Title : Mapping aquaculture ponds using deep learning for indian subcontinent
Abstract:
This paper presents a novel approach for mapping inland aquaculture ponds in the Indian subcontinent using Sentinel-2 MultiSpectral Instrument (MSI) satellite imagery. The proliferation of pond aquaculture in this region, driven by the ever-increasing global demand for fish and seafood, underscores the urgency of accurately mapping these ponds. India's role as a significant seafood exporter amplifies the importance of precise pond localization for processing companies. Our algorithm leverages Sentinel-2's NIR, SWIR, and Red bands to initially label ponds in a spatio-temporal context. It employs a ResNet-backed U-Net architecture trained on preprocessed imagery, optimizing pond boundary delineation for small-scale aquaculture ponds. To efficiently process vast satellite images, we partition them into 256x256 patches with a 128-pixel overlap, subsequently stitching them together and applying post-processing techniques using image processing to isolate the desired pond boundaries. By incorporating historical water body statistics from Global Surface Water Extent dataset, we effectively filter out lakes, rivers, and water over agricultural land, thus ensuring that our mapping system is exclusively focused on aquaculture ponds. Our automated pipeline achieves an impressive Intersection over Union (IoU) score of 85%, enabling the detection and mapping of both existing and newly established inland aquaculture ponds with precision, thereby addressing the pressing need for reliable pond data in the region.
Keywords: aquaculture, ponds, sentinel-2 MSI, segmentation, deep learning
