{"id":299562,"date":"2020-01-16T19:59:25","date_gmt":"2020-01-17T03:59:25","guid":{"rendered":"https:\/\/www.esri.com\/about\/newsroom\/?post_type=arcnews&p=299562"},"modified":"2024-03-26T17:02:56","modified_gmt":"2024-03-27T00:02:56","slug":"australian-state-automated-large-area-land-classification-with-machine-learning","status":"publish","type":"arcnews","link":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning","title":{"rendered":"Australian State Automated Large-Area Land Classification with Machine Learning"},"author":5752,"featured_media":0,"menu_order":0,"template":"","format":"standard","meta":{"_acf_changed":false,"sync_status":"","episode_type":"","audio_file":"","podmotor_file_id":"","podmotor_episode_id":"","castos_file_data":"","cover_image":"","cover_image_id":"","duration":"","filesize":"","filesize_raw":"","date_recorded":"","explicit":"","block":"","itunes_episode_number":"","itunes_title":"","itunes_season_number":"","itunes_episode_type":"","_links_to":"","_links_to_target":""},"categories":[551,23422,1031],"tags":[20422,295112,237901,282882,12662],"arcnews_issues":[409212],"class_list":["post-299562","arcnews","type-arcnews","status-publish","format-standard","hentry","category-imagery","category-machine-learning-capability","category-natural-resources","tag-arcgis-pro","tag-australia","tag-land-use","tag-landcover","tag-remote-sensing","arcnews_issues-winter-2020","arcnews_sections-news"],"acf":{"short_description":"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.","pdf":{"host_remotely":false,"file":"","file_url":""},"flexible_content":[{"acf_fc_layout":"content","content":"The state of Queensland, in northeastern Australia, is remarkably geographically diverse. It includes coastal rain forests, widespread eucalypt and acacia woodlands, tropical savannas, ephemeral inland rivers, deserts, and rich agricultural belts. With an area of 1,730,000 square kilometers (668,000 square miles), it is approximately seven times the size of Great Britain.\r\n\r\nTo map and assess land-use patterns and changes throughout the state, Queensland\u2019s Department of Environment and Science (DES) formed the Queensland Land Use Mapping Program (QLUMP) more than 20 years ago."},{"acf_fc_layout":"image","image":299552,"image_position":"right","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":"\u201cLand use has been identified as a foundational spatial dataset that the government considers vital for the progression and development of Queensland,\u201d said Andy Clark, senior scientist at the DES Remote Sensing Centre. \u201cThe state is large, and it is important that we continue to improve the speed at which we collect this data, as well as its accuracy. In addition, we must keep our procedures in accordance with the standards established by the Australian Land Use and Management Classification system so that it is consistent with data collected throughout the entire country.\u201d\r\n\r\nTraditionally, the methodology used to keep QLUMP up-to-date relied on a team of skilled spatial scientists to manually digitize land-use features from satellite imagery. Because of the size of Queensland, this process took a lot of time and resources.\r\n\r\n\u201cPreviously, we have made various attempts to automate QLUMP, all of which ultimately proved to be unsuccessful,\u201d said Clark. \u201cDecision tree models were used to infer land-use features from ancillary data; however, this method did not provide an accurate representation of what was on the ground. We tried using the spectral information from satellite imagery to conduct a supervised classification but determined that this procedure could not successfully distinguish between features because, spectrally, they appear very similar. Also, object-based image analyses tended to be just as resource-intensive as manually drawing land-use features.\u201d"},{"acf_fc_layout":"quote","image":"","text":"Computer vision in fusion with high-performance supercomputing and integrated with ArcGIS represents a paradigm shift that increases our capacity to compile and publish timely land-use information.","author_name":"Andy Clark","author_profession_organization":"Senior Scientist, Queensland Department of Environment and Science\u2019s Remote Sensing Centre"},{"acf_fc_layout":"content","content":"In recent years, however, machine learning\u2014a subdiscipline of artificial intelligence (AI)<\/a>\u2014has progressed to the point that using computer vision and deep learning in image analysis and classification is now viable.\r\n\r\n\u201cWith advanced programming tools and computer hardware, the speed and capabilities required to successfully apply machine learning to accurately classify large areas of land looks very promising,\u201d said Clark.\r\n\r\nIn fact, he and his team developed a model that can automatically classify different types of land use throughout Queensland. Here\u2019s how they got it working."},{"acf_fc_layout":"content","content":"

Refining the Data and Training the Process<\/h2>\r\nIt requires a lot of information to train a model to do machine learning. Fortunately for DES, it had been using QLUMP for years to collect data.\r\n\r\n\u201cIt was just a matter of refining it so that it could be used in the machine learning process,\u201d said Clark."},{"acf_fc_layout":"image","image":299532,"image_position":"left","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":"He used ArcGIS Pro and ArcPy to generate and refine the training data. Clark also applied a range of geoprocessing tools to postprocess the prediction probability from the computer vision model.\r\n\r\n\u201c[The] Reclassify [tool was] used to convert the prediction to a binary raster,\u201d he explained. \u201cRaster to Polygon [was used] to convert the data to a feature class. Union [was used] to derive change. And Eliminate [was used] to merge small features into larger ones.\u201d\r\n\r\nThe QLUMP team independently verified the accuracy of the process by randomly generating thousands of points and assessing the land use at each point.\r\n\r\n\u201cArcGIS Pro also generates the error matrix for us, as well as the creation and publication of web maps, apps, and reports to communicate with our stakeholders,\u201d Clark added.\r\n\r\nIn the machine learning process, DES uses a convolutional neural network (CNN) based on a U-net architecture to help the model visually recognize land cover. CNNs are algorithms that mimic the functions of the human brain. By being exposed to large amounts of visual data, the model can learn to distinguish between similarities and dissimilarities in the data.\r\n\r\n\u201cWe borrowed the idea from Olaf Ronneberger, who developed it for biomedical image segmentation, which is a way to identify cells in microscopy images,\u201d Clark explained. \u201cUsing this architecture, we created an algorithm with 87,153,153 parameters.\u201d\r\n\r\nThe team iteratively fed thousands of satellite imagery patches through the neural network to produce a prediction. The algorithm then self-evaluated and refined the prediction, and the cycle was repeated until it ultimately achieved a 97-percent-accuracy rate."},{"acf_fc_layout":"image","image":299542,"image_position":"right","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":"\u201cPython was used to develop the computer vision part of the project,\u201d said Clark. \u201cWe used NumPy, a library with a large collection of high-level mathematical functions for Python, to handle the multidimensional array and the Geospatial Data Abstraction Library (GDAL) to read the imagery and convert it to the NumPy array\u2014the format required for the neural network.\u201d\r\n\r\nGDAL can then take an output array and convert it back to an image.\r\n\r\n\u201cThere was a small component of GDAL reading vector data, but ArcGIS Pro was the main tool for processing vectors,\u201d said Clark. \u201cWe also used Keras, a Python library, for developing and evaluating deep learning models. TensorFlow ran in the back end. It is an artificial intelligence library for data flow and the creation of large-scale neural networks.\u201d\r\n\r\nIn machine learning, because of the amount of data that has to be processed and refined quickly and repeatedly, processing speed is critical. That\u2019s why DES is using eight Tesla V100 graphics processing units (GPUs) that are connected to its high-performance computing infrastructure for deep learning data processing.\r\n\r\n\u201cThe processing speed is amazing,\u201d said Clark. \u201cEach GPU essentially provides us with the equivalent of a year\u2019s worth of conventional CPU processing in about 2.5 days.\u201d"},{"acf_fc_layout":"content","content":"

Expanding Machine Learning to Other Land Uses<\/h2>\r\nOriginally, the model was trained to identify and map banana plantations in the Johnstone River catchment in north Queensland. It was then used to infer banana plantations in the Tully catchment. This allowed the scientists involved in image analysis to focus on the interpretation of the imagery the model produces so they can better inform department decision-makers about appropriate biosecurity responses to plant diseases.\r\n\r\nPanama Tropical Race 4 is a serious disease that can spread rapidly through a banana plantation. In 2015, Queensland\u2019s Department of Agriculture and Fisheries (DAF) detected it when examining plant samples."},{"acf_fc_layout":"image","image":299572,"image_position":"left","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":"At the time, DES had not implemented its computer vision-based image analysis process. So determining the potential spread of the pervasive fungus required a team of five scientists per year to manually map and analyze all the banana plantations and other land-use classes in Queensland.\r\n\r\nThese scientists\u2019 mapping efforts were subsequently used to help train the CNN model. In 2019, DES received new imagery and updated the banana plantation mapping, which took four days for the computer to complete. Because of its speed and accuracy, the CNN model is currently being trained to map other land-use classes.\r\n\r\n\u201cComputer vision in fusion with high-performance supercomputing and integrated with ArcGIS represents a paradigm shift that increases our capacity to compile and publish timely land-use information,\u201d said Clark. \u201cThe methods are sustainable for any image segmentation task and have been applied to mapping wooded vegetation in Queensland, which is quite a different application, as these areas range from dense rain forest to scattered-tree landscapes.\u201d\r\n\r\nDES intends to expand its methods into most land uses, including other crop types, forestry plantations, and urban land-use classes.\r\n\r\n\u201cAnything that you can see in the imagery, you can train an algorithm to find. You just need lots of good-quality training data,\u201d said Clark. \u201cBased on the methods developed to date, computer vision has the capability to bring efficiencies to large-area mapping and monitoring programs that inform natural resources management and monitoring by governmental and nongovernmental organizations.\u201d"},{"acf_fc_layout":"sidebar","layout":"standard","image_reference":null,"image_reference_figure":"","spotlight_image":null,"section_title":"","spotlight_name":"","position":"Center","content":"Since the completion of this project, ArcGIS Pro and ArcGIS API for Python have undergone several enhancements. ArcGIS API for Python now natively supports the U-net model, and in addition to being able to train a model using ArcGIS Notebooks, users can now train deep learning models natively through ArcGIS Pro using a geoprocessing tool. Both ArcGIS Pro and Notebooks support end-to-end deep learning workflows, from labeling and preparing data to training a model and running inferencing. This\u2014combined with ArcGIS Image Server technology that manages imagery data efficiently\u2014significantly simplifies workflows. Moving forward, DES is looking into running its land-cover classification workflow through ArcGIS.","snippet":""}],"references":null},"yoast_head":"\nAustralian State Automated Large-Area Land Classification with Machine Learning<\/title>\n<meta name=\"description\" content=\"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Australian State Automated Large-Area Land Classification with Machine Learning\" \/>\n<meta property=\"og:description\" content=\"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning\" \/>\n<meta property=\"og:site_name\" content=\"Esri\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/esrigis\/\" \/>\n<meta property=\"article:modified_time\" content=\"2024-03-27T00:02:56+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.esri.com\/about\/newsroom\/app\/uploads\/2020\/01\/arcnews-banner-withmatchinelearing-wide.jpg\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:title\" content=\"Australian State Automated Large-Area Land Classification with Machine Learning\" \/>\n<meta name=\"twitter:description\" content=\"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.\" \/>\n<meta name=\"twitter:image\" content=\"https:\/\/www.esri.com\/about\/newsroom\/app\/uploads\/2020\/01\/arcnews-banner-withmatchinelearing-wide.jpg\" \/>\n<meta name=\"twitter:site\" content=\"@Esri\" \/>\n<meta name=\"twitter:label1\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data1\" content=\"8 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\n\t \"@context\": \"https:\/\/schema.org\",\n\t \"@graph\": [\n\t {\n\t \"@type\": \"WebPage\",\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning\",\n\t \"url\": \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning\",\n\t \"name\": \"Australian State Automated Large-Area Land Classification with Machine Learning\",\n\t \"isPartOf\": {\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/#website\"\n\t },\n\t \"datePublished\": \"2020-01-17T03:59:25+00:00\",\n\t \"dateModified\": \"2024-03-27T00:02:56+00:00\",\n\t \"description\": \"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.\",\n\t \"breadcrumb\": {\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning#breadcrumb\"\n\t },\n\t \"inLanguage\": \"en-US\",\n\t \"potentialAction\": [\n\t {\n\t \"@type\": \"ReadAction\",\n\t \"target\": [\n\t \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning\"\n\t ]\n\t }\n\t ]\n\t },\n\t {\n\t \"@type\": \"BreadcrumbList\",\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning#breadcrumb\",\n\t \"itemListElement\": [\n\t {\n\t \"@type\": \"ListItem\",\n\t \"position\": 1,\n\t \"name\": \"Home\",\n\t \"item\": \"https:\/\/www.esri.com\/about\/newsroom\"\n\t },\n\t {\n\t \"@type\": \"ListItem\",\n\t \"position\": 2,\n\t \"name\": \"ArcNews Articles\",\n\t \"item\": \"https:\/\/www.esri.com\/about\/newsroom\/arcnews\"\n\t },\n\t {\n\t \"@type\": \"ListItem\",\n\t \"position\": 3,\n\t \"name\": \"Australian State Automated Large-Area Land Classification with Machine Learning\"\n\t }\n\t ]\n\t },\n\t {\n\t \"@type\": \"WebSite\",\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/#website\",\n\t \"url\": \"https:\/\/www.esri.com\/about\/newsroom\/\",\n\t \"name\": \"Esri\",\n\t \"description\": \"Esri Newsroom\",\n\t \"potentialAction\": [\n\t {\n\t \"@type\": \"SearchAction\",\n\t \"target\": {\n\t \"@type\": \"EntryPoint\",\n\t \"urlTemplate\": \"https:\/\/www.esri.com\/about\/newsroom\/?s={search_term_string}\"\n\t },\n\t \"query-input\": {\n\t \"@type\": \"PropertyValueSpecification\",\n\t \"valueRequired\": true,\n\t \"valueName\": \"search_term_string\"\n\t }\n\t }\n\t ],\n\t \"inLanguage\": \"en-US\"\n\t },\n\t {\n\t \"@type\": \"Person\",\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/#\/schema\/person\/41c803b2ea8734c36f9c4e9586d1449d\",\n\t \"name\": \"Amy Ambard\",\n\t \"image\": {\n\t \"@type\": \"ImageObject\",\n\t \"inLanguage\": \"en-US\",\n\t \"@id\": \"https:\/\/www.esri.com\/about\/newsroom\/#\/schema\/person\/image\/\",\n\t \"url\": \"https:\/\/secure.gravatar.com\/avatar\/f356480172f8ad0bc8d72b855e84171c52f1944c7c7779f3e425d73bf3efa3c7?s=96&d=blank&r=g\",\n\t \"contentUrl\": \"https:\/\/secure.gravatar.com\/avatar\/f356480172f8ad0bc8d72b855e84171c52f1944c7c7779f3e425d73bf3efa3c7?s=96&d=blank&r=g\",\n\t \"caption\": \"Amy Ambard\"\n\t },\n\t \"url\": \"\"\n\t }\n\t ]\n\t}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Australian State Automated Large-Area Land Classification with Machine Learning","description":"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning","og_locale":"en_US","og_type":"article","og_title":"Australian State Automated Large-Area Land Classification with Machine Learning","og_description":"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.","og_url":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning","og_site_name":"Esri","article_publisher":"https:\/\/www.facebook.com\/esrigis\/","article_modified_time":"2024-03-27T00:02:56+00:00","og_image":[{"url":"https:\/\/www.esri.com\/about\/newsroom\/app\/uploads\/2020\/01\/arcnews-banner-withmatchinelearing-wide.jpg","type":"","width":"","height":""}],"twitter_card":"summary_large_image","twitter_title":"Australian State Automated Large-Area Land Classification with Machine Learning","twitter_description":"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.","twitter_image":"https:\/\/www.esri.com\/about\/newsroom\/app\/uploads\/2020\/01\/arcnews-banner-withmatchinelearing-wide.jpg","twitter_site":"@Esri","twitter_misc":{"Est. reading time":"8 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning","url":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning","name":"Australian State Automated Large-Area Land Classification with Machine Learning","isPartOf":{"@id":"https:\/\/www.esri.com\/about\/newsroom\/#website"},"datePublished":"2020-01-17T03:59:25+00:00","dateModified":"2024-03-27T00:02:56+00:00","description":"In Queensland, Australia, a computer vision model was trained to automatically classify different types of land use.","breadcrumb":{"@id":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.esri.com\/about\/newsroom\/arcnews\/australian-state-automated-large-area-land-classification-with-machine-learning#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.esri.com\/about\/newsroom"},{"@type":"ListItem","position":2,"name":"ArcNews Articles","item":"https:\/\/www.esri.com\/about\/newsroom\/arcnews"},{"@type":"ListItem","position":3,"name":"Australian State Automated Large-Area Land Classification with Machine Learning"}]},{"@type":"WebSite","@id":"https:\/\/www.esri.com\/about\/newsroom\/#website","url":"https:\/\/www.esri.com\/about\/newsroom\/","name":"Esri","description":"Esri Newsroom","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.esri.com\/about\/newsroom\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/www.esri.com\/about\/newsroom\/#\/schema\/person\/41c803b2ea8734c36f9c4e9586d1449d","name":"Amy Ambard","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.esri.com\/about\/newsroom\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/f356480172f8ad0bc8d72b855e84171c52f1944c7c7779f3e425d73bf3efa3c7?s=96&d=blank&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/f356480172f8ad0bc8d72b855e84171c52f1944c7c7779f3e425d73bf3efa3c7?s=96&d=blank&r=g","caption":"Amy Ambard"},"url":""}]}},"sort_order":"1","_links":{"self":[{"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/arcnews\/299562","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/arcnews"}],"about":[{"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/types\/arcnews"}],"author":[{"embeddable":true,"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/users\/5752"}],"version-history":[{"count":0,"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/arcnews\/299562\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/media?parent=299562"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/categories?post=299562"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/tags?post=299562"},{"taxonomy":"arcnews_issues","embeddable":true,"href":"https:\/\/www.esri.com\/about\/newsroom\/wp-json\/wp\/v2\/arcnews_issues?post=299562"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}