"}],"short_description":"See how new prediction enhancements in ArcGIS Pro 3.6 create more reliable models for spatial prediction.","flexible_content":[{"acf_fc_layout":"content","content":"Predicting where diseases may emerge is an impactful use of spatial data science. In this workflow, we will explore how new prediction evaluation enhancements in ArcGIS Pro can help produce more realistic and trustworthy models. Disclaimer<\/span>: this blog is meant only for demonstrative purposes of new features in ArcGIS Pro.<\/em><\/p>\n"},{"acf_fc_layout":"content","content":" For this example, we will be studying the tick-borne illness anaplasmosis. Human anaplasmosis is transmitted by ticks and is unevenly distributed across the United States. For this analysis, we start with data for all U.S. counties from the CDC<\/a>, but to simulate a realistic prediction challenge, we will intentionally exclude Minnesota from model training. This setup mirrors a common real-world scenario: we want to predict disease risk in a region where we don\u2019t yet have confirmed observations.<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955806,"id":2955806,"title":"eval12","filename":"eval12.jpg","filesize":71856,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval12","alt":"map of anaplasmosis presence in counties for the entire continental US, and the same map with Minnesota removed","author":"437350","description":"","caption":"","name":"eval12","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 22:31:42","modified":"2026-02-09 22:33:01","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":578,"height":780,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","medium-width":193,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","medium_large-width":578,"medium_large-height":780,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","large-width":578,"large-height":780,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","1536x1536-width":578,"1536x1536-height":780,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","2048x2048-width":578,"2048x2048-height":780,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12-345x465.jpg","card_image-width":345,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval12.jpg","wide_image-width":578,"wide_image-height":780}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" We begin by training a Gradient-Boosted model using the Forest-based and Boosted Classification and Regression tool. The model incorporates:<\/p>\n The tool automatically evaluates model performance using randomly selected hold-out data. At first glance, the results look reasonable:<\/p>\n Sensitivity is 57%, meaning the model correctly identifies a little more than half of counties with reported anaplasmosis that it has not seen during training.<\/p>\n Based on this metric, we would expect the model to identify roughly half of the affected counties when predicting into Minnesota.<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955793,"id":2955793,"title":"eval4","filename":"eval4.jpg","filesize":37113,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval4","alt":"output messages of the FBBCR tool showing 57% sensitivity for disease counties","author":"437350","description":"","caption":"","name":"eval4","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:04","modified":"2026-02-09 22:34:20","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":685,"height":331,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","medium-width":464,"medium-height":224,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","medium_large-width":685,"medium_large-height":331,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","large-width":685,"large-height":331,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","1536x1536-width":685,"1536x1536-height":331,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","2048x2048-width":685,"2048x2048-height":331,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","card_image-width":685,"card_image-height":331,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval4.jpg","wide_image-width":685,"wide_image-height":331}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" When we apply the trained model to Minnesota, the results are unexpected. The model identifies none of the known anaplasmosis counties.<\/p>\n So what happened? The reported performance suggested the model was adequate, yet it completely failed in a new spatial context.<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955794,"id":2955794,"title":"eval5","filename":"eval5.jpg","filesize":68311,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval5","alt":"side by side comparison of the predicted counties (all negative) and the observed counties (many positive)","author":"437350","description":"","caption":"","name":"eval5","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:06","modified":"2026-02-09 22:34:51","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":707,"height":668,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","medium-width":276,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","medium_large-width":707,"medium_large-height":668,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","large-width":707,"large-height":668,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","1536x1536-width":707,"1536x1536-height":668,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","2048x2048-width":707,"2048x2048-height":668,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5-492x465.jpg","card_image-width":492,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval5.jpg","wide_image-width":707,"wide_image-height":668}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" There are two issues at play.<\/p>\n The initial evaluation used random validation, meaning training and test data were drawn from geographically similar areas. This allows the model to perform well on locations that closely resemble what it has already seen.<\/p>\n However, Minnesota represents a new spatial region, and random validation does not adequately test the model\u2019s ability to generalize across space. Recall Tobler’s First Law of Geography: Everything is related to everything else, but near things are more related than distant things.<\/p>\n"},{"acf_fc_layout":"content","content":" Anaplasmosis occurs in a relatively small number of counties. This class imbalance makes it easy for a model to be overwhelmed by the majority (non-disease) counties and still appear accurate under standard evaluation metrics.<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955795,"id":2955795,"title":"eval6","filename":"eval6.jpg","filesize":27760,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval6","alt":"a pie chart showing that anaplasmosis has less than 5% presence","author":"437350","description":"","caption":"","name":"eval6","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:08","modified":"2026-02-09 22:35:13","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":620,"height":592,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","medium-width":273,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","medium_large-width":620,"medium_large-height":592,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","large-width":620,"large-height":592,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","1536x1536-width":620,"1536x1536-height":592,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","2048x2048-width":620,"2048x2048-height":592,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6-487x465.jpg","card_image-width":487,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval6.jpg","wide_image-width":620,"wide_image-height":592}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" To address these issues, we turn to the Evaluate Predictions with Cross-Validation tool.<\/p>\n This tool introduces two critical improvements for rare, spatial phenomena:<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955796,"id":2955796,"title":"eval7","filename":"eval7.jpg","filesize":96155,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval7","alt":"a map showing a balanced dataset and the Evaluate Predictions with Cross Validation tool","author":"437350","description":"","caption":"","name":"eval7","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:10","modified":"2026-02-09 22:36:05","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":1165,"height":640,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","medium-width":464,"medium-height":255,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","medium_large-width":768,"medium_large-height":422,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","large-width":1165,"large-height":640,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","1536x1536-width":1165,"1536x1536-height":640,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","2048x2048-width":1165,"2048x2048-height":640,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7-826x454.jpg","card_image-width":826,"card_image-height":454,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval7.jpg","wide_image-width":1165,"wide_image-height":640}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" Instead of random splits, the tool performs spatial cross-validation, forcing the model to prove itself in geographic areas it has not encountered during training. This validation happens many times in many spatially contiguous areas to really test the model’s ability to generalize.<\/p>\n This produces a much more realistic estimate of how the model will behave when predicting into new regions.<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955817,"id":2955817,"title":"eval0","filename":"eval0.jpg","filesize":14738,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval0","alt":"a graphic showing how cross validation works","author":"437350","description":"","caption":"","name":"eval0","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 23:00:23","modified":"2026-02-09 23:00:34","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":422,"height":282,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","medium-width":391,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","medium_large-width":422,"medium_large-height":282,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","large-width":422,"large-height":282,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","1536x1536-width":422,"1536x1536-height":282,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","2048x2048-width":422,"2048x2048-height":282,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","card_image-width":422,"card_image-height":282,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval0.jpg","wide_image-width":422,"wide_image-height":282}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" To prevent the model from being dominated by non-disease counties, we apply k-medoids undersampling. This balances the training dataset, so there are equal numbers of disease and non-disease counties for each validation run, while preserving the patterns in the data.<\/p>\n Note<\/span>: Once you prove that balancing your data gives you reliable results, the Prepare Data for Predictions tool will help you take the necessary data engineering steps to balance data for training your final predictions.<\/em><\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955818,"id":2955818,"title":"eval00","filename":"eval00.jpg","filesize":23059,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval00","alt":"a graphic showing undersampled data with a spatial split","author":"437350","description":"","caption":"","name":"eval00","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 23:01:09","modified":"2026-02-09 23:01:21","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":667,"height":305,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","medium-width":464,"medium-height":212,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","medium_large-width":667,"medium_large-height":305,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","large-width":667,"large-height":305,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","1536x1536-width":667,"1536x1536-height":305,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","2048x2048-width":667,"2048x2048-height":305,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","card_image-width":667,"card_image-height":305,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval00.jpg","wide_image-width":667,"wide_image-height":305}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" With these two factors in place:<\/p>\n (1) Sensitivity improves to ~67% under spatial cross-validation<\/p>\n"},{"acf_fc_layout":"image","image":{"ID":2955798,"id":2955798,"title":"eval8","filename":"eval8.jpg","filesize":48793,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval8","alt":"output messages showing the average cross-validation sensitivity of 66%","author":"437350","description":"","caption":"","name":"eval8","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:12","modified":"2026-02-09 22:36:30","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":675,"height":588,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","medium-width":300,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","medium_large-width":675,"medium_large-height":588,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","large-width":675,"large-height":588,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","1536x1536-width":675,"1536x1536-height":588,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","2048x2048-width":675,"2048x2048-height":588,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8-534x465.jpg","card_image-width":534,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval8.jpg","wide_image-width":675,"wide_image-height":588}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" (2) When predicting into Minnesota, the model now correctly identifies about 60% of disease counties<\/p>\n \n"},{"acf_fc_layout":"image","image":{"ID":2955799,"id":2955799,"title":"eval9","filename":"eval9.jpg","filesize":69683,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval9","alt":"side by side maps showing the new predictions lining up better with observed disease counties","author":"437350","description":"","caption":"","name":"eval9","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 21:55:14","modified":"2026-02-09 22:37:03","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":722,"height":662,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","medium-width":285,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","medium_large-width":722,"medium_large-height":662,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","large-width":722,"large-height":662,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","1536x1536-width":722,"1536x1536-height":662,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","2048x2048-width":722,"2048x2048-height":662,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9-507x465.jpg","card_image-width":507,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval9.jpg","wide_image-width":722,"wide_image-height":662}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":" This alignment between reported performance and real-world prediction is exactly what we want to see.<\/p>\n"},{"acf_fc_layout":"content","content":" In most applied prediction problems, we don\u2019t have \u201cMinnesota data\u201d waiting in the wings to validate our results. We must rely on model diagnostics to decide whether predictions can be trusted.<\/p>\n By accounting for spatial extrapolation and rare events, the new prediction evaluation tools in ArcGIS Pro provide performance metrics that reflect how models will actually behave in the real world. With accurate metrics come confidence, which is key when models are being used to inform decisions and policy.<\/p>\n"},{"acf_fc_layout":"content","content":"The Challenge: Predicting to New Places<\/h2>\n"},{"acf_fc_layout":"content","content":"
Training an Initial Model<\/h2>\n"},{"acf_fc_layout":"image","image":{"ID":2955807,"id":2955807,"title":"eval22","filename":"eval22.jpg","filesize":93373,"url":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","link":"https:\/\/www.esri.com\/arcgis-blog\/products\/arcgis-pro\/geoai\/predicting-to-new-geographic-regions-with-spatially-aware-model-evaluation\/eval22","alt":"maps of tick observations, whitetailed deer observations, and mouse observations","author":"437350","description":"","caption":"","name":"eval22","status":"inherit","uploaded_to":2955769,"date":"2026-02-09 22:31:44","modified":"2026-02-09 22:32:23","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/www.esri.com\/arcgis-blog\/wp-includes\/images\/media\/default.png","width":423,"height":787,"sizes":{"thumbnail":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22-213x200.jpg","thumbnail-width":213,"thumbnail-height":200,"medium":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","medium-width":140,"medium-height":261,"medium_large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","medium_large-width":423,"medium_large-height":787,"large":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","large-width":423,"large-height":787,"1536x1536":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","1536x1536-width":423,"1536x1536-height":787,"2048x2048":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","2048x2048-width":423,"2048x2048-height":787,"card_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22-250x465.jpg","card_image-width":250,"card_image-height":465,"wide_image":"https:\/\/www.esri.com\/arcgis-blog\/app\/uploads\/2026\/02\/eval22.jpg","wide_image-width":423,"wide_image-height":787}},"image_position":"center","orientation":"horizontal","hyperlink":""},{"acf_fc_layout":"content","content":"
\n
What Went Wrong?<\/h2>\n
1. Random Validation Can Inflate Performance<\/h2>\n
2. Anaplasmosis Is a Rare Event<\/h2>\n
Improving Evaluation with Spatial Cross-Validation<\/h2>\n
Testing True Spatial Generalization<\/h2>\n
Handling Rare Events with Undersampling<\/h2>\n
Results That Match Reality<\/h2>\n
Why This Matters<\/h2>\n
Further Reading<\/h2>\n
\n