Expansion of agricultural areas is the largest contributor to deforestation worldwide. This is especially the case in the tropics where large swaths of forest are cleared for soy, palm oil, cocoa and other soft commodities.

Consumer goods manufacturers are incentivized to decrease deforestation within their supply chains by public pressure and campaigning NGOs. More recently, new legislation from e.g. the European Union obliges companies to stop sourcing from areas where deforestation has happened since a specified cut-off date.

In order for these companies to stop deforestation within their supply chains, two key pieces of information are required. First, companies need to know in detail their full supply chain, including all growers as well any potential traders in between. Second, they need to know if any (recent) deforestation is linked to any of the growers they source from. Both prerequisites represent a challenge for the vast majority of manufacturers.

While the latter problem has become increasingly manageable, due to recent advances in available satellite imagery, open source algorithms and cloud computing, knowledge of the full supply chain remains a complex problem. For many commodities, actual links between farms, production facilities and, ultimately, the consumer goods manufacturer’s products are unknown. In many cases, only the sourcing regions are known. This is a critical issue, since companies are now faced with possible sanctions if they cannot prove that their commodities are sourced from deforestation-free areas.

To address these challenges, at Satelligence we developed a combination of methods where we look at a combination of open data, multi-sensor satellite data, machine learning and auxiliary data to construct complete overviews of companies’ supply chains. These supply chains include smallholder areas for commodities such as palm oil and cocoa.

We will present some of the methods that we use to construct these supply chain overviews, and their related deforestation, and we will demonstrate how we scale these methods to the entire globe.

Assessing land-use following deforestation is vital for reducing emissions from deforestation and forest degradation. However, the assessment of land-use is often rather generic and at coarse levels and does not fully differentiate land-use from land-cover. The lack of land-use/cover differentiation generates confusion when forest cover change statistics derived from satellite imagery are compared directly against land-use statistics reported by governments in their national statistics. This is especially true in the pantropics, where forest loss is increasing each year. Unfortunately, data on the causes of forest loss are limited, due to a lack of accurate methods and data. Using deep learning, we can unlock the potential of satellite images to assess and monitor changes and report on proximate drivers of forest loss. In this work, we present our research results, where we assess the potential of using deep learning methods, high-resolution, medium-resolution, single-date, and time series of satellite imagery to assess the proximate drivers of forest loss (land-use following deforestation) at the pantropical and national scale (Ethiopia and Ivory coast).

First, at a pantropical scale, we assess the potential of spatial, temporal and spatio-temporal deep learning methods for large-scale classification of land-use following tropical deforestation using dense satellite time series over six years on the pan-tropical scale. Here, Our main assumption is that the choice of type of deep learning model is crucial to increase the classification performance on land-use following deforestation or direct drivers of deforestation. Based on an extensive reference database of six forest to land-use conversion types, we find that the spatio-temporal models achieved a substantially higher accuracies than models that account only for spatial or temporal patterns. Although all models performed better when the scope of the problem was limited to a single continent, the spatial models were more competitive than the temporal ones in this setting.

Second, on a national scale, we expanded our research to assess the potential of different satellite data modalities (single-date, multi-date, multi-resolution, and an ensemble of multi-sensor images) for classifying land use following deforestation in Ethiopia using deep learning. The experiments were conducted on a real-world dataset acquired from freely available Planet NICFI, Sentinel-2, and Landsat satellite data. The experiments aimed at an analysis of (a) single-date images from individual sensors to account for the differences in spatial resolution between image sensors in detecting land-uses, (b) Ensemble of images from different sensors (Planet NICFI/sentinel/Landsat) with different spatial resolution to account for the benefit of integrating multiple satellite data across scales in detecting land-uses, and finally, (c) the use of multi-date data to account for the contribution of temporal information in detecting land-uses. Our main assumption is that the choice of satellite imagery (resolution) type is crucial to increase the classification performance on land-use following deforestation or direct drivers of deforestation.

Based on an extensive visually interpreted reference dataset of eleven types of post-deforestation land-uses, we find that classification of land-use following deforestation from higher-resolution, an ensemble based on multi-sensor imagery (Planet NICFI, sentinel-2, and Landsat-8), and multi-date imagery achieved substantially higher F1-score accuracies than single date medium-resolution images. The F1-score accuracy for medium-resolution images was enhanced when incorporating multi-date images as opposed to only using a single-date image. Its performance was equally good as the high-resolution ones in this setting. We also found that an ensemble of high and medium-resolution images produces substantially higher F1-score accuracies than a single sensor image. These findings suggest that either high detailed spatial patterns or detailed temporal patterns are required for identifying land-use following deforestation with medium-resolution single-date imagery not being sufficient. It also suggests that an ensemble of multiple satellite sensors provides more useful spatial information than a single sensor alone.

We demonstrate the potential use of high and medium resolution satellite images derived from Planet-NICFI, Sentinel-2, and Landsat sensors, in concert with deep learning, as a powerful tool for an efficient and automated assessment of land-use following deforestation in a scalable and cost-effective manner. Our results and approaches also paves the way for national scale land-use monitoring at the local level to support national and regional forest conservation policy development, and reporting.

Côte d'Ivoire and Ghana are the largest producers of cocoa in the world. In recent decades, vast tracts of forest areas have been converted into cocoa plantations. Efficient and accurate methods for remotely identifying cocoa plantations are essential to the implementation of sustainable cocoa practices and for the periodic and effective monitoring of forests. In this study, a method for cocoa plantation identification was developed based on a multi-temporal stack of Sentinel-1 and Sentinel-2 images and a multi-feature Random Forest (RF) algorithm. The Normalized Difference Vegetation Index (NDVI) and second-order texture features were assessed for their importance in an RF classification, and their optimal combination was used as input variables for the RF model to identify cocoa plantations in both countries. The RF model-based cocoa map achieved 82.89% producer's and 62.22% user's accuracy, detecting 3.69 million hectares (Mha) and 2.15 Mha of cocoa plantations for Côte d'Ivoire and Ghana, respectively. The results demonstrate that a combination of an RF model and multi-feature classification can distinguish cocoa plantations from other land cover/use, effectively reducing feature dimensions and improving classification efficiency. The results also highlight that cocoa farms largely encroach into protected areas (PAs), as 20% of the detected cocoa plantation area is located in PAs and almost 70% of the PAs in the study area house cocoa plantations. This study highlights the urgent need for governments, private companies and cocoa buyers to address both the distal and the proximal causes of cocoa-related deforestation. Other policy applications of the produced cocoa map include further environmental and socio-economic studies regarding the productivity, the quality and the sustainability of Ivorian and Ghanaian cocoa in both regions. This study also demonstrates a successful method to map cocoa farms at national level and shows potentials to be upscaled temporally and spatially.

Much concern about tropical deforestation focuses on oil palm plantations, but their impacts remain poorly quantified. Using nation-wide interpretation of satellite imagery, and sample-based error calibration, we estimated annual expansion of large-scale (industrial) and smallholder oil palm plantations and their impacts on forests from 2001 to 2019 in Indonesia, the world’s largest palm oil producer.
Over nineteen years, the area mapped under oil palm doubled, reaching 16.24 Mha in 2019 (64% industrial; 36% smallholder), more than the official estimates of 14.72 Mha, but less than our omission-adjusted estimate: 18.83 Mha (CI: 18.30-19.36). Replacement by oil palm expansion accounted for nearly one-third (2.85 Mha) of forest losses in this period (9.79 Mha). Industrial plantations replaced more forest than detected smallholder plantings (2.13 Mha vs 0.72 Mha). New plantations peaked in 2009 and 2012 and declined thereafter. Expansion of industrial plantations and forest loss were correlated with palm oil prices. A price decline of 1% was associated with a 1.08% decrease in new industrial plantations and with a 0.68% decrease of forest loss. Deforestation fell below pre-2004 levels in 2017-2019 providing an opportunity to focus on sustainable management. If prices rise, effective regulation will remain key to minimising deforestation.

Agricultural expansion is the main driver of deforestation in the Brazilian Amazon. It causes substantial losses of a unique biodiversity and severe emissions of greenhouse gases. In addition, post-deforestation land use often uses agricultural fires that frequently cause severe degradation in adjacent forests. In the past, national and international efforts reduced deforestation in the Brazilian Amazon and aimed on the prevention of agricultural fires. However, since 2016 the political developments in Brazil lowered environmental regulations and law enforcement (de Area Leão Pereira et al. 2020). Increasing rates of deforestation and fires have been reported and its intensity attracted the attention of international media. Nevertheless, it often remains unclear how different stakeholders contribute to the observed changes in deforestation and burning. In particular for fragmented landscape with smallholder agriculture the spatial resolutions of many fire- or burned area products is too coarse to differentiate between fires used for deforestation and those used for agricultural land management.
To better address this knowledge gap, we assessed the occurrence of deforestation and burned areas in Pará state. Specifically, the agricultural frontier of the Novo Progresso region is known for its high deforestation rates and raised media attention in 2019, when agricultural producers promoted a “Day of Fire” in response to president Bolsonaro’s call to use the Amazon more economically (Caetano 2021).

We used the Framework for Operational Radiometric Correction for Environmental monitoring (FORCE, Frantz 2019) to process an integrated time series of all Landsat and Sentinel-2 observations between January 2014 and October 2020. We derived clear observation sequences (COS), classified the COS with a Random Forest and aggregated class probabilities into annual COS scores (Jakimow et al. 2018) that were then translated into annual maps of burned area and land-cover. Based on these maps we assessed the average treatment effect of the last three Brazilian presidencies (2011 – 2016 Dilma Rousseff, 2016-2018 Michel Temer, since 2019 Jair Bolsonaro) on deforestation and burning. We used propensity score weighting to account for confounding variables like the slope and distance to road and then fitted and compared deforestation and burning between different land use zones and four size classes of rural properties.
Our results showed a four-fold increase of deforestation from 2017 to 2020, i.e. after the impeachment of President Rousseff. Conservation areas showed significantly less deforestation than non-designated areas and agrarian settlements. Alarmingly this effect has been significantly lowered during the presidencies of Temer and even more so Bolsonaro. Strikingly, deforestation rates per property increased in particular on larger properties.

The burned area mapped was largest in 2017 (4,343 km²) and 2020 (4,805 km²), but the majority (> 80 %) of it was mapped on land that had been already deforested. The relative fraction that was burned compared to the individual property sizes was highest in agrarian settlement projects and smallholder properties. More broadly, our approach shows the suitability of combining maps of land-use changes and burned areas with medium to high spatial resolution with quasi-experimental methods for improving our understanding of how political changes influence the land-use of different agricultural stakeholders.


REFERENCES
Caetano, M.A.L. (2021). Political activity in social media induces forest fires in the Brazilian Amazon. Technological Forecasting and Social Change, 167
de Area Leão Pereira, E.J., de Santana Ribeiro, L.C., da Silva Freitas, L.F., & de Barros Pereira, H.B. (2020). Brazilian policy and agribusiness damage the Amazon rainforest. Land Use Policy, 92
Frantz, D. (2019). FORCE—Landsat + Sentinel-2 Analysis Ready Data and Beyond. Remote Sensing, 11
Jakimow, B., Griffiths, P., van der Linden, S., & Hostert, P. (2018). Mapping pasture management in the Brazilian Amazon from dense Landsat time series. Remote Sensing of Environment, 205, 453-468

Agricultural extensification at the expense of natural vegetation loss is a central environmental issue in the tropics as well as a major source of greenhouse gas emissions. A prominent goal of policies mitigating climate change and biodiversity loss is to achieve zero-deforestation in the global supply chain of key commodities, such as palm oil and soybean. However, the extent and dynamics of deforestation driven by commodity expansion are largely unknown. Here we mapped annual soybean extent at 30-meter spatial resolution over the southern hemisphere of South America from 2000 to 2019. Our study area encompasses all major biomes where soybeans are cultivated: Amazonia, Atlantic Forest, Cerrado, Chaco, Chiquitania, Pampas, and more recently, Pantanal and Caatinga biomes. Our maps were produced using wall-to-wall Landsat and Moderate Resolution Imaging Spectroradiometer satellite data, a stratified random sample of Sentinel 2 satellite data, and three years of continent-wide field observations. We assessed the accuracy of the maps using data collected from field visits obtained at sample pixels selected by a stratified two-stage cluster sampling design. To understand the shifting dynamics of land-use change in South America in the 21st century we quantified the amount of primary forests, non-primary forests, non-forest natural vegetation, pre-existing croplands and pastures that were replaced by soybean. We also integrated the annual soybean maps with annual forest loss maps, and quantified deforestation caused by soybean as a direct and latent driver, highlighting emerging hotspots of soybean expansion as a direct driver of deforestation.

Our results showed that from 2000-2019, the area cultivated with soybean in South America more than doubled from 26.4 Mha to 55.1 Mha. Most soybean expansion occurred on pastures originally converted from natural vegetation for cattle production. The most rapid expansion occurred in the Brazilian Amazon, where soybean area increased more than 10-fold, from 0.4 Mha to 4.6 Mha. Across the continent, 9% of forest loss was converted to soybean by 2016. Soy-driven deforestation was concentrated at the active frontiers, nearly half located in the Brazilian Cerrado. Our results suggest that efforts to limit future deforestation must consider how soybean expansion may drive deforestation indirectly by displacing pasture or other land uses. The targeting of single commodities and single geographies for monitoring omits leakage effects, inter-commodity transitions, and land banking, all of which may result in concurrent increased forest loss and increased soybean cultivated area. Holistic approaches that track land use across all commodities coupled with vegetation monitoring are required to maintain critical ecosystem services.

The 30-meter resolution South America soybean map product is being updated on an annual basis, and freely available at: https://glad.earthengine.app/view/south-america-soybean.