TerraCover - Terra Global Capital

Overview

TerraCover is a land-use and land-cover (LULC) classification tool designed to support accurate and consistent mapping of land-use changes to mitigate environmental changes. It enables users to interpret satellite imagery, classify ground cover types, and validate classifications using expert input and machine learning algorithms.

Accurate Classification: Enables high-accuracy LULC classification using expert-driven methodologies
Multi-Sensor Analysis: Leverages different remote sensing sources (MSI, LiDAR, and SAR) to enhance classification precision
Collaborative Validation: Allows user-driven classification and validation through sample-based visual outputs for continuous improvement
Scalable Mapping: Facilitates large-scale land cover mapping to support conservation and planning efforts

Acquisition of Remote Sensing/GIS Imagery

TerraCover supports the collection of high-resolution remote sensing and GIS data.

Retrieval of Multi-Source Imagery: Integrates satellite, aerial, and drone imagery
Historical & Real-Time Data: Allows comparison of land cover changes over time
Scalable Data Acquisition: Supports programs of varying geographic extents

Data Preprocessing and Preparation

Before classification, data is cleaned and structured to maximize data and computational efficiency while maximizing the accuracy of data products.

Standardized Data Processing: Applies a L2 spectral normalization to MSI and HSI data, compress the information into efficient data structures, minimize the noise-to-signal ratio in the data the variance in the data, and create multisensory data-cubes that leverage information from different imagery sources as scale
Spectral Band Clustering: Uses clustering techniques (Iso-cluster) and widely recognized indexes (NDVI, NDWI) to perform features augmentation for MSI and GRD imagery; uses dimensionality reduction (PCA, t-SNE, Autoencoders) to perform dimensionality reduction from sensors with large number of autocorrelated spectral bands (e.g. HSI)
Automated Data Correction: Minimizes georeferencing errors (Orthorectification) and inconsistencies in the spectral signatures (radiometric and atmospheric)

Wall-to-Wall Classification

TerraCover performs wall-to-wall land cover classification at scale using a variety of AI/ML architectures.

Machine Learning Classification: Adapts to different computational configurations by adopting different AI/ML architectures for landcover classification, including Deep Neural Networks (2D-CNN and ViT), traditional ML (Random Forests, Gradient Boosting, Neural Networks), multimodal or multi-architecture ensemble
Predictions Uncertainty Estimation: Adopts Probability Calibration to map the probability assigned to pixel classification to the real likelihood of the event, making a more reliable estimate of uncertainty
Feature Importance Assessment: Limited to less computationally demanding ML models, allows for calculating the relative importance of different data sources used for the classification
Custom Classification Schemes: Adapts to program-specific land cover categories
Seamless Integration with GIS Tools: Allows direct export for further analysis

Post-Classification

After classification, results are refined and validated to improve the accuracy of wall-to-wall maps.

Spatial Speckle Filter: A filter that identifies and reclassifies granular interference patterns that degrade image quality by replacing each pixel with the median of its neighborhood
Temporal Filter: A filter that smooths temporal patterns by adjusting the probability of a pixel conditioned to geophysical or temporal constraints, to avoid non-feasible transitions
Error Detection & Correction: Leverages the use of probability density maps to identify regions with higher uncertainty and focus new training to improve the model
Change Detection Analysis: Allows for tracking spatial shifts in land cover over time, and is used to extract data to train land-use change forecasting models
Uncertainty Estimation: Estimation of LULC transitions based on held-out visual interpretation of random samples of imagery stratified across different expected transition pools.