Elsevier

Ecological Informatics

Volume 66, December 2021, 101454
Ecological Informatics

Towards the automatic monitoring of deforestation in Brazilian rainforest

https://doi.org/10.1016/j.ecoinf.2021.101454Get rights and content

Abstract

Deforestation is considered one of the main environmental threats to the ecological balance on the planet. At the same time, monitoring changes in forest cover is a major challenge, especially in Brazil, a country with continental dimensions, a vast coverage of tropical forests, and an accelerated ongoing process of illegal deforestation. This study aims to propose and present an integrated automatic methodology for monitoring changes in forest cover, to enable “near real-time” monitoring of vast territorial extensions. Based on the application of Fully Convolutional Neural Networks (FCNs) combined with a logistic growth model, the methodology is aimed at allowing accurate detection of changes in forest cover based on the multitemporal assessment of satellite images. The results show that the combination of the two approaches make the methodology able to pinpoint deforestation processes. The applicability of the methodology is demonstrated for the Amazon and Atlantic Rainforest biomes, which are important areas of tropical forests in Brazil. In addition to enabling the agile and accurate identification of forest cover losses and providing efficient computing, the comparative results show that the methodology can be applied to issue alerts of suspected deforestation activity in standalone automatic monitoring systems, and also as a complementary tool to existing systems currently under operation.

Introduction

Deforestation is one of the main threats to environmental and ecological balance on the planet, as well as one of the primary sources of greenhouse gas emissions (de Bem et al., 2020, Le Quéré et al., 2018, Goulart et al., 2013). In Brazil, the illegal removal of vegetation represents a worrying environmental problem that affects all regions of the country, in particular the Amazon and Atlantic Forest biomes, which are predominantly constituted by tropical forests. The Atlantic Forest biome, for example, has suffered significant degradation of its vegetation cover, to the point that less than 15% of its original forest area remained (Dean, 1997, Wagner et al., 2020). According to Ribeiro et al. (2009), more than 80% of the fragments remaining in the Atlantic Forest were <50 ha, and further, natural reserves protected only 9% of the remaining forest, and 1% of the original forest. The annual deforestation report referring to the year 2019 identified the loss of 12,187 km2 of vegetation cover throughout the Brazilian territory (MapBiomas, 2020). In addition to the direct environmental and economic consequences, deforestation has also caused indirect economic damage due to several initiatives that have suspended the import of Brazilian products from activities under suspicion of contributing to illegal deforestation (Sousa, 2016, Assunção and Rocha, 2019, Arruda et al., 2019).

Given this scenario, the generation of information that allows the monitoring of forest resources at the most different scales of interest and spheres of activity is fundamental to enable the adoption of effective measures to mitigate and control illegal deforestation. Several deforestation mapping methodologies based on the use of multispectral satellite images have been proposed and are used in current ongoing projects in Brazil (Coutinho et al., 2013, MapBiomas, 2020, INPE, 2008, Diniz et al., 2015). Since 1988, by means of PRODES (Program for the Monitoring of the Amazon Forest by Satellite), the Ministry of the Environment (MMA) and the National Institute for Space Research (INPE) monitor deforestation in the Amazon forest by means of the application of land use classification methods based on satellite imagery (INPE, 2008). The PRODES methodology is based on visual interpretation of land use classes and it allows the identification of deforestation polygons with minimum area from 6.25 ha. Based on such classification, a photointerpreter delimits new deforestation polygons taking as a starting point the deforestation maps of previous years in which areas of non-forest and hydrography are properly identified. The deforestation rates calculated by PRODES are released annually. More recently, the DETER-B program (Diniz et al., 2015) started to provide the detection of deforestation and other changes in vegetation cover with a minimum area of about 1 hectare and a temporal resolution of 5 days, in which the updates in the vegetation cover in the Amazon area are generated almost daily. Other complementary systems have been developed to control and prevent deforestation, such as the Selective Logging Monitoring System (DETEX) (INPE, 2008), and the TerraClass Project (Coutinho et al., 2013). Recently, Map Biomas Alert was launched in Brazil. It consists of an online platform that incorporates a system for validating and refining alerts on deforested areas that covers all biomes in the Brazilian territory. The tool runs machine learning algorithms that are applied to refine the deforestation polygons with high-resolution satellite imagery from monthly deforestation alerts collected by the different systems in the country: DETER (Diniz et al., 2015), which includes the biomes of the Amazon and Cerrado, SAD/IMAZON (De Souza et al., 2008) and SIPAMSAR (CENSIPAM, 2008), both covering the Amazon biome, and Global Land Analysis and Discovery (GLAD) (Hansen et al., 2016), from the University of Maryland, comprising the other biomes. Such alerts are analyzed by means of visual inspection to eliminate false positives, and are then forwarded to the supervised classification process, from Google Earth Engine, with images from Planet satellites, which have a spatial resolution of 3 m, with daily frequency.

With the improvement of machine learning techniques applied to image processing, new potential methodologies started to be studied for the detection of forest cover changes. In particular, convolutional neural networks (CNNs) and, subsequently, fully convolutional neural networks have achieved promising results (de Bem et al., 2020, Grings et al., 2020, Khan et al., 2017, Wagner et al., 2020), having even surpassed human performance in image classification tasks (Yu et al., 2017). Amongst the different architectures of fully convolutional neural networks, the U-Net stands out for several factors, being the most important its segmentation and computational performances (Bragagnolo et al., 2021a). The U-Net architecture has already been successfully applied in studies involving the identification of deforestation (de Bem et al., 2020, Lee et al., 2020, Wagner et al., 2020, Bragagnolo et al., 2021c). de Bem et al. (2020) applied three CNN architectures (U-Net, SharpMask, and ResNet) using images from Landsat-8 to detect deforestation in the Amazon Forest. The authors found similar results for the three architectures evaluated, having outperformed classical machine learning algorithms, such as random forest and multilayer perceptron. The resulting metrics showed that the ResNet achieved slightly better performance, with 0.9252–0.9358 (precision), 0.9617–0.9508 (recall), while the U-Net achieved 0.9223–0.9175 (precision) and 0.9003–0.9508 (recall) in the validation sets of the same areas for two different time spans. In turn, Lee et al. (2020) mapped deforested areas using high-resolution remote sensing (Kompsat-3 images with 0.7 m resolution) and deep learning. They applied two CNN architectures (SegNet and U-Net) to find landscape affected by human-induced deforestation. The results showed that the U-Net architecture was approximately 11% more accurate than the SegNet architecture trained in the same database. Further, Wagner et al. (2020) mapped forest cover and changes and two species of pioneer trees, Cecropia hololeuca, and Tibouchina pulchra. The authors combined a dataset from WorldView-2 and WorldView-3 (spatial resolution of 0.5 m and 0.3 m, respectively) and applied a U-Net architecture. The resulting metrics in his study were precision (0.808–0.993), recall (0.801–0.995) and F1-score (0.804–0.994), which indicate once more the fitness of the architecture for the segmentation of satellite imagery. Abrams et al. (2019) developed an FCN called Habitat-Net based on the U-Net architecture in order to map images of tropical forest habitats using images of canopy and understory for training. The higher mean Dice coefficient of Habitat-Net (0.94 for canopy and 0.95 for understory) indicates that accuracy of Habitat-Net is higher than U-Net (0.89, 0.94). Finally, Bragagnolo et al. (2021c) evaluate six different FCN architectures (U-Net, SegNet, DeepLabV3+, PspNet, VGG-PspNet, and FCN32) to map forest and non-forest areas in the Amazon biome, using satellite imagery from Sentinel-2 (using the 10 m resolution bands). The U-Net achieved the best performance among the architectures tested, having scored 0.9470, 0.9356, 0.9676, and 0.9513 for accuracy, precision, recall, and F1-score, respectively. Further metrics related to computational performance also pointed to a superior performance of the U-Net architecture, thus indicating that, beyond its better segmentation performance, its training time was considerably shorter. Comprehensive and comparative evaluations have also shown that the U-Net architecture can be applied advantageously in the identification of landslide scars from satellite images (Bragagnolo et al., 2021a, Bragagnolo et al., 2021c), among other image segmentation problems that rely on satellite imagery.

Although methodologies based on fully convolutional neural networks and, in particular, the U-Net architecture, have been shown to provide accurate results in the task of identifying and segmenting deforestation areas (Bragagnolo et al., 2021c), there are limitations to be overcome. The occurrence of false positives is the most important one in this specific application, since they are responsible for most of performance degradation observed in practice. In a way, the occurrence of false positives and false negatives is expected and, from a certain point on, very difficult to be overcome by merely fine-tuning the network. The main routes available for seeking improved accuracy following current conditions and existing architectures of CNNs are (i) larger and more comprehensive image databases, with more thematic layers and with better resolution, (ii) variations in training parameters (loss functions, optimizers, learning rates), (iii) variations in architectures (number of layers, layer characteristics, number of neurons, attention mechanisms) and (iv) better equipment with greater processing power, which may allow the training of a greater number of candidate CNNs. Thus, although there is room for improving the performance of the method in relation to previous attempts (Bragagnolo et al., 2021c, Bragagnolo et al., 2021d, de Bem et al., 2020, Wagner et al., 2020), it is understood that, within the current state of the art, such improvements are of an incremental nature.

Considering these limitations, we offer a hybrid approach aimed at improving image segmentation accuracy over deforestation areas. We show that the identification of deforestation areas can benefit from the application of a logistic model combined with vegetation indexes, which have already been studied to assess forest growth (Richit et al., 2019, Richit et al., 2021). In general, logistic models describe the growth of a given population that is subject to limited resources (Verhulst, 1845, Verhulst, 1847), and allow the characterization of the evolution of the population by means of well-defined, intuitive, and physically understandable quantitative parameters. In the case of a forest environment, the model can be used to assess the rate of growth of vegetation density over time, also allowing to identify spots with negative growth in plant density or any other anomalous condition whatsoever. Thus, the application of the logistic growth model allows to improve the results obtained from semantic segmentation and to promote increased classification accuracy through the overlap of two different and complementary types of assessment of forest cover. This hybrid solution is aimed at contributing to the implementation of automatic systems for monitoring deforestation in vast territories and in ‘near real-time’, thus enabling the timely issuance of deforestation alerts. Such a tool can be of great help to authorities against illegal deforestation by potentially identifying and pinpointing ongoing activities, thus allowing more effective action by environmental authorities.

Having presented the state-of-the art on the application of machine learning techniques to the segmentation of forest and non-forest areas from satellite imagery and having pointed the limitation of existing methodologies in what regards real-world applications, we now formulate the main scientific question under study in this work: would a hybrid methodology based on a blend of image segmentation and mathematical modelling offer a simpler and yet accurate and robust framework to serve as backbone of highly automated deforestation monitoring systems? Our initial hypothesis is that it would, based on the fact that image segmentation and growth modelling are complementary in the sense that while the former accounts for learned resemblances among the spectral signatures, size and shape of deforestation polygons, the latter allows the assessment of parameters whose values may point to anomalies in forest growth, including the case of clear cut.

This study aims to present and discuss a highly automated methodology for monitoring deforestation based on the application of fully convolutional neural networks and the logistic growth model. While the U-Net architecture is considered due to advantageous segmentation accuracy and computational performance shown in previous studies (Bragagnolo et al., 2021c), it is the case that the methodology is flexible to admit any other architecture whatsoever. As illustration of the methodology and of its application, case studies are carried out in the tropical forests of the Amazon and Atlantic Forest biomes in Brazil under the actual application conditions. In order to evaluate the results, we compared the deforestation polygons output by the proposed methodology with those produced and made available by the MapBiomas Alert project (MapBiomas, 2019), which uses and refines data from PRODES, the official program supported and conducted by scientific institutions of the Brazilian government (INPE, 2008). The results indicate that the proposed methodology can contribute to increasing the effectiveness and accuracy of the monitoring systems currently employed. Since the proposed methodology is shown to enable the agile and accurate identification of forest cover losses and to provide efficient computing, it is argued that it can allow the issuance of alerts of suspected deforestation activity in standalone automatic monitoring systems, or else as a complementary tool to existing systems currently under operation.

The text is organized as follows: Section 2 presents the materials and methods; Section 3 presents the results obtained in the study and comparisons with the results of MapBiomas Alert (MapBiomas, 2019); Section 4 discusses the implications of the results; finally, Section 5 makes final considerations and summarizes the research contributions.

Section snippets

Materials and methods

The methodological procedure is presented in three main stages: (a) description of the U-Net architecture and its training, which include the mapping of the forest regions and the identification of clouds; (b) description and application of a logistic model to identify areas with decreasing vegetation density; and (c) presentation of the automated system for mapping deforestation areas.

U-Net training

The U-Net trained with images from the Amazon and Atlantic Forest biome achieved performance metrics values as presented in Table 2, which are consistent with those found a previous study (Bragagnolo et al., 2021d). Also, Fig. 4 presents the training history considering the values of the loss function and the metrics evaluated for the trained U-Net. It is observed that, for all the metrics evaluated (second column), all values are close to 1, indicating a good performance of the networks,

Deforestation mapping

Based on the combination of different techniques (U-Nets and logistic model) and satellite scenes from the Sentinel-2 mission, which have an interesting frequency and resolution for the mapping and monitoring of many environmental processes, this study developed a methodology that provides highly accurate mapping and monitoring of deforestation areas. The computational efficiency of the implementation allows that such mapping can be readily obtained upon the availability of new satellite

Final remarks

This article aimed to present and discuss a methodology for mapping deforestation using a fully convolutional neural network architecture called U-Net coupled with the logistic growth model. By means of application of free and openly available multispectral imagery from the Sentinel-2 satellite for regions of the Amazon and Atlantic Forest biomes, an automated processing chain for the generation of deforestation polygons was proposed, tested, demonstrated and discussed. Case studies

Declaration of Competing Interest

The authors declare no conflict of interest.

Acknowledgements

The authors thank CAPES for the support.

References (44)

  • M. Abadi et al.

    TensorFlow: Large-scale Machine Learning on Heterogeneous Systems

    (2015)
  • D. Arruda et al.

    Amazon fires threaten Brazil's agribusiness

    Science

    (2019)
  • J. Assunção et al.

    Getting greener by going black: the effect of blacklisting municipalities on amazon deforestation

    Environ. Dev. Econ.

    (2019)
  • P.P. de Bem et al.

    Change detection of deforestation in the Brazilian amazon using landsat data and convolutional neural networks

    Remote Sens.

    (2020)
  • G. Bradski et al.

    Opencv. Dr. Dobb's Journal of Software Tools 3

    (2000)
  • L. Bragagnolo et al.

    Clouds dataset for semantic segmentation

    (2020)
  • L. Bragagnolo et al.

    Amazon and Atlantic Forest image datasets for semantic segmentation

    (2021)
  • L. Bragagnolo et al.

    U-Net Convolutional Neural Network Applied to Forest Mapping in Two Different Brazilian Biomes

    (2021)
  • CENSIPAM

    Relatório Técnico Final Missão SIPAM/SAR-MMA 2008

    (2008)
  • F. Chollet

    Keras

    (2015)
  • A.C. Coutinho et al.

    Uso e cobertura da terra nas áreas desflorestadas da Amazônia Legal TerraClass

    (2013)
  • C.,M. De Souza et al.

    Near real-time deforestation detection for enforcement of forest reserves in Mato Grosso

    (2008)
  • Cited by (0)

    View full text