Identifying Land Patterns from Satellite Imagery in Amazon Rainforest using Deep Learning
IIdentifying Land Patterns from Satellite Imagery inAmazon Rainforest using Deep Learning
Somnath RakshitJalpaiguri Govt. Engg. College [email protected]
Soumyadeep DebnathJalpaiguri Govt. Engg. College [email protected]
Dhiman MondalJalpaiguri Govt. Engg. College [email protected]
Abstract —The Amazon rainforests have been sufferingwidespread damage, both via natural and artificial means. Everyminute, it is estimated that the world loses forest cover the size of48 football fields. Deforestation in the Amazon rainforest has ledto drastically reduced biodiversity, loss of habitat, climate changeand other biological losses. In this respect, it has become essentialto track how the nature of these forests change over time. Imageclassification using deep learning can help speed up this processby removing the manual task of classifying each image. Here,it is shown how convolutional neural networks can be used totrack changes in land patterns in the Amazon rainforests. Inthis work, a testing accuracy of 96.71% was obtained. This canhelp governments and other agencies to track changes in landpatterns more effectively and accurately.
Index Terms —Amazon rainforest; Land pattern; Computervision; Deep Learning; Satellite Imagery; VGG.
I. I
NTRODUCTION
The Amazon rainforest is a moist broadleaf forest in SouthAmerica that covers a land area of more than 7,000,000km . There are nine countries that fall under the Amazonrainforests including countries like Brazil, Peru, Columbia,Venezuela, Ecuador, Bolivia etc. as mentioned in Fig. 1 . Morethan half of the worlds total rainforest cover is representedby the Amazon. Worlds largest and most biologicallydiverse tract of tropical rainforest is thought to be presenthere with almost 400 billion individual trees of more than15,000 species. The Amazon rainforest is thought to bein existence for at least 55 million years and is richer inwildlife biodiversity too than the African or Asian wet forests.Fig. 1: Map of the Amazon BasinThe Brazilian part of the Amazon was largely intact till theconstruction of the Trans-Amazonian Highway began in the 1970s [1]. It triggered a high level of deforestation and sincethen, this rate has fluctuated but has always remained highconsistently as shown in
Fig. 2 . This makes it important todevelop an automated approach of monitoring the land patternin the Amazon rainforests that will help the governmentalagencies to easily and effectively monitor the deforestationand other activities taking place in the Amazon.Fig. 2: Annual Forest Loss in the AmazonComputer vision is an important field these days and deeplearning using neural networks have been at the forefront ofit. Using a wide range of deep learning techniques, it hasbeen possible to achieve higher accuracy than the previousmodels by a noticeable amount. Also, deep learning triumphsconventional methods of computer vision by not requiringhand-crafted features. Thus, this is a technique that can beapplied to a wide variety of problems. In the proposed work, anautomated way to label satellite images with their correspond-ing class of land cover has been developed. The VGG16 modelhas been used after fine tuning and achieved high accuracy.This work may be extended by using other well-known imageclassification models. Deforestation in the Amazon rainforestshas increased considerably in the past few decades. This hasaffected its biodiversity and climate adversely. Using satelliteimagery, it has been possible to track the changes taking placewithin a region. However, due to the presence of huge amountof data, it requires considerable manual resources for properlabelling. https://rainforests.mongabay.com/amazon/amazon destruction.html a r X i v : . [ c s . C V ] S e p I. RELATED WORKSNeural networks have been used to solve a wide range ofproblems. It has been quite successfully used in problemssuch as the prediction of diabetes [2], face detection [3],object localization [4], etc. Visual image recognition hasgained spectacular popularity in the current scenario owingto the tremendously fast-paced research on this topic. Deepconvolutional networks started gaining popularity from theyear 2012 when Krizhevsky et al.’s [5] AlexNet won theImageNet challenge and beat other models by a big margin.Since then, all winners of ImageNet challenge have used adeep convolutional neural network architecture. Simonyan andZisserman [6] developed the VGG model that improved theerror rate in the ImageNet challenge.However, deep learning was not mathematically understoodwell by the community. In this regard, the works of Zeiler at al.[7] helped in throwing light in how deep learning models work.Here, it was shown that how deep learning models learnedfeatures with respect to each layer.Rajat et al. [8] introduced the concept of transfer learning inthe year 2007. Using this, it is possible to use the weights ofthe model that was trained on a particular dataset and retrainthe last layer before passing its output through a classifier.This allows to get the results that are nearly as good as thesuccessful models without requiring the enormous computingresources that were used in training the original model.Gardner et al. [9] used the ResNet50 model to classify landpatterns in the Amazon rainforests and obtained a F-Score of0.91. They used a number of data augmentation and ensembletechniques for this purpose.Longwell et al. [10] used the near-IR channel of the imageand through a deep residual architecture, obtained a F2 scoreof 0.9.To speed up the development using neural networks, quite alot software libraries have been developed. Some examplesinclude Scikit-Learn [11], Tensorflow [12], Keras [13] formachine intelligence and Pandas [14], [15] for data analysis.These libraries are often used not only in deep learning butalso in many other tasks.III. DATASETThe dataset for this work has been derived from Planetsfull-frame analytic scene products using its 4-band satellites insun-synchronous orbit (SSO) and International Space Station(ISS) orbit.
A. Chip (Image) Data Format
The set of chips were captured with four bands of dataeach, viz. red, green, blue and near-infrared. The GeoTIFFdata was originally captured along with the chips. However,they have been removed with ground control points (GCPs)for the purpose of this experiment as the data was not foundto be essential.The images in this dataset have a ground-sample distance(GSD) of 3.7 m and orthorectified pixel size of 3 m. Planets Flock satellites have been used to collect this data in bothsun-synchronous and ISS orbits between January 1, 2016 andFebruary 1, 2017. In the dataset, the TIFF files were convertedto JPG files for easier processing using the Planet visualproduct processor . B. Data Labeling Process and Quality
The Crowd Flower platform was used to label this datasetusing crowdsourced labour. Planet acknowledges the fact thatalthough the utmost care was taken to correctly label thedataset, not all labels are accurate in nature.While labelling, the whole dataset was divided into twosets, a ”hard” and an ”easy” set. Scenes having easier-to-identify labels like primary rainforest, agriculture, habitation,roads, water, and cloud conditions were placed in the easycategory. Shifting cultivation, slash and burn agriculture, blowdown, mining, and other phenomena were placed in the hardcategory. C. Class Labels
Planet’s Impact team was consulted while labelling thedataset. This dataset reasonably represents the places of inter-est in the Amazon basin. The labels can broadly be brokeninto three groups. They are: (a) (b) (c)(d) (e) (f)(g) (h) (i) Fig. 3: Sample Chips and Their Labels (a) primary (b) roads+ primary (c) partly cloudy + primary (d) haze + primary(e) cultivation + primary (f) water + primary (g) habitation +partly cloudy (h) agriculture + roads + primary (i) agriculture+ pasture + primary + partly cloudy ) Atmospheric Conditions2) Common Land Cover (Land Use Phenomena)3) Rare Land Cover (Land Use Phenomena)The whole dataset has 17 type of labels which had to beidentified for each of the chips showed in Figure 3. They aredescribed below: • agriculture • artisinal mine • bare ground • blooming • blow down • clear • cloudy • cultivation • habitation • haze • partly cloudy • primary • road • selective logging • conventional mine • slash burn • waterAs a single image can have multiple classes in this datasetso, in the algorithm, all such classes were tried to predictcorrectly for each of the images.IV. DATA ANALYSISSome basic data analysis was performed on the datasetwhich have been described in details below. A. Distribution of Training Labels
Firstly, the histogram as present in Figure 4 showing thedistribution of training labels was constructed. It has beenfound that the dataset is not balanced in nature, i.e. , alllabels are not present in uniform quantity. Labels such asprimary, clear and agriculture are present in significantly morenumber than the other ones. Whereas, some other labels likeslash burn, blow down and conventional mine are present invery less quantity. Note that in the dataset, a single image mayhave multiple classes. The histogram must be seen keeping thisin mind. Fig. 4: Distribution of Training Labels
B. Correlation Matrix
The correlation matrix was plotted, as shown in Figure 5,to understand the occurrence of the classes with respect toeach other. Here, redder is the label, more is the value of thecorrelation for any given pair of classes. After studying this plot, some interesting results were observed. Some of themare: • The label primary is associated with almost all classes.This means that most chips have some degree of primaryforests along with other labels. • The label agriculture is also associated with a few labelslike road, habitation and cultivation.Fig. 5: Distribution of Training LabelsV. PREPROCESSING OF DATASETEven after converting to JPG, the dataset was quite large insize. It would have been computationally expensive to train themodel on such a large dataset. Besides, the obtained datasetcontained images of various dimensions. Hence, all imageswere resized to a standard size, in this case, 128x128 pixels.This is also an important step as it helps in speeding upthe training. Since the downloaded VGG16 model did notcontain the top layer, it was possible to train with imageswith dimensions (128x128x3) that were different from thedimensions of images used in the original VGG16 model(224x224x3). In this dataset, 40479 images for training and40669 images for testing were used. Each image may beclassified into multiple classes.VI. METHODOLOGYIn the proposed work, the VGG16 model has been usedto classify images into various classes. Figure 6 shows theoriginal diagram of the VGG model.Fig. 6: Distribution of Training LabelsIn the model, a batch normalization layer was added to theinput layer and then fed to the VGG16 model. The last blockf the original VGG16 model was removed and the outputof the penultimate block of the VGG16 model was flattened.It was then passed on to a softmax classifier to present theoutput with respect to 17 classes. Here, 20% of the trainingdata was used for validation after training. The architectureof this model is present in Table I.TABLE I: Architecture of VGG16 model
Layer (type) Output Shape Parameterinput 1 (InputLayer) (None, 128, 128, 3) 0batch normalization 1 (None, 128, 128, 3) 12vgg16 (Model) (None, 4, 4, 512) 14714688flatten 1 (Flatten) (None, 8192) 0dense 1 (Dense) (None, 17) 139281
Here, the Adam optimizer [16] has been used to minimizethe loss, which is measured by binary cross- entropy, with alearning rate of 104. Batch size of 128 was used here and thismodel was trained for 15 epochs. By this time, the trainingloss had converged. Using an NVIDIA Tesla K80 GPU, thistook around one hour to train. The plot between the trainingloss vs epoch is shown in Figure 7.Fig. 7: Plot of Training Loss vs EpochVII. RESULTThe following metrics were evaluated in our workPrecision = T PT P + F P
Recall = T PT P + F N
Accuracy = T P + T NT P + T N + F P + F N
With TP , FP , TN , FN being number of true positives, falsepositives, true negatives and false negatives, respectively.F-Beta Score = F β = 1 β +1 1 precision + ββ +1 1 recall = (1 + β ) precision.recallprecision + recall Categorical Cross Entropy = n (cid:88) i K (cid:88) k − y ( k ) true log ( y ( k ) predict ) In the experiment, a training loss of 6.88%, training accu-racy of 97.35% and testing accuracy of 96.71% were obtained.Also, an F-beta score of 92.69% was obtained. The F-betascore is a weighted harmonic mean of the precision and recall.An F-beta score reaches its best value at 1 and worst score at0. VIII. CONCLUSIONIn this work, a way to classify satellite imagery in anautomated manner using deep learning with the help of theVGG16 model has been shown. High accuracy was consumingone hour while training with an NVIDIA Tesla K80 GPU.This model can be successfully applied to track the changingland pattern in the rainforests of Amazon. This data about thelocation of deforestation and human encroachment on forestscan help governments and local stakeholders respond morequickly and effectively. Besides, this model can be used totrack natural calamities like floods, forest fires, etc.IX. FUTURE SCOPEA few additions may be made to this work for improvementsmentioned below:Using a larger neural network is likely to give a betterresult. Models like ResNet and Inception, which are deeperin nature may give better results than the VGG16 model.Also, increased preprocessing of the dataset may helpin better classification. In this work, it has been shown howresizing the provided image to 128x128 pixels can be madeto obtain good performance. No preprocessing involving thetexture and nature of the image itself was performed.Performing data augmentation to make the system morerobust may be another way of getting better results. Sincethe satellite images may vary in terms of lighting effect,rotation, shifting, etc., it may be a good idea to perform dataaugmentation to enlarge the dataset for better training.These things may be investigated in the upcoming future toimprove the accuracy and robustness of this model.R
EFERENCES[1] Nunes Kehl, Thiago, Viviane Todt, Mauricio Roberto Veronez, andSilvio Csar Cazella, “Amazon rainforest deforestation daily detectiontool using artificial neural networks and satellite images,”
Sustainability4 , vol. 10, pp. 2566–2573, 2012.[2] Somnath Rakshit, Suvojit Manna, Sanket Biswas, RiyankaKundu, PritiGupta, Sayantan Maitra, and Subhas Barman, “Prediction of DiabetesType-II Using a Two-Class Neural Network,”
International Conferenceon Computational Intelligence, Communications, and Business Analyt-ics,Springer,Singapore , pp. 65–71, 2017.[3] Rowley, Henry A., Shumeet Baluja, and Takeo Kanade, “Neuralnetwork-based face detection,”
IEEE Transactions on pattern analysisand machine intelligence , vol. 20, no. 1, pp. 23–28, 1998.4] Pierre Sermanet, David Eigen, Xiang Zhang, Michal Mathieu, RobFergus, and Yann LeCun, “Overfeat: Integrated recognition, local-ization and detection using convolutional networks.” arXiv preprintarXiv:1312.6229 , 2013.[5] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton., “Imagenetclassification with deep convolutional neural networks,”
Advances inneural information processing systems , pp. 1097–1105, 2012.[6] Simonyan, Karen, and Andrew Zisserman, “Very deep convolu-tional networks for large-scale image recognition,” arXiv preprintarXiv:1409.1556 , 2014.[7] Zeiler, Matthew D., and Rob Fergus, “Visualizing and understand-ing convolutional networks,”
European conference on computer vi-sion,Springer, Cham , pp. 818–833, 2014.[8] Raina, Rajat, Alexis Battle, Honglak Lee, Benjamin Packer, and AndrewY. Ng. , “Self-taught learning: transfer learning from unlabeled data.”
Proceedings of the 24th international conference on Machine learning”,volume = .[9] Gardner, Daniel, and David Nichols, “Multi-label Classification ofSatellite Images with Deep Learning.”[10] Longwell, Scott, Tyler Shimko and Alex Williams, “DeepRootz: Clas-sifying satellite images of the Amazon rainforest,” .[11] Pedregosa, Fabian, Gal Varoquaux, Alexandre Gramfort, VincentMichel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel et al., “Scikit-learn: Machine learning in Python.”
Journal of machine learning re-search , vol. 12, pp. 2825–2830, 2011.[12] Abadi, Martn, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis,Jeffrey Dean, Matthieu Devin et al. , “TensorFlow: A System for Large-Scale Machine Learning.”
In OSDI , vol. 16, pp. . 265–283, 2016.[13] Chollet, Franois, “Keras,” , 2015.[14] McKinney, Wes. , “Data structures for statistical computing in python.”
Proceedings of the 9th Python in Science Conference , vol. 445, pp. 51–56, 2010.[15] McKinney, Wes, “pandas: a foundational Python library for data analysisand statistics,”
Python for High Performance and Scientific Computing , pp. 1–9, 2011.[16] Kingma, Diederik P., and Jimmy Ba, “Adam: A method for stochasticoptimization,” arXiv preprint arXiv:1412.6980arXiv preprint arXiv:1412.6980