Enhanced FCN for farmland extraction from remote sensing image

As farmland being the foundation of national agribusiness, it is of paramount significance to obtain data more efficiently about the distribution of farmland for further agricultural resource monitoring. Through classification of Remote Sensing (RS) images combined with deep learning approaches, however, previous studies did not attach enough attention to boundary ambiguity, thus achieving relatively low accuracy and demands artificial refinements in farmland extraction. To remedy flaws in current approaches and improve overall accuracy, our work reviewed relevant literature and utilized K-Means model, U-Net model and DeelLabV3 model respectively, to refine and make adjustments to farmland extraction model of RS image afterwards. After model training and parameter tuning, the final result of the classification model reached 95.76% in terms of overall accuracy, and the average cross-comparison ratio in farmland recognition rate reached 85.44%. We closed our paper with future directions and possible improvements to our work.