Semantic segmentation of agricultural aerial images using encoder-decoder models

The agricultural sector is the backbone of almost all economies in the world. Agriculture supports 70% of the population and covers about 40% of the Earth's surface. A huge quantity of aerial images of agricultural lands is available due to the invention of agricultural drones (unmanned aerial vehicles). Deriving useful information from the captured images can help the farmers in several ways to know the problems in their land. Semantic segmentation is the process of assigning a label to every pixel in the image and clustering the parts of images together, which belong to the same object. Semantic image segmentation identifies where an object is located in the image, the shape of that object, which pixel belongs to which object. This research work aims to segment six types of anomalies (that includes cloud shadow, double plant, planter skip, standing water, water-way and weed cluster) in the aerial farmland images that are most important to farmers. This is carried out in a subset of agriculture-vision - a large aerial image database for agricultural pattern analysis with nine types of anomalies. This is carried out with encoder-decoder architectures aiming to give mean intersection over union (mIoU - metric for model evaluation) greater than the existing work.