Water availability modulates maximum canopy heights of low-elevation Amazonian second-growth forests
The Big Picture: Second-growth tropical forests are vital carbon sinks. By leveraging NASA GEDI spaceborne lidar data across the Brazilian Amazon, our lab discovered that water availability plays a massive role in dictating the maximum canopy heights of these regenerating forests. Specifically, for 30-35-year-old forests at elevations under 500 meters, canopy heights peak alongside precipitation thresholds of up to 1500 mm. These insights help us understand how changing precipitation patterns and climate-induced droughts will impact the carbon sequestration superpowers of the Amazon.
Tropical Ecology (2026)
Mohan, M., Pastorello, G.Z., Feng, Y., Adrah, E., Keller, M., Ewane, E.E., Longo, M., Csillik, O., Ferraz, A., Roy, A.D., Meng, L., Chambers, J.Q.
Characterizing the Vertical Structure of Forests in the Brazilian Amazon
The Big Picture: Traditional satellite imagery often misses what happens below the top of the canopy. By analyzing massive airborne LiDAR datasets covering ~188,000 hectares, we mapped the true 3D vertical structure of the Brazilian Amazon. We discovered that forest fires disproportionately devastate the lowest strata (1–10 m), whereas selective logging and windthrows impact the canopy more evenly. We also found that while the understory of secondary forests recovers in just 5–10 years, the upper canopy (20–30 m) remains stunted even after 35 years of regrowth. Finally, we classified the Amazon into 5 distinct Forest Structural Types (FSTs), driven by proximity to roads, elevation, soil pH, and precipitation.
Communications Earth & Environment (2026)
Valle, D., Haneda, L., Brack, I., Ometto, J., Csillik, O., Longo, M., Keller, M., Almeida, D.
Thermal stress in degraded forests in the Brazilian Amazon Arc of Deforestation
The Big Picture: As climate change accelerates, degraded tropical forests are approaching critical temperature thresholds. Using ECOSTRESS thermal data and GEDI canopy metrics, our team discovered that burned forests in the Amazon are significantly hotter and more thermally volatile than intact forests. In the warmest burned patches, 87% of canopy leaves exceed the temperature where respiration surpasses photosynthesis (33.0°C). Worse, the risk of reaching temperatures that cause permanent leaf tissue damage and necrosis (49.9°C) is ten times higher in burned forests. Surprisingly, selectively logged forests regulate their canopy temperatures almost exactly as well as perfectly intact forests.
Environmental Research Letters (2025)
Cooley, S.S., Keller, M., Longo, M., Csillik, O., Dias, A.P., Silgueiro, V., Carvalho, R., Anderson, D., Gilbreath, M., Duffy, P., Adami, M., Cawse-Nicholson, K., Menge, D.N.L.
Degradation and deforestation increase the sensitivity of the Amazon Forest to climate extremes
The Big Picture: Around 40% of the Brazilian Amazon has already been altered by deforestation or structural degradation like selective logging and fires. By combining data from 545 airborne lidar transects with the Ecosystem Demography Model (ED2), our lab simulated how these structural changes affect the forest’s ability to handle climate extremes. We found that degraded, low-biomass forest patches suffer severe drops in gross primary productivity (GPP) and evapotranspiration (ET) during hot, dry conditions (high vapor pressure deficit). Intact, high-biomass forests, however, remain highly resilient to these same extreme droughts. Alarmingly, modeling a future “Degradation” scenario reveals that continued canopy loss will drastically expand the Amazon’s vulnerability to climate extremes, pushing massive new areas toward heat and water stress.
Environmental Research Letters (2025)
Longo, M., Keller, M., Kueppers, L.M., Bowman, K., Csillik, O., Ferraz, A., Moorcroft, P.M., Ometto, J., Soares-Filho, B.S., Xu, X., Assis, M.L.R., Görgens, E.B., Larson, E.J.L., Needham, J., Ordway, E.M., Pereira, F.R.S., Rangel Pinagé, E., Sato, L.Y., Xu, L., Saatchi, S.S.
Environmental drivers of spatial variation in tropical forest canopy height: Insights from NASA’s GEDI spaceborne LiDAR
The Big Picture: What limits the height of the world’s tropical forests? Using high-resolution data from NASA’s GEDI spaceborne LiDAR mission, we discovered that just three factors—climate, topography, and soil properties—account for 75% of the spatial variation in tropical canopy heights globally. Of these, Elevation, Dry Season Length (DSL), and Solar Radiation are the absolute most dominant drivers. For example, forest canopies remain relatively stable during dry seasons up to 4 months long, but if the dry season extends beyond that threshold, average canopy height drops significantly by about 8 meters. We also mapped structural differences across continents, revealing that Southeast Asia boasts the tallest tropical canopies (averaging 33.0 m), followed by Africa (29.6 m) and the Amazon (27.0 m).
Proceeding of the National Academy of Sciences (2025)
Liu, S., Csillik, O., Ordway, E.M., Chang, L.-L., Longo, M., Keller, M., Moorcroft, P.R.
A large net carbon loss attributed to anthropogenic and natural disturbances in the Amazon Arc of Deforestation
The Big Picture: The Brazilian Amazon is currently releasing more carbon than it absorbs. By deploying repeated, high-resolution airborne laser scanning (lidar) across the Arc of Deforestation, our team achieved an unprecedented mapping of the forest’s carbon budget. The findings reshape how we view forest threats: degradation (such as selective logging and fires) now drives more carbon loss than outright deforestation (clearing). While forest growth claws back 44.1 Tg of carbon per year, it is overwhelmed by multi-front losses, resulting in a staggering net carbon loss of 90.5 +-16.6 Tg annually for the region. Crucially, the data proves that territorial protection works—indigenous territories and conservation units remain highly effective shields against these anthropogenic carbon losses.
Proceeding of the National Academy of Sciences (2024)
Csillik, O., Keller, M., Longo, M., Ferraz, A., Pinagé, E.R., Görgens, E.B., Silgueiro, V., Ometto, J.P., Saatchi, S.

Featured in 99 News Stories
from 91 Outlets


2 Policy Documents


#8 of 777 outputs of similar age from PNAS

Effects of forest degradation classification on the uncertainty of aboveground carbon estimates in the Amazon
The Big Picture: When scientists estimate how much carbon a forest holds, they usually map the forest type from space and then estimate the biomass. But what if the map is slightly wrong? By combining very high-resolution PlanetScope satellite imagery, machine learning, and airborne lidar across three Amazonian sites, our lab proved that ignoring map uncertainty leads to dangerous overestimations of forest carbon. Because degraded forests are messy—a logged forest might also have hidden fire damage—factoring in the probability of misclassification actually dragged our final mean carbon density estimates significantly downward in heavily degraded areas. This highlights a critical blind spot in how global carbon budgets are currently calculated.
Carbon Balance and Management (2023)
Rangel Pinagé, E., Keller, M., Peck, C.P., Longo, M., Duffy, P., Csillik, O.
Surface energy dynamics and canopy structural properties in intact and disturbed forests in the Southern Amazon
The Big Picture: A healthy Amazon forest acts like a giant air conditioner, using its canopy to sweat (evapotranspiration) and cool the surrounding environment. But what happens when that canopy is damaged? By fusing NASA GEDI structural data with thermal satellite imagery from Landsat and MODIS, our lab tracked the “sweat” and surface temperatures of intact, logged, and burned forests in the Southern Amazon. We found a massive divergence during the dry season: while intact and selectively logged forests effectively maintain their cooling systems, burned and secondary forests suffer severe water stress, plummeting evapotranspiration, and spiking surface temperatures. Crucially, we discovered that the physical 3D structure of the canopy exerts a much stronger control over these water and energy fluxes in heavily disturbed forests than it does in intact ones, highlighting how fragile degraded ecosystems become under changing climates.
Journal of Geophysical Research: Biogeosciences (2023)
Rangel Pinagé, E., Bell, D.M., Longo, M., Silva, C.A., Csillik, O., Huete, A.
Rapid remote monitoring reveals spatial and temporal hotspots of carbon loss in Africa’s rainforests
The Big Picture: African rainforests are losing carbon, but this destruction isn’t a slow, steady leak—it happens in massive, predictable bursts. By combining all-weather radar data (which peers right through dense tropical clouds) with spatial carbon estimates, our lab mapped carbon loss across 23 African countries at a high 10m resolution. We discovered intense “temporal hotspots”—short windows tied to the dry season when remote forests suddenly become accessible to smallholder agriculture and logging. In the Central African Republic, a staggering 75.7% of the entire year’s carbon loss happened in just three months (January to March 2020). Meanwhile, Madagascar lost 89% of its annual carbon within the final five months of the same year. This near-real-time monitoring proves that if we know when the deforestation spikes will hit, we can finally deploy enforcement exactly when and where it matters most.
Communications Earth & Environment (2022)
Csillik, O., Reiche, J., De Sy, V., Araza, A., Herold, M.
Phenology-based sample generation for supervised crop type classification
The Big Picture: Training machine learning algorithms to map agricultural crops usually requires massive amounts of expensive, hand-collected field data. To bypass this bottleneck, our team developed a fully automated method that generates its own training samples by looking at the “phenology”—the seasonal growth timeline—of crops using Sentinel-2 satellite imagery. By analyzing how a crop greens up and dies back over the year using Time-Weighted Dynamic Time Warping (TWDTW) and Random Forests, we successfully mapped new target areas without any local ground labels. The algorithm excelled at identifying crops with distinct growing seasons, achieving 98% user accuracy for sugar beets and grains. However, it also revealed a core challenge in remote sensing: crops that share identical growth timelines, like potatoes and maize, are highly prone to algorithm confusion.
International Journal of Applied Earth Observation and Geoinformation (2021)
Belgiu, M., Bijker, W., Csillik, O., Stein, A.
Near-real time aboveground carbon emissions in Peru
The Big Picture: Traditional carbon accounting often relies on slow, yearly estimates that obscure the actual timing of forest loss. To fix this, we built a deep learning framework that tracks aboveground carbon emissions (ACE) in Peru at an unprecedented 1-hectare spatial resolution on a quarterly and even monthly basis. By feeding millions of hectares of airborne LiDAR data, daily Planet Dove satellite imagery, and Sentinel-1 radar data into our neural networks, we discovered that Peru emitted 20.08 Tg of carbon between the third quarters of 2017 and 2018—a figure 23.4% higher than estimates from the FAO Global Forest Resources Assessment. Furthermore, breaking this down chronologically revealed massive seasonal spikes, such as a localized burst of 11.5 Tg C lost in just the fourth quarter of 2017 alone. This high-frequency tracking allows us to pinpoint exactly when and where carbon is bleeding from tropical ecosystems, right down to the vulnerabilities surrounding protected national parks.
PLoS ONE (2020)
Csillik, O., Asner, G.P.
Challenges in estimating tropical forest canopy height from Planet Dove imagery
The Big Picture: Mapping the height of tropical forests from space usually involves a tough compromise: you can have high-resolution images, or you can have frequent updates, but rarely both. We tackled this by building a country-scale machine learning toolchain that estimates the Top-of-Canopy Height (TCH) for the entirety of Peru at an incredibly sharp 1-hectare resolution. By feeding textures from the daily-revisiting Planet Dove satellite constellation and SRTM elevation data into a customized Gradient Boosted Regression algorithm, we trained the model against actual airborne LiDAR. The result is a highly generalizable model that predicts Peruvian forest heights with a root mean square error (RMSE) of just 4.36 meters. Crucially, the model’s relative error drops significantly for the tallest trees (20–25m), proving highly effective where the vast majority of the forest’s carbon is stored.
Remote Sensing (2020)
Csillik, O., Kumar, P., Asner, G.P.
Aboveground carbon emissions from gold mining in the Peruvian Amazon
The Big Picture: Beneath the immense biodiversity of the Peruvian Amazon lies a layer of gold-rich alluvium, sparking a modern gold rush that pits short-term artisanal mining against the long-term survival of the rainforest. To measure the true cost of this illicit extraction, our lab deployed a deep learning model fusing airborne LiDAR with high-cadence Planet Dove satellite imagery across 750,000 hectares of the Madre de Dios region. We mapped the area at a stunning 1-hectare resolution and found that the forest is bleeding carbon at an alarming rate: 1.12 Teragrams of carbon were lost in a single year (2017-2018), destroying over 23,600 hectares of forest. This isn’t just a remote problem; the data reveals that illegal mining is weaponizing infrastructure, with nearly all emissions occurring within 5 kilometers of the Interoceanic Highway and major rivers, actively eating into the buffer zones of critical protected areas like the Tambopata National Reserve.
Environmental Research Letters (2020)
Csillik, O., Asner, G.P.
Monitoring tropical forest carbon stocks and emissions using Planet satellite data
The Big Picture: Accurately measuring how much carbon a nation’s forests hold—and how fast it is being lost—is essential for global climate initiatives like REDD+. To create a near-real-time monitoring system, our lab mapped the Aboveground Carbon Density (ACD) of the entire country of Peru at an unprecedented 1-hectare resolution. By running millions of hectares of airborne LiDAR data and high-frequency Planet Dove satellite imagery through a Random Forest machine learning workflow, we calculated that Peru’s forests harbor 6.928 Pg of carbon. However, the data exposes a massive vulnerability: only 2.9 Pg of that carbon is secured inside protected areas or ecological buffers. Between 2012 and 2017 alone, we tracked 0.096 Pg C in direct emissions driven by aggressive oil palm and cacao expansion, urbanization, and illegal gold mining.
Scientific Reports (2019)
Csillik, O., Kumar, P., Mascaro, J., O’Shea, T. and Asner, G.P.
Object-based time-constrained dynamic time warping classification of crops using Sentinel-2
The Big Picture: Mapping exactly what is growing across vast agricultural landscapes is incredibly difficult because satellite imagery is often obstructed by clouds, and different crops grow at different speeds depending on the weather. We solved this by bringing a speech-recognition algorithm called Dynamic Time Warping (DTW)—which matches audio patterns even if someone speaks too fast or too slow—into the realm of satellite crop mapping. By looking at a full year of Sentinel-2 satellite imagery over highly complex farms in California and Texas, our object-based DTW model learned the unique “growth rhythms” (phenology) of various crops. Because crops don’t grow completely out of season, we discovered that forcing the algorithm to adhere to “time constraints” of 30 or 45 days drastically improved its ability to correctly identify crops, achieving up to 89.1% accuracy without the need for massive amounts of traditional training data.
Remote Sensing (2019)
Csillik, O., Belgiu, M., Asner, G.P., Kelly, M.
Identification of Citrus Trees from Unmanned Aerial Vehicle Imagery Using Convolutional Neural Networks
The Big Picture: Precision agriculture increasingly relies on high-resolution drones to monitor crop health, but manually counting thousands of individual trees from these images is impossibly tedious. Our lab automated this process by training a Convolutional Neural Network (CNN) to detect citrus trees from 12.8 cm-resolution drone imagery in California’s San Joaquin Valley. However, AI models often struggle with large trees, confusing a single large, textured canopy for multiple smaller trees. To fix this “multiple crown” error, we added a post-processing step using a Simple Linear Iterative Clustering (SLIC) algorithm, which groups the image into “superpixels”. This refinement forced the algorithm to recognize symmetrical, circular groupings as single trees, skyrocketing the overall accuracy of the system to 96.24%.
Drones (2018)
Csillik, O., Cherbini, J., Johnson, R., Lyons, A., Kelly, M.
An automatic Random Forest-OBIA algorithm for early weed mapping between and within crop rows using UAV imagery
The Big Picture: In precision agriculture, catching weeds early is everything. While drones have gotten great at spotting weeds growing between crop rows, finding weeds hiding inside the crop row has remained a massive challenge. Our lab developed a fully automated Object-Based Image Analysis (OBIA) algorithm that solves this by teaching the drone to think in 3D. By generating a high-resolution Digital Surface Model (DSM) from UAV imagery, the algorithm automatically calculates the average height of the crop. It then uses a Random Forest machine learning classifier to identify any vegetation inside the row that falls below that height threshold as a weed. Tested on sunflower and cotton fields, the system proved incredibly accurate, allowing us to generate custom “prescription maps” that tell tractors exactly where to spray—resulting in herbicide savings of up to 79%.
Remote Sensing (2018)
De Castro, A., Torres-Sánchez, J., Peña, J., Jiménez-Brenes, F., Csillik, O., López-Granados, F.
Sentinel-2 cropland mapping using pixel-based and object-based time-weighted dynamic time warping analysis
The Big Picture: In order to manage global food security, we need reliable maps of what crops are growing where. But satellite imagery is often obstructed by clouds, and farmers plant and harvest on unpredictable schedules. To tackle this, we tested an algorithm called Time-Weighted Dynamic Time Warping (TWDTW) across three massive agricultural test areas in Romania, Italy, and California using Sentinel-2 imagery. TWDTW is brilliant because it doesn’t just look at a single snapshot in time; it learns the “growth rhythm” of a crop and can match it even if the timing is distorted. We proved that by grouping pixels into whole field “objects” before running the algorithm, we could eliminate messy “salt-and-pepper” classification errors and boost accuracy up to 96.19%, all while drastically cutting processing time.
Remote Sensing of Environment (2018)
Belgiu, M., Csillik, O.
Fast segmentation and classification of very high resolution remote sensing data using SLIC superpixels
The Big Picture: When responding to natural disasters like earthquakes or floods, emergency managers need high-resolution satellite maps immediately, but traditional Object-Based Image Analysis (OBIA) takes hours to process millions of individual pixels. To shatter this bottleneck, we introduced “superpixels”—specifically the Simple Linear Iterative Clustering (SLIC) algorithm—to remote sensing workflows. Instead of analyzing an image pixel by pixel, the algorithm instantly groups similar pixels into perceptually meaningful “atomic regions” or superpixels before the heavy processing begins. The results were staggering: processing a massive 12.2-million-pixel WorldView-2 satellite image took 5 hours and 35 minutes using traditional methods, but only 13 minutes using our SLIC superpixel approach. Crucially, feeding these superpixels directly into a Random Forest machine learning classifier maintained or even slightly improved the overall geometric accuracy of the maps.
Remote Sensing (2017)
Csillik, O.
Flash-flood potential: a proxy for crayfish habitat stability
The Big Picture: Freshwater crayfish are excellent indicators of a healthy ecosystem, but because they live in shallow waters, they are highly vulnerable to being washed away by severe floods. To understand how they survive, our team developed a novel digital metric called “Flash-Flood Potential” (FFP). Using Digital Elevation Models (DEMs), FFP calculates the risk of erosive floods by multiplying the local slope of a river by the steepness of the entire upstream catchment that feeds into it. By testing this metric across 243 rivers in the Romanian Carpathians, we found that as FFP increases, crayfish populations drastically shrink. Ultimately, we discovered a strict hierarchy of survival: during intense flash floods, only the largest, dominant crayfish survive because they fight for and occupy the deepest, most “durable” rock crevices, leaving smaller crayfish in unstable shelters to be swept downstream.
Ecohydrology (2016)
Pârvulescu, L., Zaharia, C., Groza, M.I., Csillik, O., Satmari, A., Drăguţ, L.
Transformation (normalization) of slope gradient and surface curvatures, automated for statistical analyses from DEMs
The Big Picture: In environmental science, powerful statistical analyses (like least-squares regression or principal component analysis) generally assume your data follows a neat, symmetrical bell curve (a “normal” distribution). The problem? Nature rarely cooperates. Terrain features extracted from Digital Elevation Models (DEMs)—like how steep a slope is or how the ground curves—are notoriously messy, featuring massive spikes and extreme “long tails” that break parametric statistical models. To fix this, we built a fully automated algorithm that mathematically massages this wild topography data into perfect bell curves. Instead of relying on manual trial-and-error, our tool evaluates the terrain’s specific geometry and automatically iterates Box-Cox and arctangent transformations until statistical outliers are neutralized. The result is a plug-and-play tool for Earth scientists to safely run complex statistics on landscape data without fear of geometric distortions.
Geomorphology (2015)
Csillik, O., Evans, I.S., Drăguț, L.
Automated parametrization for multi-scale image segmentation on multiple layers
The Big Picture: Analyzing high-resolution satellite imagery relies heavily on breaking an image down into meaningful “objects”—like separating buildings from forests. However, the tool most commonly used to do this, Multiresolution Segmentation (MRS), requires scientists to manually guess the “Scale Parameter” (SP) through tedious trial-and-error, making the process highly subjective. Our team created a tool that completely automates this decision. By analyzing the “local variance” across multiple spectral layers of an image, our tool automatically detects the natural, hierarchical scales of the landscape. Tested across environments ranging from a refugee camp in Sudan to mixed residential zones in Austria, the tool successfully self-tuned to identify small features (like individual huts) at a fine scale, and massive features (like large agricultural fields) at a broader scale, without any human intervention.
ISPRS Journal of Photogrammetry and Remote Sensing (2014)
Drăguț, L., Csillik, O., Tiede, D., Eisank, C.
