What is the land for?
Problem being addressed
Understanding land usage is an important part of governance and regulation. Commercial land is different from residential land, and they are both different from agricultural land. How can remote sensing be applied to land use classification?
Convolutional neural networks (CNN) that have a network in a network structure are applied to the land classification problem of putting geotagged images into usage categories. Two classification architectures GoogLeNet and CaffeNet are put to the test on a number of different land image datasets; with different configurations. The datasets cover the urban US, and Brazilian coffee plantations; and the algorithms are trained to differentiate in fine detail between anything from golf courses to parking lots. It is found that fine-tuning a pertained CNN is the optimal process for this task (as opposed to training on from scratch) and that overhead high-res images work best.
Advantages of this solution
This is a very accurate solution in terms of land use categorisation and could be useful for urban planning.
Solution originally applied in these industries
Possible New Application of the Work
This task involves putting words to pictures, which is pretty much the same process as for language acquisition for children (at least for named objects). Perhaps the classification tasks could be adapted to build early age kids games aimed at language teaching.
Blind people cannot watch movies, but they can experience a movie as a story if the scenes are described to them carefully enough. What about creating an image to text architecture that "looks" at a movie and describes it for a blind person scene-by-scene.
Travel and Tourism Industry
When you book a hotel you don't get to find out much about the neighbourhood around it. What about using image classifiers to tell you a bit about the place? It could be the difference between a good experience as a traveller and a great one.
Source URL: #############