background, cultivated, and non-cultivated areas simultaneously. This study
explored four segmentation approaches using both U-Net and the Segment Anything Model (SAM). Among
them, the U-Net Type 4 model that could segment multi-classes demonstrated better scalability and ap
plicability with the validation accuracy of 94.01% and dice coefficients of 0.9145, 0.9179, and 0.8898 of
background, cultivated, and non-cultivated areas, correspondingly. The pixel-based analysis was done in the
Manthai East division, Northern Province, Sri Lanka, to determine the actual area estimation and verify
the real-world relevance of this model. The suggested model shows that deep learning has the potential to
segment crops accurately and provides a cost-effective approach for monitoring agricultural land use prac
tices. This research enables agrarian officers to estimate paddy lands using freely available high-resolution
satellite images to monitor paddy fields, plan resources, and assess seasonal yields. Future work will expand
this framework to include other crop types and longitudinal analysis for prediction. The limitation is per
formance variation due to seasonal changes and cloud interference in satellite images.
