DEEP LEARNING BASED MULTI TASK ROAD EXTRACTOR MODEL FOR PARCEL EXTRACTION AND CROP CLASSIFICATION USING KNOWLEDGE BASED NDVI TIME SERIES DATA

: The role of agriculture in food security is critical by the rapid increase of the global population. Food production should be strongly increased to secure and provide life necessities. Remote sensing provides a fruitful tool in agriculture management, and it provides advantages in terms of saving time and effort. This study aims to examine the feasibility of deep learning edge detection algorithm in order to automatically extract the agricultural parcel boundaries from the open-source Google Earth data. The potential of using Normalized Difference Vegetation Index (NDVI) in crop classification crop was also presented after analyzing the pattern of Rabi major crops (Wheat and Sugarcane) in Haridwar district, Uttarakhand, India. The advantage of Google Earth Engine cloud-based platform was exploited in generating NDVI data from Sentinel-2 satellite between October 2022 and February 2023 in order to save time and effort. To check the accuracy of the deep learning model, the value of the Mean Intersection of Union (mIoU) was tested and reached 0.79.To examine the results, ground truth data were collected in the study area using Unmanned Aerial Vehicles. The overall accuracy of the rule set-based classification reached 91.17%, and the kappa coefficient value was 0.82.


INTRODUCTION
The 2.3 billion people increase in global population is playing a significant role in the rapid increase of global food necessities from agricultural crops (Tilman et al., 2011).The anticipation that the global population by 2050 is around 9.8 billion; thus, measures should be taken such as increasing food production (Tilman et al., 2011).The reports indicate that around 1 billion people suffer from famine every day as a result of the lack of food supply, and this number is to be increased by 2050 to 2 billion.According to this storyline, agriculture production should be enforced to be increased by 70% in African and Asian developing countries in the future (FAO, 2009).Since the beginning of the 1970's, remotely sensed data provided by satellites have been adopted to be used in agriculture ( Tempfli et al., 2013;Mulla, 2013).At the large-scale regions, remote sensing techniques with effective spatial and temporal information play a significant role as an important factor in the management of agricultural fields and crop monitoring (Gao et al., 2017).In order to automatically extract the cadastral information from remotely sensed data, deep learning methodology was adopted in many applications, such as road extraction, object extraction, parcel delineation, etc.In terms of agriculture, deep learning techniques is used recently to extract the agricultural parcel boundaries as the manual methods represented by digitization is time and effort-consuming.Notable number of studies used the conventional supervised classification to extract the crop information using only one single image.Because of crop variability and pixel heterogeneity, the application of traditional pixel-based classification techniques is limited as a result of the "salt and pepper" effect.The spectral similarity between various crops can be avoided by analyzing time series data provided by space platforms (Li et al., 2015a).The basic idea of remote sensing of vegetation is acquiring information on the interaction between electromagnetic energy and vegetation on the surface of the earth via passive sensors.Many factors play an important role in the difference in the reflectance of electromagnetic spectra, such as plant class, water content, etc. (Mulla, 2013;Liu, Sun and Liu, 2016).For the process of many quantitative and qualitative evaluations, such as growth dynamics, vegetation cover, and vigor, the derived vegetation indices from remote sensing-based canopies are straightforward and widely used.The remote sensed spectral characteristics and information from the canopy and plants are obtained by the differences in the green leaves (Xue and Su, 2017).Normalized Difference Vegetation Index (NDVI) is one of the most widely used in multi-spectral remote sensing, and it is derived by dividing the difference between red and nearinfrared bands on the sum of them (Karnieli et al., 2010;Yin et al., 2012;Jain and Pandey, 2021;Pandey and Jain, 2022).NDVI is used in many applications, such as health condition assessment and deforestation analysis (Singh and Kushwaha, 2021;Yogender et al., 2022).As long as the NDVI value is positive, the class is crops, grass, forests, and other vegetation cover, in contrast, the negative values indicate non-vegetation cover such as urban area, sand rocks, etc. (Pettorelli et al., 2005).The major problem in remotely sensed image classification is the crop variability represented by the pixel-level spectral heterogeneity which is the source of the pixel-based classification techniques limitation as the "salt and pepper" phenomena is common in the crop pixel-based classification (Vieira et al., 2012;Hu et al., 2013;Li et al., 2015b).The fundamental requirement for conventional supervised image classification approaches is collecting training data (Tempfli et al., 2013).According to (Huang, Davis and Townshend, 2010) the imbalance of the training samples could affect the quality of the classification; in addition, the effect of training samples might be more than the algorithm used itself.In this study, the feasibility of deep learning methods was examined in order to automatically extract the cadastral agricultural field using the free availability of Google Earth Data, then the capability of time series NDVI extracted from Sentinel-2 satellite with the synergistic of the Google Earth Engine cloud-based platform was tested to extract the major Rabi crops (Sugarcane and Wheat) in Haridwar district, India.

Study Area
Haridwar district is a part of Uttarakhand state in northern India.

Google Earth Data
Google Earth is a popular geospatial tool developed by Google that allows users to explore the Earth's surface using satellite imagery, aerial photography, and 3D models.Since its launch in 2005, Google Earth has been widely used in various fields, such as geography, environmental science, urban planning, and archaeology.
In this investigation, a patch of the latest image of the study area was collected from Google Earth on the date of 24 th November 2022.The procedure of exporting the image was done using the Save Image tool available in Google Earth software with the maximum resolution (1529*925), after exporting the image from Google Earth, it was imported to ArcGIS in order to relate the internal coordinate system of a that digital map to a ground system geographic coordinates.Georeferencing was conducted in GIS environment.The image shows the boundary between the agriculture parcels in Rabi season when the season of harvesting period is already done, and the boundaries between agriculture parcels are easily interpreted.The Google Earth image of the study area is shown in Figure 2.

Sentinel-2 Data
The advantage of 10 m medium spatial resolution of Sentinel-2 multispectral satellite was exploited in this study in order to classify the crop types in the study area.
The feasibility of Google Earth Engine platform were used in order to collect the NDVI data during the study period.Google Earth Engine is a geospatial processing online platform saves time and effort.

METHODOLOGY WORKFLOW
Google Earth image was downloaded as an input to generate the boundaries between the agriculture parcels.The next step of the investigation is to apply the knowledge based in order to perform the crop type identification based on time series NDVI data extracted from Sentinel 2. Figure 3 illustrates the methodology workflow.

Data Acquisition
In this investigation, Sentinel-2 Level 2A cloud-free data were collected between October 2022 and February 2023, with the exception of data from January 2023, as the cloud cover was dressing the entire data from this month.The advantages of Google earth Engine platform were exploited in order to save time and effort in data preparation as the conventional methods using available software is time-consuming; as a result, a code was processed on Google Earth Engine in order to obtain the time series NDVI data directly for each month from October 2022 to February 2023.

Parcel Delineation
After downloading the patch from Google Earth, clipping procedures were implemented in the experimental area.In order to obtain the agricultural cadastral information of the study area, an automated parcel delineation process was performed with the assistance of deep learning methods available on ArcGIS Pro.The deep learning model was performed after collecting training data and performing data training.

Crop Type Extraction
After research and investigation regarding the agricultural condition, an NDVI fixed value of 0.2 has been adopted as a No-Crop Field (Bare Soil) (NDVI EOS Data Analytics) .According to the crop calendar of Uttarakhand state (Ministry of Water Resources, 2016), the harvesting season of rice starts in October-November, which indicates that the field is empty without any crop in this particular period; in addition, wheat crop cultivation usually starts in October-November and harvesting season in April.The threshold values for vegetation and non-vegetation objects were estimated using NDVI.The NDVI threshold value for vegetation was > 0.25.This threshold value was used to discriminate between vegetation and non-vegetation objects.
According to this calendar, the rule set was framed to extract the crop type and it is shown in Table 3.Using GIS environment, the generation of the parcels was done, and the most proper shapes were automatically generated.The final and complete shape of the generated parcels is presented in Figure 7.The result of edge detection between crop and non-crop parcels was shown in Figure 7a, and the edge detection results between crop parcels was shown in Figure 7b.The visual appearance of the extracted parcels is suitable and satisfying.The final results of parcel delineation is shown in Figure 8 after processing in GIS environment.

NDVI Parcel Mean Generation
After exporting the NDVI data from GEE platform, a model was formed using modeling functions and GIS capabilities on ArcGIS software to prepare the NDVI time series data for the span of time.NDVI value was calculated for every parcel using GIS tools.Figure 4 shows the calculated NDVI from the satellite image; meanwhile, Figure 9 shows the mean value (parcel-wise) which is generated based on the automated parcels delineation.

Crop Area Assessment
According to the knowledge-Based classification, the crop area was assessed for both sugarcane and wheat.Figure 12 is presented the area of sugarcane, which reached 731250.3sq m; in contrast, the area of wheat reached 751350 sq m.Analyzing the NDVI trajectory and pattern of each crop was conducted in this study before performing the knowledge-based crop classification.The results of the parcel delineation were satisfactory and visually interpreted after comparison with the RGB image provided by Google Earth.After validation with ground truth data, the accuracy assessment of the knowledgebased classification reached 91.17% and Kappa coefficient value was 0.82.However, the drawback of this framework was represented by the limitation of the availability of updated data from Google Earth which as the parcels boundaries don't remain as the same status for every season which lead to uncertainty in crop classification, in addition, the process of reshape some cadastral parcels which leads to manual interference of the user to fix some boundaries.This work was done using the Multi Task Road Extractor Model, other models should be examined in the future such as SAM (Segment Anything Model), also the feasibility of other deep learning models for parcel extraction should be on larger scales in different places should also be included in the future work.
The area of Haridwar is around 2360 sq km between 29º35' N to 30º40' N and 77º43' E to 78º22' E. The district has a hot subhumid (dry) eco-region climate with a moderate to humid subtropical climate.Monsoon is the rainy season in Haridwar with annual precipitation of 1174.3 mm approximately (Ministry of Water Resources, 2016).The main cereal crop during the Kharif season for more than 54% of Uttarakhand's total cultivated area is rice with a 115-120 days life cycle, Sugarcane is the predominant crop in Uttarakhand, India, and has around 84 thousand hectares (Korikanthmath, Manjunath and Manohara, 2010;Kumar, Kashyap and Tyagi, 2018).The most predominant cultivated crop in the highlands of Uttarakhand state-north India is Wheat; moreover, its growth cycle extends between 180-210 days in which the sowing time is November-December and the harvesting time is April-May .The study has been carried out in a small experimental area near Sirchandi located between 77º 44' E to 77º45' E and 29º55' N to 29º56' N. The study area and experimental area are shown in Figure1.

Figure
Figure 1.Study Area

Figure 2 .
Figure 2. Google Earth Image of the Study Area

Figure 4 .
Figure 4. NDVI of the Study Area Figure 5.Samples Results

Figure 9 .
Figure 9. Mean NDVI of the Study Area

Figure 12 .
Figure 12.Crop Area4.5 Accuracy Assessment of the Crop ClassificationIn order to check the accuracy of the knowledge-Based crop classification, ground truth data were collected from the study the potential of using the deep learning algorithms to automatically delineate the agricultural cadastral parcels from Google Earth open source data.For crop identification, the investigation highlighted the feasibility of using knowledge-based times series of Normalized Difference Vegetation Index (NDVI) in crop classification compared to conventional methods represented by normal machine learning 0 on one single image, which leads to uncertainty and misclassification due to salt and paper impact.

Table 2 .
Data Used

Table 4 .
Table 4 and the UAV image is shown in Figure 13.Reference DataFigure 13.UAV RGB Image of Study AreaThe confusion matrix is shown in Table5.The error matrix is shown in Table5, the overall accuracy of the knowledge-based classification reached 91.17%, and the kappa coefficient value is 0.82, which indicates the good performance of the classification.Producer's Accuracies were as the following: 93.3% and 89.47% Sugarcane and Wheat, respectively; meanwhile, the User's Accuracies of our classification were as the following: 87.5% and 94.4% of Sugarcane and Wheat.