Semantic Segmentation of Land Use and Land Cover for Mapping Agricultural Activities

Over the decades, deep neural network computer vision architectures have improved their image segmentation and classification performance based on multidisciplinary efforts answering a variety of research questions and tasks. This project aims to create the possibility to improve land use and land cover analysis by employing one of the deep convolutional neural network architectures named DeepLabV3+ to overcome the restriction of statistical models in applied and agricultural economic fields. The problem derived from the small sample size due to the protection of farm operators' privacy, which could be unavoidable from the perspective of ethics.

The research, therefore, builds the pipeline based on the DeepLabV3+ to conduct prediction analysis to classify the agricultural activities in Carroll County, Missouri. After the training with the remote sensing images randomly chosen from the county area, the machine returns fair predictions about the land cover classification under the limitation of data size and computational performance. The project can provide the possibility of using this machine to estimate the farm operators' land use decisions. Furthermore, it indicates that predictive observations can be used for the various socioeconomic impacts on crop production and agricultural policies, such as the Conservation Reserve Program, after improving the initial architecture with more data acquisition.