Automated LULC map production using deep neural networks

Abstract

This article presents an approach to automating the creation of land-use/land-cover classification (LULC) maps from satellite images using deep neural networks that were developed to perform semantic segmentation of natural images. This work is important since the production of accurate and timely LULC maps is becoming essential to government and private companies that rely on them for large-scale monitoring of land resource changes. In this work, deep neural networks re trained to classify each pixel of a satellite image into one of a number of LULC classes. The presented deep neural networks are all pre-trained using the ImageNet Large-Scale Visual Recognition Competition (ILSVRC) datasets and then fine-tuned using approximately 19,000 Landsat 5/7 satellite images of resolution 224 × 224 taken of the Province of Manitoba in Canada. The result is an automated solution that can produce LULC maps images significantly faster than current semi-automated methods. The contributions of this article are the observation that deep neural networks developed for semantic segmentation can be used to automate the task of producing LULC maps; the use of these networks to produce LULC maps; and a comparison of several popular semantic segmentation architectures for solving the problem of automated LULC map production.

A deep learning framework for land-use/land-cover mapping and analysis using multispectral satellite imagery

“…Furthermore, a context module that uses dilated convolutions to systematically aggregate multiscale-contextual information while retaining resolution was introduced. In addition results are reported in [10] based on the architectures described in [14]. Conditional Random Fields as Recurrent Neural Networks (CRFasRNN) described in [19], tackles dense pixel prediction using Conditional Random Fields (CRFs).…”

“…In [9,10], the fully convolutional network (FCN) originally developed by [8] for semantic segmentation was modified and adapted for automating the production of LULC maps. In this paper, DCNNs originally designed for image classification and object detection tasks are adapted into FCNs that take arbitrary sized input and produce image segmentations [8], i.e.…”

“…This paper presents a case study in exploring recent work on DCNNs for LULC map production of Landsat 5/7 multi-spectral satellite images using LULC classes defined by GeoManitoba 1 . The presented work is an extension of [10], where FCN-8 VGG-16, CRF-RNN, and Dilation 8 (both frontend and context) networks were used on the Landsat 5/7 Manitoba satellite image data for LULC map production. The best solution reported in [10] produced an average accuracy of 88%.…”

“…The presented work is an extension of [10], where FCN-8 VGG-16, CRF-RNN, and Dilation 8 (both frontend and context) networks were used on the Landsat 5/7 Manitoba satellite image data for LULC map production. The best solution reported in [10] produced an average accuracy of 88%. The motivation of this work is to improve the results reported [10] by: i) investigating other deep learning architectures to further improve the classification accuracy, ii) propose a generalized deep learning framework to reduce the time spent pro- ducing LULC maps, and iii) to provide more effective DCNN architectures for LULC map production of 16-bit Landsat 8 GeoManitoba data.…”

“…The best solution reported in [10] produced an average accuracy of 88%. The motivation of this work is to improve the results reported [10] by: i) investigating other deep learning architectures to further improve the classification accuracy, ii) propose a generalized deep learning framework to reduce the time spent pro- ducing LULC maps, and iii) to provide more effective DCNN architectures for LULC map production of 16-bit Landsat 8 GeoManitoba data. The contribution of our work is as follows: i) an improvement in the classification of global accuracy from 88% to 90.46%, ii) extensive experimentation of different FCN architectures with extensions on a unique dataset, and iii) a dramatic reduction in the time it takes to produce these LULC maps (from 4800 hours to 8 minutes and 42 seconds).…”