Planet Labs PBC · NYSE:PL
Planet Labs PBC — Technical Presentation (April 2021)
30 slides from Planet Labs PBC's technical presentation, April 2021
Slide 1: Planet Analytic Feeds Performance Metrics
PLANET ANALYTIC FEEDS PERFORMANCE METRICS PLANET.COM April 2021
Slide 2: Overview
Planet Analytic Feeds | Performance Metrics April 2021 OVERVIEW Planet Analytic Feeds leverages computer vision to transform our imagery into information feeds that detect and classify objects, identify geographic features, and understand change over time across the globe. This model performance gui
Slide 3: Road Detection
Planet Analytic Feeds | Performance Metrics April 2021 ROAD DETECTION Planet’s Road Detection leverages a semantic segmentation computer vision model applied to Planet Basemaps. Semantic segmentation analyzes an image and designates a “class” for each pixel in the image. In this case, the model clas
Slide 4: Performance Overview and Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 PERFORMANCE Overview Planet’s models are generalized to work globally and have been tested and tuned extensively on specific AOIs and object definitions. As the Earth is heterogeneous in both time and space, performance will vary significantly d
Slide 5: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Performance San Diego, USA - Great atmospheric conditions - Large roads - No clouds - Urban area Hyderabad, India - Great atmospheric conditions - Small roads - No clouds - Urban area Port Saint Lucie, USA
Slide 6: Building Detection
Planet Analytic Feeds | Performance Metrics April 2021 BUILDING DETECTION Planet’s Building Detection leverages a semantic segmentation computer vision model applied to Planet Basemaps. Semantic segmentation analyzes an image and designates a “class” for each pixel in the image. In this case, the mo
Slide 7: Performance Overview
Planet Analytic Feeds | Performance Metrics April 2021 PERFORMANCE Overview Planet’s models are generalized to work globally and have been tested and tuned extensively on specific AOIs and object definitions. As the Earth is heterogeneous in both time and space, performance will vary significantly d
Slide 8: Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 Key Factors Our model will perform differently depending on key characteristics of the AOIs. Common false positives include bare earth patches, dense buildings bleed together, covered crop fields, clouds and cloud shadows, tree crowns, scattered
Slide 9: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Performance Hefei, China - Great atmospheric conditions - Large buildings - No clouds - Urban area Dubai, UAE - Great atmospheric conditions - Large buildings - No clouds - Planes - Urban area Baikonur, Ka
Slide 10: Vessel Detection
Planet Analytic Feeds | Performance Metrics April 2021 Hyderabad, India - Bad atmospheric conditions - Small buildings - No clouds - Barren - Suburban VESSEL DETECTION Planet’s Vessel Detection leverages an object detection computer vision model applied to Planet’s PlanetScope imagery. Object detect
Slide 11: Performance Overview and Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 PERFORMANCE Overview Planet’s models are generalized to work globally in ports less than 15 km offshore and have been tested and tuned extensively on specific AOIs and object definitions.As the Earth is heterogeneous in both time and space, perf
Slide 12: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Performance Nampo, North Korea - Great atmospheric conditions - Large vessels only - No islands - No end of piers - No docked ships Nampo, North Korea - Great atmospheric conditions - Small vessels - No is
Slide 13: Plane Detection
Planet Analytic Feeds | Performance Metrics April 2021 Port Dalian, China - Poor atmospheric conditions - All vessels - No islands - End of piers - Docked ships Naples, Italy - Decent atmospheric conditions - Small vessels only - No islands - Long piers - Locked ships PLANE DETECTION Planet’s Plane
Slide 14: Performance Overview and Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 PERFORMANCE Overview Planet’s Plane Detection model was trained on a distributed set of airfields and airports throughout the world, designed to expose the model to different conditions. It performs well within constrained areas of interest focu
Slide 15: Beijing Capital International Airport and Hartsfield–Jackson Atlanta
Planet Analytic Feeds | Performance Metrics April 2021 Beijing Capital International - Bad atmospheric Airport conditions - Large planes Hartsfield–Jackson Atlanta - Good atmospheric International Airport conditions - Small planes and large planes © Planet Labs Inc. 2020 15
Slide 16: Well Pad Detection
Planet Analytic Feeds | Performance Metrics April 2021 WELL PAD DETECTION Planet’s Well Pad Detection leverages an object detection computer vision model applied to Planet Basemaps. Object detection is a computer vision approach that analyzes an image and generates a “bounding box” around the releva
Slide 17: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Performance Permian Basin, USA - Clear well pads - No building - Rural area Permian Basin, USA - Clear well pads - Building - Urban area Permian basin, USA - Faint well pad - No building - Rural area © Pla
Slide 18: Silo Bag Detection
Planet Analytic Feeds | Performance Metrics April 2021 SILO BAG DETECTION Planet’s Silo Bag Detection leverages an object detection computer vision model applied to Planet Basemaps in Argentina. Object detection is a computer vision approach that analyzes an image and generates a “bounding box” arou
Slide 19: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors - Example of silo bags off the side of farmland - Edges of land features can resemble silo bags which can cause false positives Permian basin, USA - Multiple silo bags get detected within one detection © P
Slide 20: Road Change Detection
Planet Analytic Feeds | Performance Metrics April 2021 ROAD CHANGE DETECTION Planet’s Road Change Detection leverages temporal de-noising techniques applied to our road rasters . Since our road rasters are derived from our semantic segmentation models, which classify if a pixel belongs to either the
Slide 21: Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 Key Factors Our model will perform differently depending on key characteristics of the AOIs. Common false positives for detection of new roads include dirt roads, ephemeral streams, crop lines, dry creeks, roads with changing conditions such as
Slide 22: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Melbourne, Australia - Examples of new roads correctly detected in the outskirts of Melbourne Himalayas, Pakistan - A dry creek incorrectly detected as a new road on a rugged mountain terrain New Mexico, U
Slide 23: Building Change Detection
Planet Analytic Feeds | Performance Metrics April 2021 BUILDING CHANGE DETECTION Planet’s Building Change Detection leverages temporal de-noising techniques applied to our building rasters . Since our building rasters are derived from our semantic segmentation models, which classify if a pixel belon
Slide 24: Key Factors
Planet Analytic Feeds | Performance Metrics April 2021 Key Factors Our model will perform differently depending on key characteristics of the AOIs. Common false positives for detection of new buildings include mines, quarries, shipping containers, parking lots, existing buildings in varying shadows
Slide 25: Representative Examples
Planet Analytic Feeds | Performance Metrics April 2021 Representative Examples AOI Key Factors Doha, East Qatar - Examples of new buildings correctly detected in the harbor near Lusail Bridgetown, Australia - A mine / quarry incorrectly detected as a new building, due to shadows similarity to buildi
Slide 26: Building change detection false positive examples
Planet Analytic Feeds | Performance Metrics April 2021 Doha, Qatar - A docked ship incorrectly detected as a new building due to proximity to the pier and shadows similarity to buildings Shanghai, China - A shadow incorrectly detected as a new building to to changing shadows from tall buildings Hous
Slide 27: Performance Metrics Methodology
Planet Analytic Feeds | Performance Metrics April 2021 PERFORMANCE METRICS METHODOLOGY Planet Analytic Feeds leverages computer vision to transform our imagery into information feeds that detect and classify objects, identify geographic features, and understand change over time across the globe. To
Slide 28: Object Detection
Planet Analytic Feeds | Performance Metrics April 2021 Object detection Objects are typically represented as bounding boxes that can have any size and position. The model outputs box predictions along with confidence scores for each. The confidence score is a number between 0 and 1. To compute metri
Slide 29: Semantic Segmentation
Planet Analytic Feeds | Performance Metrics April 2021 Based on these matches, and given the minimum IOU and confidence thresholds, we calculate precision, recall and F1 metrics. Semantic segmentation Each pixel is assigned one class among a set of classes. Thus, truth and prediction datasets are co
Slide 30: Get in Touch
Planet Analytic Feeds | Performance Metrics April 2021 Change detection Similar to semantic segmentation, these binary classification metrics are based on class matches between pixels in ground truth and prediction datasets. Here, the classes are simply change or no-change. In addition, the time axi
Planet Labs PBC — Technical Presentation (April 2021) on 6ix