Mapping internal farm roadways to identify runoff accumulation areas using an integrated GIS, aerial imagery and deep learning approach in grassland farms

In temperate regions, farm roadway networks exist on grassland farms to provide an efficient means to move dairy and beef animals between grazed paddocks and farmyards. During their movement, excreta is deposited on the roadway surface, resulting in phosphorus (P) enriched roadway runoff during rainfall events, which could impact water quality. Although national roadway networks have been mapped using deep learning approaches, internal farm roadway network mapping remains a knowledge gap. The objectives of this study were to develop an integrated workflow tailored to agricultural roadway mapping and runoff risk assessment through training a deep learning model to automatically detect and map internal farm roadways using high-resolution aerial imagery and further evaluating roadway sections that have the potential to generate roadway runoff with associated risk of water pollution. Three (3) model architectures (U-Net, PSPNet and DeepLab V3+) were tested and the best performing model was PSPNet with ResNet-50 backbone. The selected model demonstrated an overall performance of 0.79, 0.86, 0.82, 0.69 and 0.90 for precision, recall, F₁ score IoU and overall accuracy, respectively. A total of 34.6 km of internal farm roadways on 10 grassland farms were extracted using the deep learning model. Further analysis of the roadway network from each of the farms indicated “high runoff susceptibility” ranging from 8.3 % to 20 %. Farm roadway sections with “very high runoff potential” ranged from 0.6 % to 4.9 %. In comparison with the existing dataset of P flow delivery paths in Ireland, this study identified new ‘hotspots’ in farm roadways. The study showed that the developed automated models provide an important and efficient tool to assess farm roadway runoff accumulation areas.