Transfer Learning Analysis For Predicting Soil Texture Classes From Soil Images

— Soil texture is one of the crucial characteristic in determining soil health. Classifying soil texture manually 1 is expensive, time consuming and requires experienced experts who are often limited available. Multiple machine 2 leaning algorithms are proposed in the recent past to hold up a fully automated soil texture classification in 12 or lesser 3 classes using soil images. Among such algorithms research on deep neural networks (DNNs) has been explored less. 4 Wherever these DNNs are applied, they are used in isolation. Limited efforts are made to transfer the knowledge from 5 DNN of some other application and reuse the pre-trained network. In this work, concept of transfer learning is 6 investigated in soil texture prediction. Inceptionv3, ResNet50 and ResNet152 are trained on soil image dataset which 7 consists of images acquired from agricultural fields of multiple crops. It also shows analysis of different image 8 processing based segmentation techniques.


I. INTRODUCTION
To maintain healthy food production, the fundamental measure is to sustain soil health.It is responsible to provide all favorable elements like nutrients, oxygen and water in order to keep up the development of plants for food productivity.Important nutrients strengthens soil structure, also controls plant disease, pests and weed, and this is possible because soil is home to variety of complex organisms.Healthy soil accords to mitigate climate change by maintaining the soil organic carbon.For all these reasons, preserving the soil is very important and to initiate a global cultural movement that gives soils the priority that they deserve.In the past few years, there In the recent years, there is escalation in research interest for soil related properties like texture classification [1], [2], [3] organic matter prediction, moisture content prediction [4] and ph level prediction [5], [6], [7].Soil properties play a very important role in defining which kind of crops can be grown in a particular soil.Soil texture is one of the important characteristic.It refers to the relative proportion of various soil particles of soil.It can also be said that it is the degree of fineness or correctness of various soil particles present in a given soil sample.Fraction of clay, sand and silt in a soil sample determines the textural class of that sample.The various soil properties that are influenced by soil texture Rate of flow of water through a saturated soil is determined by soil texture, it drains freely through in the soil with more percent of sand in it that some clayey soils.The availability of water to the plant is also influenced by texture of the soil, water gripping capacity of clay is greater than the soil with large fraction of sand.Healthy crop is maintained by healthy root which is nourished by well drained soil and it has good aeration which means that the soil holds air similar to atmospheric air.Sensitivity to erosion is also in of the property that depends on soil texture.Sandy soil has lower erodibility whereas soil with higher fractions of silt and clay have greater erodibility in the same conditions.Soil organic matter is also affected by variation in textures.In fine textured soil, organic matter breaks slower whereas in sandy textured soil it splits faster.This is possible with similar fertility and tillage management, environmental conditions.pH buffering intensity in a soil also varies according to rise in fraction of clay and organic matter content.There are different methods to determine soil texture like texture by feel, hydrometer method, pipette method, particulate organic matter method, rapid method etc.These methods generate accurate results but they are very time taking as well as labour consuming.Along with this, these lab methods require equipment and chemicals.Therefore, methods that can generate precise results in lesser time and need limited accessories/chemicals would be better befitted.Soil texture can be predicted on various parameters like void ratio liquid limit, moisture content [8] etc.One of the method to predict soil texture is collecting soil images to acquire soil images and apply different image processing and computer vision techniques.Inazumi et al. [9] proposed an artificial intelligence based model to classify soil into three classes, clay, sand and gravel.Images were captured using a smartphone camera (Iphone 7) and a digital camera (FUJIFILM X-T4).
The workflow was designed using convolutional neural network layers.Azizi et al. [10] proposed a research to aggregates of soil samples.In this work VggNet16, ResNet50 and InceptionV4 architectures were trained, out of which ResNet50 performed best with 98.72% accuracy.Swetha et al. [11] presented a novel setup made up of a smartphone camera, a dark chamber and a smartphone application for soil texture prediction.Image local, color, texture features are used along with random forest and CNN to predict the final soil texture.In this work, soil samples are collected from five different crops namely, wheat, rice, potato, mustard and sugarcane.To determine the soil texture of the samples in laboratory hydrometer test was performed.The objective is to propose a model that first classifies the crop type and then predicts the soil texture for each soil sample.Here, instead of conventional image processing and machine learning methods, deep neural networks are used.Conventional neural network (CNN) architectures named, InceptionV3, ResNet50 and ResNet152 are incorporated.

II. STATE-OF-THE-ART
Most of the work done in soil texture determination is done using traditional computer vision and image processing methods.
These methods use the classifiers like support vector machines (SVM), random forest, artificial neural network (ANN) etc which classify the soil into three or twelve texture classes as depicted in USDA Triangle.Wu et al. [12] proposed a comparison of different classifiers for predicting classes of soil texture on the basis of some terrain parameters located in Three Gorges of Yangtze River, China.For this objective, ANN, SVMs and classification trees with two different kernel functions.Some of the terrain parameters considered are elevation, slope, stream power index, slope height and length, terrain classification index for lowlands etc. Textural classes were sand, loam and clay.The various performance evaluation parameters were Receiver operating characteristics (ROC), area under the curve (AUC), kappa constant and overall accuracy.SVM with polynomial kernel function shows best performance with overall accuracy of 94.30%.Srivastava et al. [30] presented a vast review of soil texture classification which included methods with conventional image processing techniques as well as deep learning algorithms.Ajdadi et al. [13] presented a machine vision technique to classify soil aggregate size.This method precisely measures the tillage quality using image processing.Nine different size of aggregates were considered.Image acquisition was performed using a Canon digital camera.From the tilled soil images, textural information was extracted using gray level run length matrix, gray level co-occurrence matrix (GLCM), first order statistics of image histogram and local binary pattern (LBP).Artificial neural network exhibited overall accuracy of 72.04%.Morais et al. [14] proposed a method to predict and classify soil texture using soil images acquired from soil samples.Here, 63 soil samples are collected.The ground truth which is texture of the soil samples are determined by pipette method.This method attained 100% accuracy.Moreover, color features of a soil image is also helpful in predicting soil texture.Barman and Choudhury [3] processed soil sample images by generating various texture and color features like color moments, HSV histogram, color auto-correlogram etc. Soil samples were collected from paddy fields and soil texture were classified into 12 classes using multiclass support vector machines.Soil texture prediction has also been explored on remote sensing data [1].There are number of methods proposed using other parameters than image of soil sample that generate promising results in this field.
Since, these techniques have been over explored with time, there is need to explore other beneficial methods like deep neural networks.In many circumstances, DNNs will not have enough data to train the network, and data label development will be costly.In such problem statement, transfer learning is a very promising method where pre-trained networks are used to train and test other networks.Hana et al. [15] addressed the problem of limited training data by proposing a two phased method combining AlexNet for web data augmentation and transfer learning for image data.This method decreases the requirement of a large training dataset.Plant species classification is also explored using deep neural networks.Kaya et al. [16] compared the results of four different transfer learning models (including AlexNet and VGG16) on four different leaf image datasets.
Breast cancer classification [17] into four tissue subtype classes using histology images.For this pre-trained networks i.e.
InceptionV3 and ResNet50 are incorporated, where the former network outperformed the later one by achieving 97.08% accuracy.Deniz et al. [18] proposed a breast cancer detection method based on transfer learning and histopathologic image classification.For this, CNN models, AlexNet and VGG16 are employed.VGG16 is employed for feature extraction, whereas AlexNet is used for additional fine tuning.Remote sensing scene classification is also one of the research objective which has been surveyed by applying transfer learning.InceptionV3 and VGG19 [19] has been utilized on three different sized remote sensing datasets.These networks are evaluated with different depths.Hyperspectral image (HSI) classification is also one of the problem statement in remote sensing field.Xie et al. [20] proposed a HSI classification method on the basis of superpixel pooling CNN transfer learning.This method is built into 3 stages: 1. Convolution and pooling operation.2. Up sampling and superpixel pooling.3. Feeding hyperspectral data to the fully connected layer.This method attained 93% accuracy.Another research [21] proposed on HSI classification method by combining 3D separable ResNet with cross sensor transfer.Transfer learning is also being explored in agriculture field.Bosilj et al. [22] proposed a method to classify different crop types to reduce the retraining time using transfer learning.Suh et al. [23] proposed a study to classify and control volunteer potato in sugar beets.Six pre-trained nets were engaged in this objective: VGG16, ResNet50, ResNet101, InceptionV3, GoogleNet and AlexNet.These nets were employed to classify volunteer potato and sugar beet images taken under ambient varying light settings in agricultural environments.Next section discusses about the materials required for the proposed work.

A. Soil Sample Collection
The proposed work considers soil taken from fields of 5 different crops, which are wheat, rice, potato, mustard and sugarcane.
Samples are collected from fields of district Mau, Uttar Pradesh, India under the supervision of Department of Agriculture, GLA University, Mathura, India.The sample collecting site is located at 25°56'30"N latitude and 83°33'40"E longitude.The size of the fields from which the soil samples are dug out are of size 350 × 300 m.For a single crop 10 different fields are considered.From each field 5 sample are collected with a separation of 150 meter.Therefore, there are 50 samples of a single crop.So, there is a collection of 250 samples to determine the soil texture.These samples are dug out from a depth of 2 inches below the surface of the field.Determination of the fraction of clay, silt and sand in soil samples is calculated using hydrometer test.Then the soil texture is determined using USDA triangle.

B. Soil Image Acquisition
To create soil image dataset Motorola one power with android 10 version is utilized.This smartphone camera has a 16megapixels CMOS device.At the time of capturing soil images, camera settings are kept in default condition such as exposure time = 1/30s, F-stop = f/1.8,focal length = 1.12μm.Images are of size 3024 × 4032.Soil image acquisition setup was kept inside a closed room with artificial lighting condition using a table lamp.Images were captured from 20 inches above the surface.Soil samples were taken out on a white sheet.For this dataset, 8 pictures are captured per soil sample.Therefore, it has 2000 total soil images.This methodology is proposed keeping farmers as beneficiaries.So, a nominal smartphone camera is used to capture images as the beneficiaries may or may not have good quality cameras.Figure 1 shows samples of soil from the dataset.

C. Texture Determination in Laboratory
The objective of this proposed work is to classify soil texture for soil images of multiple crops.To fulfill this purpose, collected soil samples should be first tested out in the laboratory for their respective texture [24].Along with the soil images their respective soil texture is also needed as input to CNN architectures.For this purpose hydrometer test [25] and United States Department of Agriculture (USDA) technique [26] is performed in laboratory to generate the ground truth in manner of soil texture.Soil is composed of fraction of clay, sand and silt.This fraction or percentage of sand, silt and clay decides the soil texture like silty clay, loam, clay loam etc.To do so, first hydrometer test is carried out and then USDA triangle is used to generate the final result.Both of these methods are used in this work as they are very reliable and generate accurate results for percentage of silt, sand and clay.Determination of soil texture is done in the Department of Agriculture, GLA University, Mathura by hydrometer test.Entire laboratory procedure is performed as per the hydrometer test steps.A solution of 28 gms tablespoon sugar and 176 gms distilled water is made for calibrating the hydrometer [27].This solution is kept until the hydrometer's calibration temperature is reached.Then, within the limit of 2 mm coarse pieces, 50 grammes of each soil sample are dispersed with sodium metaphosphate and agitated.Finally, the percentages of sand, silt, and clay are computed using the For this purpose different methods are applied.Threshold method is first is used to segment the image.Figure 4 shows the outcome of applying different values of thresholds.It is clear from this image that the method operates best at threshold value 0.50.Figure 5 shows that the selected threshold value is best according to the fraction of pixels.Histogram from figure 5 clearly shows why the optimal value is in the range 0.40 and  Figure 6 shows the outcome of Otsu's method.Though the result of this method performs the task of segmentation but certain area in the region of interest is predicted as background.
To solve this this problem, using only Otsu's method is not enough in this particular dataset.To address this problem other color spaces can be explored.Any color is simply a combination of varying pixel intensity values in image's channel.So, to segment a specific color in an image, applying thresholding method on the particular ratios of pixel intensity values in the three channels can be applied.First the image is converted into RGB color space.In figure 7, it can be observed that the background has a high intensity value while the region of interest in all the channels.This makes segmenting the image in the background and foreground easier.However, this also makes segmenting the foreground much more challenging because of the much tighter pixel intensity.Figure 8 shows the result of segmented image using RGB color space.This improvised method also didn't segmented the region of interest properly.Next, HSV color space is also applied.Although, this color space can visualize the image much better the RGB color space but it did not segmented the image as needed.So, neither RGB nor HSV color space fulfilled the objective to segment the desired region of interest.Therefore, it can be concluded that the very first method (result shown in figure 9) generated the best segmented region of interest out of the soil image.So, this method is applied to all the image before providing as the input to convolutional neural network.In the proposed methodology, different convolution neural networks (CNNs) are used.CNN is a deep learning system that takes an image and assigns biases and weights to various objects/aspects in order to distinguish one from the other.In CNNs, requirement of pre-processing is low in comparison to other classification algorithm.These networks are composed of multiple layers of artificial neurons.Each layer generates feature maps.

A. Forward Propagation
CNN is made up of convolutional layers, pooling layers, and fully-connected layers, unlike multi-layer perceptron neural networks (MLPNN), which have all of its layers completely connected.Several convolution kernels are applied to calculate different feature maps in the convolutional layer, which is followed by a pooling layer.The j th feature map of l th convolutional layer,   , is calculated by - where   −1 denotes the i th feature map of (l-1) th layer,  −1 denotes the count of feature maps of this layer,    denotes the convolution kernel corresponding to j th map between l th layer and i th map in (l-1) th layer,    denotes the bias term of the above kernel, f (•) denotes the element-wise non-linear activation function which introduces non-linearity into the multi-layer networks.The activation functions sigmoid, tanh, and ReLU are the most common.
Following the convolutional layer, the pooling layer seeks to accomplish feature integration, parameter reduction, and shiftinvariance by lowering the resolution of the feature maps.For each feature map,    , downsample() is used to denote the pooling function - Average and maximum pooling are two common pooling operations.Pooling operation basically takes some k×k region and outputs a single number, which is the mean or maximum in that region.Following the convolutional and pooling layers, there are multiple fully-connected layers that try to learn mid-level feature maps.A high number of weight parameters is required for complete connection to be implemented.The feed forward process is the same as the regular ANN approach, which is formulated as - where    and    are the weight vector and bias term of the i th filter of the l th layer respectively.In a neural net, the softmax activation is usually given to the very last layer, which turns the output of the previous layer into an essential probability distribution.It's in where   denotes the count of nodes of L th layer, the last output layer, and  denotes the classification error of all output nodes for a single sample.Loss function in the form of Eq. 7 is called Euclidean Loss.Hinge loss, softmax with loss, sigmoid cross entropy loss, information gain loss, contrastive loss are some of the other options.

B. Backward Propagation
The backward propagation method propagates errors from the output layer's label prediction to the input layer.The weight vectors and bias term can be changed layer by layer based on these errors.This parameter updation can be written as- where  is the learning rate, Let    is the first part of right hand side of Eq. 6 as the error term on l th layer, and combines it with the result of second part, Eq. 9 can be reformulated as - If the l th layer is fully-connected layer and layer l + 1 is the output layer, the error term    is computed as - where  ′ (  −1 ) denotes the derivative of the activation function of layer l.If layer l and l + 1 are all convolutional layers, following the chain rule, the error term    is computed as- If the l th layer is a pooling layer and layer l + 1 is a convolutional layer, the error    is computed as - where  ′ (   ) is the derivative of pooling function  (   ) which is a linear function.So, the derivative term  ′ (   ) to 1, then the last term of Eq. 13 will disappear.If the l th layer is convolutional layer and layer l + 1 is a pooling layer, the error    is computed as - where upsample() represents upsampling operation.Upsampling uniformly distributes the error among the units that feed into it in the preceding layer if the pooling layer uses mean pooling.Because extremely slight changes in input will disturb the result exclusively through that unit, the unit chosen as the max receives all of the error in max pooling.The weight vector and bias term can be changed in an up-down direction based on prior updates.
There are several pre-defined CNNs which have been pre-trained on different image datasets and these can be used for image classification on other image datasets also.This method is known as transfer learning.Over the past years, this concept has been used by researchers in attaining different objectives based on image processing and computer vision.It is a method where a model developed for an objective can be reused for another objective.For transfer learning using pre-trained model, first an objective is defined, then a source model is selected which has performed on some challenging datasets.Then this model or parts of this model can be applied as a starting point of the new objective.The proposed model first classifies the crop and after that it classifies the soil texture class.So, for the first objective three different models are compared which are Inceptionv3, ResNet50 and ResNet152.For second objective, Inceptionv3 and ResNet50 are compared.

C. Inceptionv3
Inceptionv3 [28] is a convolutional neural network which is highly useful in image analysis and object detection.This architecture All the above points are consolidated into the final architecture.For this proposed work, Inceptionv3 is used to characterize the soil images on the basis crop and then it will be used to classify the soil textures.

D. ResNet50
ResNet [29] is a short for residual network.The idea behind its architecture (figure 11) was to stack some addition layers in the deep neural networks which later seen to improve the performance.The objective behind adding more number of layers is that these In this work, ResNet50 is used classify the soil images on the basis crop and then it will be used to classify the soil textures.

E. ResNet152
ResNet152 is another version of residual network with 152 layers.This architecture (figure 12) are more accurate than other ResNets by considerable margins.The performance is significantly gained with increased depth.The size of the input image is also 224 × 224 pixels.Here, ResNet152 is only used to classify the soil on the basis of crop.

VI. RESULT DISCUSSION
Experiments were conducted on the acquired dataset to evaluate the classification performance of the investigated networks.This work is proposed keeping in mind farmers as the beneficiary.Rural farmers may or may not have crucial knowledge about the soil in their agriculture land.Information about health of the soil in which crop will be cultivated is important.Soil nutrients, soil texture, soil moisture content and soil carbon content are some of the basic health characteristics of soil.In this work, texture of the soil samples extracted from agriculture farms of five different crops is predicted.For the mentioned purpose, first the dataset of the soil samples is acquired.A digital camera with 13mp resolution of a smartphone (specifications are mentioned in subsection B of section III) is used as image acquisition device.More complex camera settings were not used because the beneficiaries may not be equipped with a better camera.It was used in its original camera settings.For each of the five crops 50 samples were collected from 10 different fields (sample collection details mentioned in subsection A of section III).This method considers all the 12 texture classes of soil (as can be seen in figure 2).At this stage ResNet50 and ResNet152 is compared.For each crop 80% data is used for training and 20% is used for testing.In section V, these three architectures are overviewed.Overall prediction accuracy of the networks was assessed as ratio of the number of images classified correctly to total number of images evaluated.For first stage accuracy is recorded for 30 epochs in all the three networks.Figure 13 shows the training and loss evaluation of inceptionv3 architecture.As it can be observed that highest training accuracy is 97.5% while testing it reaches 100% multiple times.The model is generalized which suits most of the points.
This figure also shows the training and testing loss, which also suggests generalization occurring in this model.
Next for the same objective of classifying soil samples on the basis of crops, ResNet50 architecture is trained.Figure 14    number of samples will be less and in return the dataset will be small.In the proposed method, soil samples of five different crops are collected.For every crop multiple fields are considered.This work aims to benefit the rural farmers as beneficiaries and they may not have a good quality image capturing setup.So, images of collected soil samples are acquired using a smartphone camera.
The camera is used in its original setting while capturing the soil image dataset.Since soil images of five different crops are used to comprise this dataset, it has the diversity to create a robust model.

VII. CONCLUSION AND FUTURE SCOPE
A unique two-stage classification algorithm that takes advantage of transfer learning is proposed in this research.This method along with classifying the crops and soil texture classes, predicts the texture of the soil samples as well.Transfer learning reduces the need for a significant amount of training data to build a reliable classifier.There is no publically available soil image dataset.In the previous researches, authors have created soil image dataset using different image capturing devices.Different methods have considered distinct crops.This work considers soil samples collected from multiple crops and fields, which fills in the gap from existing research work, as an add on it shows the diversity in this dataset which later proves to be one of the reason to make a robust classification model.This method's strength is that it fully utilises current resources.When the training data is limited, this approach can be quiet useful.
Transfer learning improves the generalization ability of fine-tuned networks and alleviates the problem of overfitting.In both the stages CNN architectures, Inceptionv3, ResNet50 and ResNet152 are trained, out of the three ResNet152 outperforms the other two.
This research aims to make use of soil texture classification and prediction.There is still a lot of work to be done to improve the performance of deep neural networks on soil image datasets in the future.Some of the directions worth further research in future includes: Considering a larger dataset with soil images of multiple crops.Images can be captured from different devices which will make the dataset even more diverse, (2) GAN can be proven as an effective method of soil texture prediction when using with CNN, (3) Other soil properties like soil moisture content, soil organic matter, soil ph level can be used along soil texture as known labels in deep neural network while using soil sample images only.


Organic matter content  Water holding capacity  pH buffering capacity  Susceptibility to erosion  Cation exchange capacity (CEC)

Fig. 1 .
Fig. 1.Few examples of soil samples from dataset

Fig. 2 .
Fig.2.USDA Triangle for different samples Percentage of Sand = 100% -Percentage of Silt -Percentage of Clay (3) The findings of the hydrometer test, as well as the percentages of sand, silt, and clay, are tabulated.The United States Department of Agriculture (USDA) triangle is used to determine the texture of soil samples.The output of USDA triangle are further tabulated and represented as in figure 2. Basically USDA triangle displays the texture class of a soil sample out of the 12 classes of soil texture as dot in a triangle.Finally, figure 3 shows results for texture of soil samples as output of hydrometer and USDA triangle method for texture determination.It tells about the outcome of soil textures crop-wise.IV.IMAGE PRE-PROCESSING AND IMAGE SEGMENTATION Before providing images to any network, the soil images are first segmented to generate the region of interest from the image.

Fig. 3 .
Fig. 3. Decision on Soil texture of samples using Hydrometer and USDA triangle.

Fig. 4 .Fig. 7 .Fig. 8 .Fig. 9 .
Fig. 4. Outcome of applying threshold values on a soil image charge of guessing the incoming data's class label.Let oi and yi stand for predicted label and ground-truth label, respectively, for input sample.The loss function is usually formulated as - −   ∥ ∥ 2 ,   =   (7)

Fig. 10 .
Fig. 10.Architecture of Inceptionv3 are the partial derivatives of the loss function with respect to    and    separately which can be expanded to the format as -

(figure 10 )
allows deeper networks while keeping the number of parameters from growing too large.It is proposed with a depth of 50 layers.This architecture was build and trained by Google.It has a pre-trained version with weights of ImageNet and can classify up to 1000 categories.Size of the input image for this network is 299 × 299 pixels.In Inceptionv3 architecture, techniques for optimizing the existing inception network.The techniques include: 1. Factorization into smaller convolutions: To improve computing efficiency by lowering the number of parameters in a network.Smaller convolutions are used to substitute larger convolutions, resulting in faster training.

Fig. 11 .Fig. 12 .
Fig. 11.Architecture of ResNet50 layers progressively learn more complex features.With the addition of residual blocks to the design, the challenge of training extremely deep networks has been solved.It is proposed with a depth of 50 layers.ResNet50 was build and train by Microsoft.Just like Inceptionv3 model, this is also pre-trained on ImageNet database.Size of the input image for this network is 224 × 224 pixels.
While acquiring the image dataset 8 images of each sample were captured, so this dataset is consists of 2000 soil images of five different crops.The proposed method is made up of 2 stages.Since last outcome is to classify texture of a soil sample and dataset has soil images of all five crops, first stage is to classify the soil samples according to crops.Here, out of 2000 soil images 70% images are used to train the networks and rest 30% are used to test the networks.At this stage three convolutional neural networks are used with pre trained weights as transfer learning.Inceptionv3, ResNet50 and ResNet152 architectures are incorporated and compared for classifying soil samples according to crops.After this task is done, next is to predict the texture of the soil sample.

Fig. 16 .Fig. 17 .
Fig. 16.Outcome of ResNet50 accuracy and loss in classification at second stage shows the training and testing outcome of ResNet50 architecture.Highest training accuracy is 78% and that in case of testing is 88%.As can be observed from the results this model fits just fine, there is no issue of overfitting or underfitting.After this ResNet152 architecture is also trained.Figure15shows results for ResNet152 architecture.From the training and testing metrics it can be observed that this model also seem to generalize well enough.Highest training accuracy is 98% and testing accuracy reaches 100 TableI.Training and Testing accuracy in prediction of crops using soil images for Transfer learning

Table II .
Figure 16shows the training and loss evaluation of ResNet50 architecture for different crops.For mustard, Training and Testing accuracy in prediction of crops using soil images for Transfer learning TableIII.Comparative analysis of proposed model with state-of-the-art methods datasets which mostly considers soil samples of a single crop, which creates a gap in the study.When just a single crop is considered, sugarcane and wheat the training accuracy is 98.5% and 98% respectively.The testing accuracy of ResNest152 for all the crops is 100%, which clearly shows that it's better than the other network.TableIIshows performance of both the networks at second stage.In Table III presents a comparative analysis of existing methodologies with proposed method.Till now the researchers have proposed classification using a single objective only.This work suggests a two stage objective, where first objective is to classify the crops using soil images and second objective is to predict the texture of the soil crop-wise.Existing methods are built upon small

Table IV .
Table IV shows comparison of proposed model with existing methods in terms of precision and recall values.Comparison of exiting methods with proposed method in terms of other metrics