SIT.net: SAR Deforestation Classification of Amazon Forest for Land Use Land Cover Application ()
1. Introduction
Forest monitoring is important part of healthy climate and useful for sustainable development. The world’s largest ecological specifically carbon sinker and collector is the forest. Thus, precisely keeping track of variation in forest cover is reaching the entire world’s climate neutrality plan. With the help of merging different band, use mathematical modelling and apply color thresholding with segmentation to find out more accurate deforest mitigation. Due to high mitigation of deforestation compared to reforestation affecting the climate changes [1] as well as human life including global development of society, greenhouse effect and many more problems arise.
Comprising the majority of terrestrial species worldwide, tropical forests are among the planet’s most biodiverse ecosystems. Unpleasantly, the last several decades have seen an alarming acceleration of the use of both renewable and non-renewable natural resources in conjunction with industrial development. This has resulted in significant environmental changes, such as the loss and degradation of forests [2] . In the Brazilian Legal Amazon (BLA), the ongoing deforestation process is a particularly dire scenario. Approximately 5,200,000 km, or more than 59% of Brazil’s total land area is comprised of this region [3] .
In environmental research, remote sensing (RS) has proven highly valuable [4] , most notably when it comes to tracking forest ecosystems. Due to cloud coverage and heavy rain, it is more difficult to monitor deforestation accurately. That’s why, compared to optical satellite, SAR satellite is good for penetrating clouds and gives better resolution during heavy rainfall [5] . Using remote sensing imagery to track changes in the world’s forests is a very successful approach, mostly because it is inexpensive and has a short duration of interval to revisit the same area for better observation [6] .
SAR satellite is an active sensor which detects and measures their own illuminated energy for investigation purpose, good in both day and night time [7] . SAR satellite time series is good for observation due to high spatial resolution (20 m) and temporal resolution (6 days) for land monitoring [8] . Sentinel-1 has four types of data for different purposes like raw data level 0 for auxiliary information, SLC-Single Look Complex data level 1 for soil and vegetation monitoring, GRD-Ground Range Detection level-1 for land monitoring, Level-2 OCN for Floor and Ocean data monitoring [9] . Due to different acquisition mode, their polarization and resolution are difficult including range of coverage area. The surface characteristics and deforestation monitoring are revealed by the backscattering coefficient that SAR sensors record and the radar’s incidence angle determine how the surface reacts to these elements [10] .
Deforestation brought on by woodcutting, ranching, mining, oil extraction, building dams, and infrastructure projects together account for forest loss. Degradation of forests and climate related climate change shifts, wildfires, framing industrial plantations, and tree illnesses [11] .
The deforestation analysis using Sentinel-1 requires pre-processing to remove speckle noise and need stabilization used for balancing the radar signals variation in different time series images [12] . Stabilization is useful if you are using Maximum Likelihood detection approach and U-net CNN approach but not useful with Adaptive linear threshold approach [13] .
The ongoing research on deforestation, RDSTC-Reversed Depth wise Separable Transposed Convolution faces problem of segmenting and classifying mixed pixel and unable to identify early stage of fire detection [14] .
Using deep learning model with U-net adding (residual block attenuation gate unit) RAU-Net and use dataset with merging different band of Sentinel-1,2, MODIS, to increase the accuracy for detecting fire [15] .
Forest change detection is also important part of deforestation monitoring for LULC. The current scenario of research based on cloudless satellite images using deep learning CNN with attention mechanism for better analysis compared to standard U-net [16] .
Forest biodiversity is an important for monitoring vegetation and helps towards sustainable development of world. Using Squeeze Excitation with Partial Convolution including transpose for forest mapping and structuring [17] .
In SAR, backscattering defines the intensity of ground object which helps us for better understanding. Using the patch based CNN3D model gives better result compared to CNN2D as well as CNN1D with dual polarization SAR. The arising, if patch size is large creates more smoothing and neglect small objects which reduce accuracy and unnecessary computation is required [18] .
Deforestation Expansion pattern observed by frontier which reflect important information about carbon, water, biodiversity for better understanding of regional forest loss. Using fine scale spatial temporal optimization techniques with Bayesian Deep learning methodology (DnCNN + BLUC) helps to improve accuracy of SAR satellite and reduce speckle noise problem but due to variation of wet season and dry season patch size and patch framing is different which reflect different variation [19] .
Basically, Land Use Land Cover means land used by human which is directly related to human nature interrelationship for sustainable development. The remote sensing images have different spectral land which is used for different applications. For combining different spectral band on the basis of remote sensing indices like NDVI, NDBI, NDWI and apply Random Forest to monitor changes increase the accuracy and efficiency. In future Random Forest combine with Deep Learning to deal with mixed pixel problem and identifying complex boundaries as well [20] .
As per the latest literature review, author is summarizing this paper with new ideas based on Transpose Deep neural network designing including the segmentation concept and mathematical pre-processing. The structure of paper is based on Motivation and associated contribution. The main motivation is SAR satellite that gives high resolution in both day and night and penetrates clouds even not effecting during heavy rain or other climate effects.
Use Multi Temporal Speckle Filtering and compare their results to find out the best filter to reduce speckle noise which is the big disadvantage of SAR satellite. Apply Band Math Merging Concept to create new bands for analysis and implement new methodology called SIT.net to classify time series satellite images for monitoring deforestation. For calculating area of deforestation, use Color Thresholding and compare their results for better analysis. Apply segmentation and compare their results to identity the best segmentation for creating mask to analyze deforestation. Use Multi Temporal Speckle filters like Boxcar, Lee Sigma, Frost, Median and Intensity Driven Adaptive Filter (IDAN) and compare their results.
· Organization of paper
➢ Dataset Study
➢ Methodology Modelling
➢ Outcomes Exploration
2. Materials and Method
2.1. Dataset Study
In Brazil, the most responsible area of tree loss since 2010 to 2022 is Para. In Para, approximately the third most responsible region for tree loss is Novo Progresso since 2001 to 2022. In October 2023, Novo Progresso got many alert regarding deforestation and mainly due to fire [21] . Past 15 years, REDD Reducing Emission of Deforestation and Forest Degradation monitoring the Brazil and analyze that Para is the second largest area of Brazilian Amazon of deforestation [22] [23] . As per the above analysis, authors choose State of Para, Brazil dataset to analyze the proposed model for deforestation classification. Below “Figure 1”and “Table 1” visualize the study area using google earth pro.
The dataset used in this paper captured from NASA. For the analysis of proposed model, author use Sentinel-1 GRD images of year 2019 and year 2023. The Sentinel-1 has feature of Dual Pole Ground Range Detector to observe the Earth accurately in both day and night without any interruption of heavy rains or clouds. “Table 2” summarizes more about dataset used in this paper.
![]()
Figure 1. Geographical representation of study area.
![]()
Table 1. Study area geographical coordinates.
![]()
Table 2. Image acquisition description.
2.2. Methodology Modelling
The modelling of complete methodology is based on five steps. With the help of workflow diagram the whole proposed concept visualize easily. Below “Figure 2” explains the number of steps used in this paper. The proposed model divide into three sections, but pre-processing and intermediate processing done by SNAP tool to prepare data for training and testing for our proposed model operated in MATLAB 2023b, to check the accuracy and loss for validation.
The Sentinel-1 data is a type of SAR (Synthetic Aperture Radar). SAR data need to pre-processed, author use SNAP-Sentinel Application Platform software developed by ESA (European Space Agency) Copernicus. This tool is suitable for earth observation analysis. The pre-processing steps used for SAR are described in “Table 3”.
After pre-processing, need to apply filtering to reduce speckle noise. Speckle noise is the only challenge in SAR satellite images. Speckle noise is multiplicative in nature and due to this coherent of imagery is also affected. There are many Bayesian and Non-Bayesian methods are available to reduce speckle noise. In this paper, authors compare Bayesian filters to reduce noise [24] [25] [26] [27] . “Table 4” describes multi temporal speckle filters.
The next methodology is called BMM, Band Math Merging concept. Mathematical modelling based on arithmetic concept like addition, subtraction, multiplication, division. On the basis of filters analysis, apply band math in each filters for better results. With the help of multiple band math merging concept, get more understanding of which band math is perfect for deforestation analysis.
In this paper, the basic mathematical model is used called addition, subtraction, and multiplication. The dataset of 2019 and 2023 Sentinel-1 dual polarization (VV, VH) is used for Band Math merging to create new band called Add, Sub, Mul for analysis. The expression used here are as follows:
. (1)
. (2)
. (3)
The next step based on deep neural networking concept. After applying Band Math in different filtered images and use that images for classification using proposed model called SIT.net. The proposed model is the combination of Transposed Convolutional neural network using Transformation methodology to analyze the accuracy of filtered images and Band Math filtered images. Below “Table 5” gives complete overview of proposed model with description.
![]()
Figure 2. Workflow of methodology modeling.
![]()
Table 3. Sentinel-1 Pre-processing.
![]()
Table 4. Multi temporal speckle filters analysis.
![]()
Table 5. Proposed model description.
“Figure 2” workflow diagram helps us to understand the complete overview between pre-processing, intermediate processing and final processing. The proposed model analyzed by their output results and images. In next section, we discussed about the outputs of pre-processed results of 2019 and 2023 dataset. And after applying different Multi temporal speckle filtering on pre-processed images for better visualization. The filtered images used for Band math merging concept to create new band for applying proposed model on it and compare the results between filtered images and band math filtered images using accuracy and loss.
For calculating the deforested area, we use Color Thresholding and Segmentation for identifying the deforested pixels using binary images. The gray scale value of each RGB channel is used to divide the foreground and background pixels in RGB color thresholding algorithm which means the intensity of the image pixels is not independent of color chromatics gives unreliable thresholding. That’s why L*a*b or HSV is used, where the intensity parameters and chromatic characteristics are separated [28] . To convert the color thresholding images into binary and create mask for identifying the deforested area. To calculate the deforested area, binary mask is used. For the area calculation, authors convert number of pixels into kilometer per square.
. (4)
2.3. Outcomes Exploration
The experimental analysis is based on four sections. The First section, discussed about pre-processed images as an input for further analysis based on “Table 3”. The second section, discussed about the intermediate processing results based on “Table 4” and based on “Equation (1)”, “Equation (2)”, and “Equation (3)” Band Math merging of filtered images. And the third section, discussed about the proposed model results based on “Table 4” and the comparison between two different dataset. The final Forth section, based on “Equation (4)”discussed about deforested area calculation and their difference analysis.
The pre-processing of original image computation gives an overview that, due to deforestation in 2023 images takes less computation time compared to 2019 original image computation. Deforested pixel area and non-deforested pixel area have different pixel vales area computed easily and fast. With the help of “Table 6”, it reflects the differences.
For better understanding of dataset, histogram and adjusted histogram of original images see in below “Figures 3(a)-(f)”.
The next step is applying speckle filtering. The multi-temporal speckle filtering helps to understand which fitter gives better results and take less computational time see in below “Table 7” and compare it with “Table 4”.
After applying the multi-temporal speckle filtering, band math concept based on “Equation (1)”, “Equation (2)”, and “Equation (3)”applied see in below “Table 8”. The band math gives more visual images for analysing the deforestation and their computational time is seen in “Table 4”. The new bands help to train proposed model to analyze and check the accuracy and loss between two different data.
![]()
Table 7. Multi-temporal speckle filtered images.
![]()
Table 8. Band math filtered image analysis.
![]()
Figure 3. (a) Histogram of 2019; (b) Histogram of 2023; (c) Adjusted histogram of 2019; (d) Adjusted histogram of 2023; (e) Before and after image of 2019; (f) Before and after image of 2023.
The final processing based on three steps, first is applying proposed model in filtered images datast and band math images dataset, compare the accuracy and loss see in “Table 9”.
For justifying the proposed model functionality see in below “Figures 4(a)-(d)” accuracy and loss as per number of iteration of filtered images dataset and in “Figures 4(e)-(h)” is the accuracy and loss as per number of iteration of Band math images dataset. The dotted line explains the accuracy and loss and blue line and orange line explain the processing.
Once we analyze the accuracy and loss of both filtered images and band math images. The next step is to calculate deforested area. For identifying the deforested and non-deforested area, need segmentation and color thresholding. Using Color Thresholding, L*a*r gives better results compared to HSV and RGB to create mask see in below “Figure 5(a)” “Figure 5(b)”.
We use Binary Masked and apply on Original Images, compare with segmentation using Thresholding (100) and apply Morphology using circle shape pixel of radius 1 with active contouring using region based till 100 iterations for better analysis between deforested and non-deforested area. To calculate deforested area per pixel using props function which helps to give lot of information about segmented region in the images. We use area per pixel information and add them to calculate area see in below “Table 10”.
![]()
Figure 5. (a) Binary masked image 2019; (b) Binary masked image 2023.
![]()
Table 10. Deforested and non-deforested area analysis.
![]()
Figure 6. (a) Deforested and non-deforested; (b) Deforestation comparison; (c) 2019 deforestation; (d) 2023 deforestation.
For better analysis, bar graph of deforested and non-deforested area is seen in below “Figures 6(a)-(d)”. In “Figure 6(a)” and “Figure 6(b)” helps to understand actual differences between non-deforested and deforested area using original images and segmented images of 2019 and 2023.
For better visualization, combine the segmented area with orange in color and non-deforested area with blue in color see in below “Figure 6(c)” “Figure 6(d)”.
3. Conclusion
The forest monitoring is important for healthy environment. The proposed methodology helps to find out the deforested area using SAR satellite which helps to monitor in both day and night without any interruption of weather. For computational analysis of SAR satellite, imagery is the important key point of this paper due to large area covered. The Band Math concepts to create new band for dataset preparation take less time compared to multi-temporal speckle filtering and give better understanding of the images, also increase the accuracy and reduce the loss compared to filtered images. The proposed model takes less computational time for Band math image processing compared to filtered image processing. For calculating area, RGB and HSV do not give better results compared to L*a*b in color thresholding to create mask for analyzing deforestation and using segmentation with morphology and active contouring helps to eliminate unwanted mask to make more clear visualization of deforested area. In future, more images of different time series dataset are used for analyzing deforestation and applying proposed model in optical satellite to analyze deforestation.
Acknowledgements
We thank Dr. Manoj Kumar and Dr. Ankush Agarwal for their support and guidance. A special thank to GLA University and my family for supporting me and giving me an opportunity to complete one more step in my PhD.