DJI Matrice 4E Maps 29 Native Species Across 130 Hectares In Open-Source Flora Study

A surveyor running an environmental baseline study used a DJI Matrice 4E and free open-source software to identify, classify, and geolocate every individual plant of 29 native species across a 130-hectare site. Pablo Carranza shared the work in the Drone Mapping and Processing group, posting annotated orthophotos that show hundreds of tree crowns outlined and labeled by species, each one tied to a coordinate.

The flight setup is the part most operators will recognize. Carranza flew the Matrice 4E at 45 meters (148 feet), low for a mapping mission, and said he picked the 4E over the cheaper Matrice 4T specifically because the 4T’s camera is not good enough for this job. He is right about the hardware difference. The 4E carries a 4/3 CMOS 20-megapixel wide-angle camera with a mechanical shutter built for surveying, while the 4T’s wide-angle sits on a smaller 1/1.3-inch sensor tuned for thermal-first public safety work.

The post drew 87 likes and 29 comments inside a day, with biologists, farmers, and surveyors asking how the species identification actually worked. The answer is a mix of fieldwork, machine learning, and a deliberately small training set.

A field biologist trained the model before the drone did the counting

The identification was not done from the air alone. A biologist visited the site, surveyed the local flora, collected physical samples, and built a catalog of the 29 species found there. That ground catalog became the labeled training data for a supervised machine-learning model, which then classified the individual plants visible in the drone imagery.

Carranza described the site visit as necessary because his team was unfamiliar with the local vegetation. That detail matters more than it sounds. The two species visible in his labeled orthophoto, Espinillo (Vachellia caven) and Moradillo (Schinus fasciculatus), together with the Spanish-language discussion in the comments, place the work in the South American Espinal scrub-forest, most likely Argentina. It is dry, xerophytic terrain where crowns overlap and many species share a similar gray-green signature from above. A model trained on the wrong region’s plants would have guessed.

When another group member repeated the common claim that AI plant recognition needs a hundred thousand photos, Carranza pushed back with a number from practice. One hundred thousand plants to train a model is too many, he wrote, and around 300 samples produce solid results as long as the environment does not change too much. That qualifier is the whole game in vegetation classification. A model tuned to one site’s lighting, phenology, and species mix does not travel well to the next valley over.

The workflow runs on free open-source tools

Carranza processed the drone photos in WebODM Lightning, the cloud-hosted version of the open-source OpenDroneMap photogrammetry toolkit, then cropped the resulting orthophoto while preserving its geolocation using image tiles in Python. The choice is notable. This is a professional-grade deliverable built on a free toolchain rather than Pix4D, DroneDeploy, or DJI Terra.

Because the orthophoto was georeferenced, every classified plant could be pinned to a coordinate by its centroid. That single step is what turns a pretty map into a dataset. Once each individual has a location, the spatial relationships between plants and between species become measurable: clustering, spacing, dominance, and distribution across the 130 hectares. Carranza said that geolocation is exactly what enabled the study of relationships between individuals and between species, which is the actual point of an environmental impact baseline.

He was candid about flight conditions too. The day was sunny, which casts shadows that complicate crown segmentation, and he noted that an overcast day with no wind would have been better. Anyone who has tried to map tree canopy in hard afternoon light knows the problem: shadow edges read as object edges, and the model has to learn to ignore them.

Onboard AI is arriving on the same airframe

The Matrice 4E that Carranza flew already ships with a built-in AI detection model, and DJI now offers model-training tools and a third-party developer certification path to run custom models on the drone’s onboard computing. His pipeline keeps the intelligence on the ground, in WebODM and Python after the flight. The hardware he used is one firmware generation away from doing some of that classification in the air.

This is the same pattern visible across DJI’s enterprise line, where the company has pushed onboard autonomy and AI down to its more affordable airframes through the Manifold 3 computing module and obstacle-sensing add-ons rather than reserving them for dock-deployed systems. The processing power to identify a plant in flight is no longer the bottleneck. The labeled training data from a biologist on the ground still is.

DroneXL’s Take

The detail that makes this project credible is the one most AI-mapping demos leave out: a human biologist walked the site, pulled samples, and built the catalog before any model touched the imagery. That is supervised learning done correctly, and it is the difference between a classifier you can defend in an environmental report and a colorful guess.

I have watched this category mature on DroneXL for years, from the early work on drones and AI spotting farmland bird nests to more recent deep-learning systems built for counting animals from the air and drone surveys mapping waterfowl habitat across Wisconsin’s wetlands. What stands out here is the cost structure. Carranza paired a roughly $5,000-class mapping drone, reviewed in depth when DJI launched the AI-powered Matrice 4 Series, with a free photogrammetry engine and a single field survey. Compare that to the seven-figure scanning operations DroneXL covered when a Matrice 400 and Zenmuse L3 mapped lost Maya cities in Guatemala. Two very different budgets, the same underlying move: putting a georeferenced dataset where there used to be guesswork.

The honest constraint is the one Carranza named himself. His 300-sample model works as long as the environment does not change too much. That is a real limit, not a footnote. It means this is a site-specific tool, not a general species classifier, and the fieldwork has to be repeated for each new region. Whether DJI’s onboard model-training tools eventually shrink that field-survey step, or whether the biologist’s catalog stays the irreplaceable input, is the open question this kind of project keeps raising. For now, the drone counts what the biologist taught it to see.

Source: Pablo Carranza, posted in the Drone Mapping and Processing group on Facebook.

DroneXL uses automated tools to support research and source retrieval. All reporting and editorial perspectives are by Haye Kesteloo.