Tesera Systems - How We Use AI-based Machine Learning

Overview

Our High Resolution Inventory Solutions (HRIS) leverages advanced AI and machine learning to deliver an Enhanced Forest Inventory (EFI) that includes tree species identification and a comprehensive array of attributes determined at an average resolution (analysis unit size) of 0.04 hectares / 0.1 acres. These analysis units (and attributes) can then be further aggregated into larger reporting units (of any size) based on client needs and specifications. HRIS also offers Individual Tree Inventory (ITI) capabilities, all integrated within a robust, vector-based enterprise inventory solution. 

We use LiDAR, multispectral imagery, ground plot data (along with terrain, soils and other available data) and AI-based machine learning and automation to generate hundreds of features. We create models for each attribute of interest, and we also provide predictive accuracy statistics for each and every attribute prediction. 

The collected ground plot data is used to train and test the AI system using data-splitting strategies (e.g. cross-validation). Land cover classification is performed using gradient boosted trees. Forest attributes are modelled using regularized linear regression. These models have been selected for their robustness and interpretability. Various state-of-the art feature engineering and feature selection routines are also used throughout.

Our AI system is trained independently for every single inventory project, and is customized to meet project-specific requirements. Whenever possible, we use freely available datasets (e.g. NAIP imagery, Sentinel-2, USGS LiDAR). Otherwise, the client purchases the necessary data from commercial vendors. Training data that was purchased and/or collected by the client is owned by the client. Client data may be used (with permission) for internal R&D projects, but it is never used for another project.

HRIS generates a point-in-time inventory, representing the current state of the forest. We also utilize machine learning to detect change over time using our Change Detection AI-based system to account for forest disturbances, land cover change, forest growth, etc. It is also possible to update the forest landscape and natural resource inventory by repeating the analysis with a new set of inputs.

A stepwise overview of the data sources, methodology, analytical procedures, detailed statistical summaries, and identification of anomalies, outliers and bias are openly reviewed with clients. Continual feedback, questions, insights related to bias and other concerns are encouraged and carefully considered as a means to assess the AI-machine learning methods employed, and for adapting, adjusting and improving methods and performance.