Applying Convolutional Neural Networks to Per-pixel Orthoimagery Land Use Classification

摘要

Recently, the proliferation of Convolutional Neural Networks has spurred research in a wide range of fields such as image recognition, voice synthesis, and various other classification tasks. Over the last several years, the availability of satellite and other forms of orthoimagery has also increased due to the decreasing cost of capturing devices. The amount of annotated or labeled orthoimagery has not kept pace with the increased availability of imagery, largely due to the time complexity of labeling such data. Land cover usage classifications in particular would have many uses in agriculture. The United States Department of Agriculture's National Agricultural Statistics Service provides land cover usage data at a resolution of 30 meters, which compared with - for example - a 1 meter imagery resolution, leaves a large discrepancy between the quality of the raw image data and the labeling data. This research uses these low quality labels along with high quality image data to train a model that attempts to perform per-pixel land use classification, in hopes to create a classifier that is able to predict several different classes of land use, up to or beyond the resolution accuracy of the much less adequate label data set. It is important to note however, that it is very difficult to evaluate if a model provides relatively better classifications based on the semantics of the input image, due to the low resolution of the image labels. This is because, an individual pixel in the image label will only represent one class per NxN meter area - in the case of our data set, a 30x30 meter area. That individual pixel may be a poor representation of features actually represented in the higher resolution image data. Thusly, we will attempt to demonstrate that, with enough data, a model may generate higher a resolution classification than the original imagery labels with a reasonable margin of error, and attempt to define a way to evaluate the effectiveness of the model despite the poor resolution of the image labels.