ENVI Deep Learning

ENVI Deep Learning is a commercial add-on module for NV5 Geospatial's ENVI image analysis software that brings deep learning to remotely sensed imagery without programming. Training useful models usually needs data-science skills and code; the module instead provides a guide map and labeling tool in which analysts label examples with regions of interest, annotations or COCO datasets, then train models with a few clicks.

Version 4.0 trains ONNX models with seven built-in architectures for pixel segmentation (SegUNet++, SegUNet, DeepLabV3+), object detection (RT-DETRV2 Small and Large) and grid detection (ResNet50V2, ResNet101V2), and can also run ONNX models trained elsewhere, for example in PyTorch or TensorFlow. ENVI's preprocessing (calibration, atmospheric correction, colour transforms) prepares consistent inputs, and users can add spectral indices or elevation layers to strengthen classifiers. Outputs include classification rasters, class-activation maps, ROIs, shapefiles and GeoJSON.

The module supports terrestrial, UAV, airborne and spaceborne imagery. NV5 shows it extracting new flood water from Sentinel-1 SAR data and detecting aircraft in 0.5 m SkySat imagery. Korea University used it with Sentinel-2 seasonal composites to map coniferous and broadleaf forest across South Korea for national greenhouse-gas reporting, reaching 83.13% overall accuracy. Built on the ENVI Task framework, trained classifiers can run on desktops, on-premises servers or in the cloud; a Runtime edition runs inference without a separate Deep Learning licence, and models can be shared through the Analytics Repository. It runs on Windows and Linux; training calls for an NVIDIA CUDA GPU, while classification can run on GPU or CPU.