Decision Analyis Journal 3 (2022) 100041 ‘Contents lists available at ScienceDirect Decision Analytics Journal Peery Analytics ELSEVIER, journal homepage: www.elsevier.convlocate/dajour ac Smart farming using Machine Learning and Deep Learning techniques ® Senthil Kumar Swami Durai*", Mary Divya Shamili” Schaef Pg Prey Unveil Pale “bapa of CS. eho! of Prgmemg, Peny Uno, Bangi, ns ARTICLE INFO ABSTRACT co The price of calivatng the sal, producing crops and Keping Ivenonk sere to a faring. Agee Machin sing ‘sre co a country’s exonomic development. Nery 58 percent of « county’ primary source of liveliood Pre ‘sfarming. Farmers tl date bad adopted conventional farming eens, These teclgues wee nt prelse hus reduced the productivy and consused «lot of tne, Precis farming help to increase the oducts by precisely determining the steps that needs oe practiced ats due seas, Predicting the weather conditions, oalyaing the so, recommending the eps for culvation, determine the amount of fertile, pestilds that heed to be used are some elements of pression farming. Preise Farming uses advance techologes such 5 1OT, Dat Mining, Data Analytics, Machine Learning to collec the aca, an the systems and predic the ‘esuls With the hep of technologies Precise farming helps to reduce manuel labor apd increase prod "Farmers have been facing varius challenges in shes recent ties, hs neues crop ftlure due oles infertity of sol and soon. Due to the changes taking place inthe environment the proposed work helps 19 ‘detfy how to manage crops and harvest ina start way It guldes an individual fr smart farming, The aim ofthis work is to help an individual clivate crops efficiently and hence achieve high product at low ot It ls helpe to predict the totl cost needed fr cultivation, This would help an indvigual to pre-plan ‘he activites Before cltivaton resling in an integrate solution in fring 1. Introduction Agriculture, or farming asi is commonly know, isthe practice of growing crops and raising cattle. It contributes greatly to a country’s ‘economy. Many raw materials and food products are produced by agriculture. Raw materials such as cotton, jute is used by industries for ‘manufacturing various products that is used in day-to-day life. Agei- culture not only felps for food production but also produces resources ‘needed for creating commercial products. Agriculture used traditional techniques for cultivation of crops. Conventional or traditional farming ‘is mosly practiced all over the world. It involves techniques suggested by experienced farmers. These techniques are not precise hence results {in hard labor and time consumption. ‘The application of digital technologies which includes robots, elec- tronie devices, ensor and automation technologies is associated with Precision Agriculture. This technology aims to reduce workloads, in- crease profitability and decision management. Precision agriculture adeitionally referred to as precision farming i a farming control system ‘that provides @ comprehensive approach to deal with the spatial and temporal crop and soil variability to maximize profitability, optimize yield, improve quality of production (1). Precision Agriculture is an ‘ficient wey to improvise the yield. On discussing about the adoption ‘ate of precision agriculture, the high value enterprise farms adoption * coresponding au rate was more compared to low value enterprise farms. The adop: tion rate of precision agriculture also depends on the country and the geological locations, The adoption rate of Precision Agriculture fn mountain zone is less compared to farmers in the valley (2). The variation of adoption rate is due to the high investments needed. Hence there needs a way to reduce cost on machines hence all farm-size an cope precision agriculture. Precision agriculture is aided by advanced technologies such as 11, Data Mining, Artificial Ineligence, and Data Science. The Internet of Things (101) isa network of interconnected computational things like sensors and smart gadgets that can communicate with one another and share data (3). In agronomic applications, wireless sensor networks fare being used to remotely monitor ambient and soil characteristics in order to predict erop health. Using WSN as a forecasting approach, the watering schedule of agricultural fields can be predicted. Wireless Sensor Networks aequite data fom external varables such as pressure, hhumidity, and temperature, as well as soil moisture, salinity, and conductivity (41 Machine learning makes agricultural applications incredibly eff cient and simple, Data eequsition, model building, and generalization are the three stages of the machine learning process. The majority fof cases, machine learning algorithms are used to deal sith complex. Bal addres boson} 224@yaho0 coin (SKS. Dari, mary 202024000 Gpreidencyniversiy a (MD. Shami, org/10.1016/.joue. 2022,100081 Received 21 December 2021 Received in revised form 1? March 2022; Accepted 29 March 2022 2772-85227 2022 The Asthorts), Published by Elsevier Ine This ean open access article der the CC BY-NEAND lense Chup “ereaivecommonsorg/tcenses/y-ne-nd/4.07.problems when Iman competence is insufficient. Machine leering may be used in agriculture to forecast soil paramecers like organic ‘carbon and moisture conten, as well as crop yield prediction, disease and weed identification in erops, and species detection (5]. Traditional ‘machine learning i improved by Deep Learning by adding additional complexity to the model and changing the input with various func tions that allow data representation ina hierarchical manner, through, multiple levels of abstraction, depending on the network architecture employed, A significant benefit of Deep Learaing is feature leaning, ‘of Indeed the automat extraction of features from origina data. The ability to identify unknovn things such as anomalies rather than just ‘collection of existing items is a key aspect of the Deep learning ‘model, which uses the homogenous properties ofan agrieultural field to discover faraway, badly obstructed, and unknown objects [6 Blockchain has swiftly become a key technology in a variety of precision agriculture applications. The requirement for smart peer-o- eer systems capable of verifying, securing, monitoring, and analyzing agricultural data has prompted researchers to consider developing Dlockehain-based loT systems in precision agriculture. Blockchain plays 4 exitcal role in transforming traditional methods of storing, sorting, and distributing sgrieuttural dats ato a digital format a way that Is more dependable, immutable, transparent, and decentralized. The com- Dination of the Internet of Things and blockchain in precision farming, reslts na network of smart farms. More autonomy and flexibility are attained asa rest ofthis pring (71 ‘The sbove-mentioned technologies such as IOT, Data Seience, Ma- chine Learning, Deep Learning, Blockchain deals mostly with dara ‘hich are very useful for understanding and providing great insights of data, Hence, these advanced technologies are used in various agricul- ture practices sch as identifying the bes op for a pacicular locaton, identifying factors that would destroy the erops such as weeds, insets and crop diseases co obtain Insights about the erop growth and help in decision making. ‘Agriculture can be divided into 7 important steps that includes ‘Land Management, Soil Preparation, Water Monitoring, Ideatifvng the ‘weeds, Pesticides Recommendation, identifying diseased crops, and cost estimation. Land Management refers to the monitoring physical features that includes weather conditions, geological characters, ‘hiss important since there are variations in climatic conditions across the globe which would affect the crops. Rainfall san important aspect fof the earths climate, and its unpredictability has a direct impact on agriclture, water management systems, and biological systems [8]. AS ‘result, tool that assist in predeting rainfall in advance are required ‘0 tha erop management can be simplified. Soll san essential component of agriculture. Rooting, moisture and nutrient storage, mineral reserve, anchoring, and a variety of other variables that affes plant growth are all determined by soil depth [0]. ‘The initial step for Soil preparation is testing the soil. It involves identifying the soi’s current muient levels and the suitable amount ‘of nutrients to be feed to cerzin soll based on it ferdity and crop demands. The values from the sol est report are being used to catego rize a number of key soil parameters, notably Phosphorus, Passium, Nitzogen, Organic Carboa, Boron, as and soil ph (10). leigation is a type of agriculture chat plays an important role in water and soil conservation. Complicated data could be used to maintain ergation performance and consistency when assessing systems with respect to ‘water, soi ellmate, and erop faci (111 ‘Weeds are plants that is grown where its not needed. It inchudes plants that are not intentionally sown. Weeds compete for water, nu ‘Wiens, light, and space with agriculrural plants, lowering crop yields. ‘Weeds can diminish the commercial worth of agricultural regions by lowering the quality of farm products, causing irrigation water loss, and smaking harvesting machinery harder to run. To control weeds farmers often spray bomogencous herbicide spraying througout the fed twice fr three times during the growth season. However, this method has resulted in the uncontrolled use of lage volumes of herbicides, whieh is harmful to humans, non-target animals, and the environment (12) econ Apes Sora 32028) T00081 Plant diseases can have a devastating influence on food safety, as well as a considerable lass in both the quality and quantity of age cultural goods, Plant diseases can potentially prevent grain harvesting entirely in Severe circumstances. AS rest, in the field of agricultural {nformation, computerized kdentifieation and diagnosis of plant dis ceases is widely needed. Many approaches for doing this problem have been offered, with deep learning emerging as the preferred method because to its excellent performance [13 “Hence this work focuses onthe steps involved in cultivation of crop. It uses Deep Leeming and Machine Learning algorithms to deliver so lutions fo various challenges faced during cultivation. Ie mainly focuses ‘on recommending the erops based on weather parameters, sigesting the nutrients requirements and specifying the Growing Degree Days. 1 also helps in identifying the weeds and recommending herbicides for the same. Many insects ruin the erops hence pesticides are ree ‘ommended based on the insects that are present in the field. And finaly cst estimation is very much necded in these recent times. Crisis, uncertainties would resule in great loss. Hence forecasting the cost for cultivating a crop is necessary to plan for future uncertain events. This ‘work specifies various core in cukvation for future years, 2, Literature review Crop growth i primarily influenced by the sol's macronutrient and trace mineral content ofthe sil. Sol being the broad representation of several environmencal factors including rainfall, humidity, sunlight, temperature and soil ph. The use of a support vector machine and decision tree algorithm to distinguish the type of erop based on m) cronutrients and meteorological characteristics has been presented a8 an efficient means of predicting the crop. Three crops where selected suchas rice, wheat and sugarcane. Based on certain observations details bout micronutrients where been obtained. These details where foed into the classifier model that in turn predicted the crop based on the passed values. There are many Machine Learning algorithms that works in a different mancer, Hence selecting only two madels will not provide the required output. The accuracy score of SVM was greater than decision tee algorithm with @ sore of 92% [4]. In this work best out of two algorithms is selected. But there are various algoriths dedicate for classification tasks There isa need for working on other rmadels such as K Neighbors classifier, Logistic Regression, Ensemble classifiers, These algorithms are indeed applied in proposed research work. The (14 predicts only a crop based on the values entered into the SVM model. Data is mos valuable, Hence more information can be obtained apart from using them for prediction. The proposed research work not only recommends the crops and also uses the data to obeain various information that would provide a detalled view about the predicted crops this includes specifying the Growing Degree Days such as heat units, amount of heat needed for the crop growth and the amount of nitrogen, phosphorous and potassium eontent need to be supplied for the growth per 200 Ib. frilizer. “Machine Learning algorithms such as SVM and decision tee cas: sifier was used [14] but in this work Machine Learning algorithms such as Decision Troe, K Nearest Neighbor, Linear Regression mode, Neural Network, Naive Bayes and Support Vector Machine was used for recommending a erop to the user. It has provided an exposure to other algorithms compared to [14]. Linear Regression model was sed to prediet the production value against the climatic parameters such as rainfall, temperature and humidity. The scores of all these algorithms were below 90% (15]. This work was just a model implementation sing che dataset. Web interface needs to be implemented so that feven common people can use i efficiently, All che values need to be provided manually forthe model to predict the crop. The proposed work helps in extracting temperature and humidity values using Web Seraping. Hence manually entering the values are not needed. The proposed work provides an interactive web interface where the user Specifies the average rainfall and sol Ph value, The temperature andInumidity details are extracted automatically and feed into the best model that Includes 10 algorithms with hyper parameter tuning. The proposed work tends to achieve an accuracy of 95.459% with hyper ‘parameter cuning the algorithms which was not ineluded in [1]. The predicted results along with certain information are displayed in the ‘web interface which makes the user to understand the results more ceficiently ‘Base temperature of a given crop ean be used to calculate the ‘GD Growing Degree Days. The main alm of this study is to come up with easy and mathematically acceptable formulas for calculating {GDD's base temperature. Temperature data for snap beans, sweet con, land cowpea are used to propose, prove, and test mathematical for- mulas, These new mathematical formalde, in comparison to eacier approaches, can produce the base temperature quickly and correctly These formulas can be used to calculate the GDD base temperature for every rop at any developmental stage [16]. This work provides a {formula to calculate the GDD for the crops. Hence the formula specified in 16] was applied tothe predicted crop to estimate their GDD in the proposed work ‘Weeds grown along with soybean can be detected using K-means and CNN model. Kemeans were used for identifying the features ofthe images and convolutional neural network for was vsed for classifying the weeds and soybean It also suggests that accuracy can we improved by fine tuning the CNN model. CNN model provides an efficient way to detect the weeds present among crops. When used along with K- ‘means intially the images and its augmentatons are clustered and fon using CNN model helps to precisely identify the weed [17]. The proposed work uses the pretrsined model such as Resnet152V2 hence it has important layers suchas skip layer and identity layer. The main ‘goal of these layer is to make sure tha the output image is same as the input This increases the accuracy and the predictions are correct. Not ‘only predicting the image the proposed model also helps to provide ‘etals about the herbicides that can be used which is an addtional information forthe ser. Existing deep learning techniques are used for woed detection ‘This study provides information of various ML. and Deep Learning algorithms that ean be used for identifying weeds. It mainly emphasis fon pre-trained models. I¢ suggests that pre-eained models as lot of Deneits and hence can be used to image clasiieation. I also provides ‘guidance of how to work on datasets and make the datasets efficent {or bullding the models. Many public datasets are available on various platforms that can be used for this purpose. It specifies Image Resizing, data augmentation, image segmentation some ofthe techoiques would bring about accurste classifications and tendency of increasing the accuracy is also more in pre-tained models (18). Since this study provides directions to perform deep leaning techniques the proposed model has opted cerain techniques preprocessing steps sucha Image Resizing, data augmentation is opted before building the actual deep Tearing model to predict the weeds ‘Another algorithm that can be used for identifying weeds in veg- table plantation isthe CenterNet. CenterNet is used for weed idenif- ‘ation, It includes two stages. In frst stage the Bok choy images were collected and dezected. tn the second stage, colorindex based segmen- tation were performed on the images collected to identify the weeds present in the dataset. The images were collected from Nanjing, China The images were augmented to increase the dataset size and images were annotated. CenterNet algorithn was used for both training and testing the images. Iisa ground-based weed identification technique. ‘More optimization would lead to better results was suggested (19. CCenterNet algorithm is simple yet there is a need an algorithm that strives to get correct prediction. The proposed work uses ResnetS2V2. lalgoritim that strives to achieve more accuracy since it has special layers such as skip layer and identity layer that tres to get input image as output itself. Hence predictions would be absolutely correc. Hence Resnet152V2 algorithm is selected to obtain accurate prediction and ‘based on the prediction obtain the list of herbicides, econ Ans Sora 92022) 100081 s s & Farmers face a challenging task in identifying erop inseets since pest infestation destroys a substatil portion of the crop and affects ts quality.The use of highly skilled taxonomists to correcly identify ingeets based on their physical tralts is @ shortcoming of traditional Insect Kentifietion. Experiments were conducted using image char acteristics and ml algritnms such as neural networks, support vector machine, knearest neighbors, neve bayes, and convolutional neural network madel to identify twenty-four insects from the Wang and Xie dataset. To inrease the performance ofthe classification models, ‘fold cross-validation was used. The CNN model had the greatest classification rates of 91.5 percent and 90 percent, respectively. The results revealed a considerable improvement in classification acuraey and computational time when compared to state-of-the-art clasifie tion algorithms (20). This work [20] has used basic CNN model for classification as well as the same dataset used by various researches. Hence the proposed model has used a different dataset called the Pest’ dataset fom Kaggle website. This dataset consists of 9 classes of insets. Bach image is taken from different locations. This dataset was selected for the proposed model since the model is trained of images about various locations that gives more knowledge for the model to understand che image and distinguish them. The proposed model uses Resnet152V2 mode! for eassifeation, The Resnet152V2 model isthe basic model and top of which Global Average Pooling 2D, Dropouts and sare hidden layers ae been implemented. This refers to fine tuning the bse pre-trained model. Tis helps in extracting more information and Delp inefficient classification, The association between the degree of difficulty in identifying in sects and the identification key was investigated in this article, For a collection of 134 insects, the SPIPOLL database was wtlized to generate 198 characteristic value pathways. Based on the average TES of all the insects with that of characterise value was formulated. The CV's derived IES was then used to generate an estimated IES for each bug, resulting ina ranked list of insets. Finally, the anticipated bug ranking lise was compared to the actual bug ranking list. The results showed significant correlation between the estimated and actual truth TES, indicating that the CV can be used to estimate the IES of SPIPOLL. Ingects (21. This work has specified of how to consider the features ‘of an image with respect eo insects’ dataset. Its main goal isto identify key that helps in distinguishing the classes, This proposed work contributes in specifying that a key fs important for distinguishing the inset classes. Hence the proposed work uses Resnet1S2V2 algorithm for this very reason, ResnetS2V2 isa pretrained model and it auto ratially picks the important features rather than manually defining them, The Resnet152V2 base model on addition with Dropouts helps sn removing unnecessary hidden layers and selecting the relevant ones fs an advantage. Identifeation of insects does not solve the prob Jem completely. Suggesting Pesticides provides 2 complete solution ‘The proposed model els to identify the insoets as well a suggest Pesticides for the same.SKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 Fg. 2.1.1, Cop ommendations architect La Se oS aT) Smuree——tp:/vwckagae.comvathary ainglerop-ecommendaton- dataset Seot ——2200 ‘Atenas | ator Classitieaton atk come 39 Tntps/iwwockaggle.c htp:/nwwckagale.c ‘omstekharyada/rop on/aj021977/erop- ssoilesv names a 35 2 2 Casiication Clasiiation 7 3s ‘Various elements must be considered when estimating the cost of ‘crop, Ie divides agricultural costs into five categories and provides caleulations for exch, It also gives examples of how to gure out ‘how much a crop cost Is a theoretical article that always guide the Implementation of estimating the cost of cultivation (22). This ‘theoretical study was used inthe proposed mode! co calculate the cst ‘of cultivation. It was very helpfil as it provided elementary description to calculate the costs for cultivation, The formulas proposed in this study was used in the proposed system to estimate the costs tll the year 2028. From 2004 to 2015, the goal ofthis research isto evaluate the gap ‘between various costs and gross value of output (GVO), as well asthe ‘wenés of input utilization and critical factors fr gross value of output of gram crop across top production states. The findings demonstrate 0 after 2009-2010, sll sates’ GVO and overall costs inereased signif ‘cantly. The commencement of the Government of India’s agricultural waiver scheme in 2008-2008 was found to be the eause of a large {increase in operational costs from 2009 to 2010. It was aso obvious ‘thatthe compound annus growth rate was larger in 2009-2010 than in 2014-2015 when comparing 2004-2005 to 2007-2008. Profit margins ‘were high in Madhya Pradesh and Rajasthan, indeating a cost-cutting ‘end [23]. This work provided a comparative study about the cost Derween diferent states of ina. The proposed work has sed Ensemble regression algorithms that is used to forecast the costs tll 2028. 1. provides a comparative study ofa crop fora speeife stave from 2010— 2028, Hence the user would be able to identify the trends of cost From the year 2010-2028. The forecasting is explicitly applied on the India’s cost of cultivation survey data from 2010-2018. This provides an elaborative view of operational cost, xed cost, total cost, Cost ‘Concepts were displayed in form of a graph for better understanding ofthe tends ofthe casts. 3. Methodology ‘There are four modules proposed in this work Fig. 3.1 such as erop recommendation, weed identification, Pesticide recommendation, crop cost esimation. The proposed work is a Web application developed through Django framework. The Web Interface starts with the User login page. In order to access these modules, users need to faltilly register with ther basic details such a8 their name, Address, County, State, Pin code, Phone number, Username and Password, Once the account i reated, they are redirected to the login page where the user ‘needs to login using their credentials The following Sections describes the modules in deta 3.1. Module 1: Grop recommendation Datasets Used: For Crop Recommendation module the dataset used are Crop recommenderesy, soilcsy, steatifiepamesese. All these datasets were obtained from Kaggle website. Fig. 3.1.2 gives the sum sary of the datasets used in this module (See Fg. 3.1.1) ‘The Crop recommendation was used for training model since it contains attbutes such as temperature, humidity, average rainfall, soil Ph, nitrogen requirement ratio, potassium requirement ratio and phosphorous requirement ratio essential for predicting a crop. The datasets such as Soil names and Crop names are used after prediction to obtain the sol type and scientiic name of the predicted crops. Steps involved in Crop Recommendation module are as follows Step 1: Importing Libraries and Dataset In order to ulize Machine Learning algorithms and preprocessing tools specific libraries needs tobe imported. Using these libraries, che ‘model building and prediction would be performed efficiently. The Iibrares such as NumPy, pandas, pickle, matpotib, seaborn, LabelSKS Dial and MD. Stan NP K temperature 0 0 2 495 58 41 24.770862 2 60 55 44 3 74 35 40 4 72 2 2195 107 34 32 2196 99 15 27 27417112 2197 118 33.30 2198 117 32 34 2199 104 18 30 bumidity ph 26491096 80.158363 6 980401 20130175 81.604873 7.628473 262717340 rice 2.774697 66.413269 6.700064 59.636962 24191797 67225123 6.362608 26272418 52.127394 6.759793 23,602016 60,396475 6.779833 econ Ayes Sora 3 (2022) 100081 rainfall label | 20879744 82002744 6502085 202935536 rice 80.319644 7.038096 226655537 rice 3.004459 82.320763 7.840207 263964248 rice 242,964024 rice 177.774507 coftee 6.086922 127.824610 coffee 173322899 coflee 127.175289 cotter 140937041. coffee missing values " P k ‘temperature humidity ph rainfall label dtype: inte escocece Fig, 814, Mising valve dete Encoder, train.tesLsplit were imported. The models such as Naive ‘ayes Logistic Regresion, SVM, Decision Tree Classifier, Bagging Clas sifier, Random Forest Classifier, AdaBoost Classifier, Gradient Boosting Classifier, XGBoost Glaser, LGBM Classifier and KNN was imported ‘The dataset called crop recommendation was used inital for tesaing and testing the models. Fig. 3.1.3 shows the glimpse of the crop recommendation dataset. Fach crop has a set of values for temperature, humidity, rafal, Nitrogen, potassium, phosphorous. ‘Step 2: Descriptive Analyse "To obtain a best predictive model descriptive analytics i to be performed in prior. Descriptive analytics provides an idea of how the dataset looks lke and helps to draw new insights. Once the dataset is imported, missing values per attribute is checked. For erop recom- ‘mendation dataset the attributes are free of missing values, resul are displayed in Fig. 3.1.4 Once identifying that there are no missing values, n Fig. 3.1.5 the datatype of the atibutes are identified fol- Towed by isting the unique valves in the dependent variable, .e, Label attribute Fig. 31.6 ‘Step 3: Dota Visualization ‘Once the basic details about the dataset is obtained data visualiza- ton is performed o analyze the dataset ina visual format. A correlation ‘matrix isa relationship lattice fs basically a table appearance that spec ifles the correlation coefficients between atibutes, Here, the attributes fare adresse inthe fist line, and inthe fist column, Fg. 2.1.7 shove the correlation matrix forthe dataset types " intoa ° intsa x intea temperature floated funidity floated ph floatea rainfall Floatsa label object tunes abviert Fig. 3.1.5, Data of ech clam ig, 21.8 displays the distribution of crops instances in the dataset ‘The fg signifies thatthe dataset has all crops equally distributed in the dataset, Bach atebute in the dataset is potted against the dependent variable such a6 Label column. Taking into consideration Nitrogen requirement per each crop the dataset specifies that cotton requires ‘ore nitrogen forts growth compared to all ather crops. These cetails, fare obtained from Fig, 3.1.9. In the same way Fig. 21.10 specifies the Potassium requirement for each erop in which grapes and apple being the highest and orange consumption being the least followed by Fig, 3.1.11 specifying the Phosphorous requirement for each erop in which grapes and apple being the highest. ‘With respect to Temperature requirement for each crop papaya requires more temperature (Pig 31.12), ice needs more rainfall com pated to all other crops (Fig. 3.113) and coconut requires more hi ridity compared to other crops is demonstrated in Fig. 3.1.14. When analyzing the soil Ph requirement per crop almest all erops require ‘mare Ph value which is demonstrated in Fig, 2.1.15 All the atibutes except Label attribute are numerical in nature. Dis ‘eibution plots is primarily used for unvariant sets of data and depicts information using histogram. Hence, distribution plots were plotted to ‘identify the distribution of data throughout the dataset. In Pig. 3.1.16 the values of Phosphorous attribute as not equally distributed. There fare ups and dowas in the distribution of data which is specified by the blue line in the graph. With respect to Fig. 3.1.17 the potassium columa has more zero values Values lis between 20-45 and 75-120 Jn Fig, 21,18 that deserbes the distribution of data for Nitrogen. ig, 8.1.19 demonstrates distribution of data for Temperature Col- ‘umn most of the valves les between 20-35, It also shows normal Aisrbution of data In Fig, 2.1.20 the values are not much distributedSKS Dial and MD. Stan unique crops econ Ayes Sora 3 (2022) 100081 ['rice’ ‘maize’ ‘chickpea’ ‘kidneybeans’ ‘pigeonpeas* ‘othbeans” ‘mungbean’ ‘blackgram' jute’ ‘coffee"] ‘lentil’ ‘pomegranate’ ‘banana’ ‘mango’ ‘grapes’ watermelon’ ‘musknelon’ ‘apple’ "orange papaya’ "coconut" ‘cotton’ 1M. 83.6. rps presen inthe dat ig. 3.19. Carlin mae rice 100 maize 100 jute 100 cotton 100 coconut 100 papaya 100 orange 100 apple 100 muskmelon 100 watermelon 100 grapes 100 mango 100 banana 100 pomegranate 100 lentil 100 blackgram 100 mungbean 100 mothbeans 100 pigeonpeas 100 kidneybeans 100 chickpea 100 coffee 100 Name: label, dtype: intea {in rainfall, With respect to fig. 3.1.21 the values are distributed more ‘within the range 8.5 to 75 for sol Ph column All she distebution poss depicts thatthe data for each columa is not normally distibuted and sts value counts vary between ranges. Step 4 : Outir Detection and Outer Treatment ‘An outlier is a data point that is discovered to be considerably diferent from the other values for a given set. These outliers may corrupt the dataset and may provide wrong predictions. Hence before passing the data into the model outliers must be detected and tested. Im order to detect the outliers, the proposed model has used box pots and IQR Technique. Box plots are used to visualize the outliers in fan attribute. Fig. 3.1.22 depicts the box plot for Soil PH ateibure. ‘The figure shows that the values above 8.5 and below 4.5 are been categorized as outliers, Inter-Quantile Range Technique (IQR) is used to detect the outliers using a quantile range which specifies the per centage of data that is outside the quantile range between 0.75-0.25. In some cases, outliers can be removed or may be included because it is important in the business perspective. In crop recommendation. dataset the outliers detected are urefel and must be included because the observations or tuples in the dataset is obtained using experiments and these observations corresponds to particular crop growth deta {Asa result, this work has included all the observations present in the dataset since itis very important for prediction. Step 5: Label Encoding ‘The dependent attribute or variable ‘Label’ in the dataset is Label ‘encoded, This attribute has categorical and non-aumerical values. The ‘Label’ attribute contains names of the erops. Since itis non-numerieal the values need fo be encoded into numerical values as many Class fication models does not encourage the use of non-nummerical values Hence, these values are encoded into numerical values and then fed {nto the model for future predictions Step 6: Spiting the data into Train and test sts ‘After detecting the outliers and visualizing the stebutes, the fast step for building the model is splitting the dataset into training and testing set. The intial data required for training machine learning algorithms is referred to as training data. Machine earning algorithms are fed training datasets to leern how to make accurate predictions oF accomplish a desired activity. Testing set defines Set of deta used 10 provide an accurate evaluation ofa finalized model fit on the training sample. The training and testing set split ato is 5050. Sup 7: Mode Building ‘This work aims to identity the crops using Machine Learning Clas sification algorithms. The proposed work uses 10 classification algo "ithms to find the best mode for future predition. The steps performed for model bulding are as follows + The model is being imported from the library. + Model is being defined, + The training and testing data are ited into the model. + After taining the mode, the models being tested over the testing dataset, + Confusion Matrix and evaluation metrics are calculated. ‘ig, 3.23 provides results about Random Forest Classifier. The pre dicted values ofthe testing set are displayed inform ofa list pointed by y_predictions are then followed by traning set accuracy score, test et accuracy score and the accuracy score of that model for the crop recommendation dataset. Classification report is depicting theSKS Dial and MD. StanSKS Dial and MD. Stan Deco Ayes Sora $ (2022) 100081 Fg, 3.1.18, Huy eglemen of ech op. Pig. 91.14 Ral euiement of ech ep ig 31.15. So Ph element ofeach rp. ig, 21.16, Dias data for Phosphorous Clu.SKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 Accuracy Evaluation 10.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 0.00% 0.00% me Proposed Accuracy eo baisting work re ee 1 Geteooinonle ob SETI) SUIS HONS sap gig ima 2.49 in a cir tt ei 2 eae amma ie nes past 2 sitemeter ssp t ein sn 0h, 2,120, 1400 ‘eho ‘ede + cia 00200) 1 Steet ig. 21.28, Summary ofall the hyperpsracte ied o herepecive goin Aigorsthe adel aezuracy © GSCVKNN gid ob kon 85545455, 4 @SCVNE grid. obi.nb 94727279 2 GscvoT —ord_otat sastete2 3 GSCVBG —gnd.obLtg 90696364 4 GSCVNB1 —grd_obLm> 94727273 6 RSCVRF — random 95454545, © GSCVAF ard soaen 04545455, 7 GSCVADA gid cb ata 50181816 8 Gscv.csc aid obj abe 90272727 ° @scvxcs mab tune 92080800 ig. 2.1.30, Mace! performane summary with yperpcineer ing precision, recall end FI score for the epplied model. The same steps are followed fo all the 11 algorithms and the corresponding predicted values, raining set accuracy, resting set accuracy, classification report fs obtained, Fig, 3.1.24 shows the confusion matrix for the random forest clasifier. The diagonal ofthe confusion matrix shows the how ‘many times the prediction was correct. Accuracy is calculated using this formula (ig. 31.25) using the formulas obtained inthe confusion ig. 81.29 gives the summary ofthe performance ofall the models ved in the proposed work. According to this summary itis been observed that most ofthe algorithms has accuracy above 9036 In order to obtain best model for prediction all the models whose accuracy above 90% is picked and hyper parameter tuning is performed. ig, 3.1.81 shows a comparison of accuracy between Proposed work and Existing work. The existing work has used les algorithms and thelr accuracy are displayed in color orange. Any algorithm is been measured ‘using accuracy to obtain a est model. Based on the model evaluation all the algorithms used by existing authors sem to have less accuracy. ‘The proposed work has used 11 algorithms in total to obtain a best ‘model fr further analysis To further excract the model’ eapabilty this work has used hyperparameter tuning,SKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 Accuracy 120 300 eAccuracy Fig. 3.1.31. Mode! perooaocs summary wth hyperarte y Step 8: Hyperparameter Tuning ‘Hyperparameters are the key factors that determine the model architecture, and hyperparameter tuning seems to be the process of dlscovering the optimum model architecture. Hyperparameters were ‘used to tne that filtered algorithms. ig. 31.28 depicts the hyperparameters used for the algorithms above 90%, Different algorithms have diferent parameters to be de- fined hefore going for training the algorithms. Once these paremeters fare defined the model is defined along with adding these parameters to itusing Randomized CVS or Grid Seerch CV algorithm. Once the model is defined using these parameters as shown in Fig. 3.1 28 traning data {is used for training the hyper tuned model followed by testing them using the test set, Accuracy for the hyper tuned algorithms isealeulated and hence used as metric for deciding whieh algorithm to choose for Tuture predictions (see Fig. 8.1.29). Some algorithms ater hyperparameter tuning its accuracy decreased Dbut when Random Forest Clasifier was hyper tuned using Random- SearchCV increased the accuracy to 95.4545% which is shown in Fig. 8.1.29 (see Fig. 3.1.20), ‘The Fig, 3.1.26 depicts the accuracy comparison among algorithms ‘with hyperparameter tuning. This shows that most of che algorithms ‘when hyper tuned tends to provide more accuracy. Hence Random Forest Classifier hyper tuned using Randomized Search CV is been. picked and use for further analysis. ‘Step 9: Preetng the crop ‘On identifying the best model for predicting the erop. The model 4s sored ina pickle file. Pile file is used Co serialize Python objects architetuces, which isthe method of transforming an objec in memory toa byte steam tha ean be stored on drive as a binay ile. This binary file can be de-serialized back toa Python object when we load it into a Python code. Hence, the model is stored into a pickle fle so just on calling a pickle file we can predict the crops. Values of temperature, humidity, average rainfall and sol Ph nced 1 be passed into the model to predict the crops. tn order to obtain ‘the climate conditions such as temperature and humidity the latitude ‘ane longitude of the current location is identified using Web Scraping the website “hitps://pinfo.o/. This Is done by requesting the pro- ‘gram to open thi website which uses the IP address of the device to ‘obiain the details on the current location. The values extracted are Tongitude and latitude values. Extracted values ace passed 5 input 1 the OpenWeatherAPI which provides the details about the eurtent (se mn temperatre <-bse temperate (ODD (tse temper + maxi epee) - ese temper ‘GDD = (minimum temperate * maxsamum temperature) - base temperature 2 (Cae 3: manu temperate > tae man tps ‘GDD= inn tpetice + be msm epee - bse epee 2 Fig. 31.32 GD value ela, Ninoge concentration fr 20 Feliz: 00% Mean ale of Nope ts) Poo consent fo 20 Fete 0 Mean vale of Nirgen xi) ‘ee contain fr 200, Felizer=200 Mn va of Nanogen i) Fig. 3.1.33, Naviet concession cession location. Temperature, humidity, place, region details are fetched. The average rainfall and soil Ph value is entered by the wer through a Web ‘terface. On integrating all the parameters and is passed into the model {for prediction. The model prediction probability is been fetched and ordered in descending order and the erops with highest probability is been selected and values are decoded to get the actual crop names. ‘Thus, the model displays the Top 5 erop names that can be grown in the current location. This work not only provides recommendations of, ‘cxops but also suggests the GDD-Growing degree days and the amount ‘of NPK — Niteogen, potassium, phosphorous required fr op growth.SKS Dial and MD. Stan Deco Ayes Sora $ (2022) 100081 Fig. 2.2.1 Weed Mentfcatin sem anes, Black grass Cleavers Common Chickweed Scentless Mayweed Shepherd's All purse Small Flowered Cranesbill Lossy silky bent sane pera un voolael V2 planed ase i, 82.3, Daas deals for Vee Mestexon wateSKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 Mosquito Sawfly Stem-borer ig, 3.8.2 Semple ages of posts date fa ‘nsw aga com/simranolmesiajpestatset Nestmeln 3150 on mage Dutt ett Casson aco 9 Tg. 3.3. Daas dei for Foe Ideienton and Penile Recommendation ‘and Growing degree days (GOD) are @ weather-based marker that can bbe used to gauge crop development. Its a crop producer ealeuation that isa measure of temperature levels that is used to anticipate plant and pest growth rates, such as the time that a crop achieves maturity ‘The calculation of Growing Degree Days allows farmers to forceast the tate at which their plans Wil mature. Fig, 5.1.27 demonstrates of how ‘the GDD values aze calculated foreach predicted crop, Respective Minimum and maximum values are obtained fora crop from the dataset. Hach crop possesses minimum and maximum base temperatures. These temperature values were been obtained from of- ficial agricultural websites and were stored ae a dataset. These values ‘were been applied to get the GDD value for crop (see Fig, 3.1.81) Fg. 34.1. Co einaton site echitecise, Nitrogen, Phosphorous, Potassium content is necessary for erop growth. Mean value of nitogen, Phosphorous and potassium was ex tracted fora particular crop and was divided by 100 ane was multiplied by 200 to obtain the nutrient concentration for 200 Ib, fertilizer per hectare is shown in Fig. 3.1.28 see Figs. 3.1.52 and 3.1.33). 3.2, Module 2: Weed identification Dataset: The dataset used for this module is v2 plant seedlings dataset from Kaggle (see Fig. 9.21) This dataset was used since it has images of various weed growth stages, Tiss very important since it has images of all weed stages. Any weed at its intial stage can be captured and uploaded into the sytem.SKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 8 Forecasting Inps://eands dacnetnicin/Cost_of Cultivation him Serco font of Seigation ge’, “capture, Sf oepreciation on implenens & fara tuiding’, ‘Skeree o nad aplen, Yon eve (321°, ‘Mean squared error J (predicted value -actual value)? 2 ‘Mean absolute error =S_predicted value -actual value| 2 Root mean squared error = Yinean squared error ig, 94.4, Fol for etising the mode ‘The proposed work efficiently categorizes the weed type. This is due to ‘the model trained by the v2 plantings dataset. ‘Steps involved in Weed Identification module areas follows ‘Step 1: Importing Libraries and Dataset In order to utilize Deep Learning algorithms and preprocessing tools specific libraries needs to be imported. Using these libraries, the model building and prediction would be performed efficiently ‘The libraries such as NumPy, panes, rain test split, ResnetS2V2, ‘TensorFlow, keras were imported. Pig. 3.2.8 shows the glimpse of the ‘V2 Plant seedlings dataset. After importing the dataset, the fle names are been saved ina data frame along with its class name. Some of the mages present in each class are depicted using Fig, 3.2.2. There are rine weed classes present inthe dataset Bach cass has more than 200 mages. Sup 2: Spliting the Training. Testing set and Validation set ‘The dataset i split into 3 parts namely raining, testing and valida tion data. The Validation data is used while executing the epochs Training set contains 2701 images and Testing set contains 1447 mages. Validation set has 675 images, All the images were resized (0 224X224 sie Data Augmentation meaning increasing the number of images for training was performed, This inchuded in rotating, Hipping and rescal- ng the image. Additional images were obtained whichis then passed {nto the model for classification, Sup 3: Mode Bulding To classify the images Resnet152V2 pre-rained keras mode! is been sed, Deep Residual Networks are networks with convolutional, pooled, activating ad fllyconnected layers mousted on top ofeach other. The only structure that converts an accordance with the established into a recurrent neural network isthe identity link between the levels. To laddress the problem of the curse of dimensionality, this architecture facde a conception of the Residual Network. In this network, we use a techsique known a skip connections. The skip connection skips several phases of training ané connects output pin. The advantage of {incorporating this manner of residual connections i hat regularization will Bypass any layer that reduces architecture performance. A both 1 corollary, ineredibly deep learning models can be trained without the problems that vanishing/exploding gradients cause. Due to its advantages Resnet152V2 was opted for the proposed work to classify the images. Inthe proposed work Resnet152V2 was defined fllowed by Global Average Pooling2D, Dropouts and additional hidden layers Hidden layers was introduced above resnet152V2. First Hidden layer had 1024 ‘neurons, second and third iden layer with 512 neurons and fourth hidden layer with 32 neurons. All the hidden layers had activation funetion REL. Dropout were introduced in between each hidden layer 0 that it removes unnecessary neurons and tries to solect the best rameters to identify the images. The output layer had 1 neuron with ‘activation function SofMax for classification ‘The epochs executed for this module is 25 with batch size 32. The epochs are trained using the training data and are validated using validation data. Checkpoints are being defined and these checkpoints are stored for future weights updation. Step 4: Herbicide Recommendation ‘Not only predicting che image is important but also recommending Herbicides for the predicted weeds would be very useful forthe user. Herbicides are generally used fr killing the weeds. Manuslly removing the weeds are mostly impossible fr a large fam. Hence herbicides are ‘optimal way for destroying the weeds from the land. ‘The Herbicide details were collected by referring to various agricul tural website. To find a suitable herbicide the chemical component to ‘ul them sidentife from agricultural websites. Based on the chemical components the herbicide having it as an active component is being picked. The elaborate details regarding dosage, wellsulted crops, wel suited sll, active ingredients, iemtfying whether i sa presemergentSKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 4. Value of seed (both farm seed and purchased) 5. Value of manures (owned and purchased) and. Cast A, :Cost A, + rent pala for loasod-in land ig. 2455. Cato atom formal ‘or postemergent are obtained by a paf ile called ‘Product Label’. Tis product label provides the required details. These deals ae stored a5, 4 dataset whichis wsed for recommendation, For all crops same herbicide cannot be used. It could harm the arowth of the erops as well as affet the soll fectlity. Hence to ensure Accurate details crop specific herbicides are predicted using Random Forest Clasifr. ‘The input parameters for Random Forest Classifier ae crop name and the name ofthe weed. The crop name must be entered by the usr. ‘The name of the weed is identified by the Resnet152V2 fine tuning ‘model when the Image is been uploaded by the user, The Random Forest Clasiier classifies the crops based on its highest probability herbicides are listed 3.3. Module 3: Pests identification and pesticides recommendation Dataset: The dataset used for this module is pests dataset from Kagale (see Fg. 83.2). "This dataset was wed since [thas Images of various insects caprured at different locations. In many scenarios insects always merge with the crop and it is difficult to distinguish them. Hence this dataset is selected since it has insects in contact with crops. Any insect can be captured and uploaded into the system. The proposed work efficiently categorizes the insect, Tis is due to the model tained by the pest’ dataset Fig. 4. Loin page, Steps involved in Pest Identification and Pesticides Recommends: tion module areas follows Step 1: Importing Libraries and Dataset In order (0 utilize Deep Leeming algorithms and preprocessing tools specific libraries needs to be imported. Using these libraries, the model building and prediction would be performed efficiently, ‘The Ubravies such as NumPy, pandes, ransestsplit, ResnetS2V2, ‘TensocFlow, keras were imported. Vig. 33.3 shows the glimpse of the Pests dataset. After importing the dataset, the file names are been saved ina data frame along with its class name. Some of the images present in each class are depicted using Fig. 3.5.2. There are nine insect classes present in the dataset. Step 2: Spiting the Training, Testing set and Validation set ‘The dataset is split into 3 parts namely traning, testing and valiéa tion data. The Validation data is used while executing the epochs. ‘Training set contains 1766 images and Testing set contains 945, mages. Validation set has 441 images, All the images were resized to 224X224 sie Data Augmentation meaning increasing the number of images for ‘raining was performed, This fackided in rotting, Hipping and rescal- Jing Uhe image. Additional images were obtained whichis then passed into the model for classification, Step 3: Mode Bulding “To classify the images Resnet1S2V2 prerained keras model i been , pooled, and fully connected layers mounted on top ofeach othe. ‘The only structure that converts an accordance with the established Jno a reeuzent neural network isthe Identity link between the levels. ‘Toaddeess the problem of the curse of dimensionality, this architectureSKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 io ‘adés a conception of the Residual Network. In this network, we use 4 technique known as skip connections. The skip connection skips several phases of taining and connects outpst pin. The advantage of incorporating this manner of residval connections is that regularization will Bypass any layer that reduces architecture performance. Ae both 4 corollary, incredibly deep learning models can be trained without the problems that vanishing/exploding gradients cause. Due to is advantages Resnet152V2 was opted for the proposed work to classify the images, In the proposed work Resnet152V2 was defined followed by Global Average Pooling2D, Dropouts and additional hidden layers. 4 Hidéen layers was introduced above resnet152V2. Fest Hidden layer had 1024 ‘neurons, second and third hidden layer with 512 neurons and fourth hidden layer with 32 neurons, All the hidden layers had activation function REL. Dropout were introduced in between each hidden layer s0 that it removes unnecessary neurons and tries to select the best parameters to identify the images. The output layer had 1 neuron with activation function SoftMax for elasitication. "The epochs executed for this module is 20 with atch size 32. The epochs are trained using the training data and are validated using validation data. Checkpoints are being defined and these checkpoints are stored for furre weights updation. The accuracy obtained was 0.98, ie, 98%, ‘Sup 4: Pesticide Recommendation Not only predicting the image Is Important but also recommending pesticides forthe predicted insect would be very useful for the user. Pesticides ate sprayed for killing the insects so that it does not destroy ‘the crops and inerease in number in the feds ‘The Pesticide detals were collected by referring to various agricul tural websites. To find a suitable herbicide the chemieal component to Kil them Is identified from agricultural websies. Based on the chemical components the pesticides having it has an aetive component is being picked, Other details regarding well-suited crops are obtained by a pat file called “Product Label’. This product label provides the required details. These detalls are stoced a8 dataset whieh is used for recommendation. Same as herbicides for all erops same pesticides cannot be used. It could harm the erop growth as well s affect the soil fertility. To ensure correct pesticides crop names along with predicted insect are passed Random Foret Classifier, The crop name must be entered by the user. ‘The name of the weed is identified by the Resneti52V2 fine tuning ‘model when the image is been uploaded by the user. The Random Forest Classifier classifes the crops based on its highest probability estes ace listed 5 Basahacessitr oasis S2Sesiss 258 { GrasentocsigCossir gb 05727273 oOABHSAS on AH545 . anos xg S05 BITENT at7zTZI 0 gncasater gm 1000"00 SOUSKSAS aa AHSHS ig. AL Suny det 34, Cost estimation ‘The main aim of cost estimation module is to predict the cost concepts and total cost needed for cultivation, This module is very {important in recent times due to sudden criss, natural calamities, price rise of cultivation resources, Indian cost of eallvation survey data is ‘considered for this module. Fight years data was collected. Som 2010— 2018, The data was aggregated since each year was a separate fle 8s shown in Fig. 3.4.2. In order to forecast values Machine Learning, Regression models are been considered (See Fig. 34.1) Discussing about the dataset the attribute are depicted in Fig. 2.4.3 ‘The dataset can be viewed as two parts: Firstly, is the item wise break up attributes and cost of cultivation attributes. Kem-wise breaks includes all the operational cost attributes and fixed cost aiteibutes. Cost of cultivation attrinates includes Cultivation Al, A2, B1, B2, C1, C2 and C3, To forecast the cost details itis important to rin each tem wis atrbute separately. The year, state and crop attribute are used 35 the independent variables and each item-wise atebute separately or example, let us consider the Operational cost of Fertilizer & ‘manure attribute, For predicting this atibute following steps are fa loved Step 1: Importing Wbraries and dataset ‘The necessary Ubraries such as NumPy, pandas, tain test split, Label Encoder, Machine Learning Ensemble Regressors are being im: ported. The 8 separate datasets corresponding to years berween 2010— 2018 are been loaded Step 2: Data Preprocessing and Model Buling The imported datasets are aggregated to obtain « single dataset. ‘The crop and state column ace non-numerieal in nature. Hence LabelSKS Dial and MD. Stan econ Ayes Sora 3 (2022) 100081 2 oscves ost .ny msn 4 CSOVNB.1 gute 72273 6 SVE ores sen 8 escvose omtenae ws2rza 8 csovxc8 pane emt Encoding is performed to obtain numerical values forthe same. Year, ‘encoded crop name, encoded state name is been feed into the Machine Learning and Ensemble regressor models. Step 3: Model Building ‘With regard to Regressor models XGBoost repressor is considered {or explanation. Year, encoded crop name, encoded state name is been considered as the independent variable and Operational cost of Fertil ier & manure is considered asa dependent value. The dataset is split, into test and training data, The regressor model is been defined and. the training datas fitted into the model to tain the model. The model is then tested sing the testing score. The r2 sore, root mean squared error, mean squared etror, mean absolute error is been estimated for each model a8 shown in Fig, 04.4. Swap 4: Cost Caleulation Based on these values the best model is selected with the highest s2score, The same procedure is been practiced forall the itemewise attributes and the best models are stored in a pickle file. The pickle file is then used to forecast the values tll 2028. Once all the attribute forecasting values are obtained then these values are stored dynami= cally ina data frame. This data frame is converted into a csv file and, is used for future calculation, ‘When discussing aboot cost calculation the formulas are used to ‘obtain the values of Operational cos, fixed cost, total cost, cos con cepts values AL, A2, Bl, 82, Cl, C2 and C3. The forecasted values till, 2028 are been applied to get the results. The formalas as depicted in Hig. 3458 and Fig. 3.45. 4. Results & discussion ‘The proposed work is a web interface through which the user can access the models efficiently. Pig. 4.1 represents the login page Crop Predictor ig AA, Frnt en omer the des are tempered nl. through which the user needs to login with their credentials to access the models. Fig. 42 shows the user dashboard after logging in, thus enabling users to access the models. The results of each model are discussed in each section. 4.1. Grop recommendation ‘Ten Algorithms were used for crop recommendation. The accuracy score above 90% was selected a8 shown in Fig. 41.1. The accuracy drifts are been clearly observed using Fig. 41.2. For these selected algorithms Hyperparameter Tuning was applied from which best model with highest accuracy was obtained as shown in Vig. 4.1.2 According. to Fig. 4.1.3 the Random Forest clasifier hyper tuned with Randomized CV is opted as best mode since its accuracy is 95.4536 and stored asa pickle file for further analysis. Frontend: ‘The following figure shows Crop recommendation system. The user needs to provide details about the average rainfall and ph. The other parameters are derived. Initially longitude and latitude values are extracted from the current location using Web Scraping and is passed fa parameters to the Weather APL. The weather API provides the temperature, humidity details of the current location. Based on che parameters passed, the system predicts the 5 crops and provides deta, According to Fig. 4.1.4 the user has entered the values 500 for average rainfall and 7 as the soil ph, ‘The page after clicking predict directs to the results page which contain the details provided section depicted by Vg 41.6. This shows the longitude and latitude values of the current location. I also shows the place, region. Followed by the basic details the temperature, ridity obeained fom passing longitude and latitude values are shown. Average rainfall and ph. values are entered by the user (se Fig. 4.1.5).yes tral 3 (2022) 100081 Crop betas sotnpe Tena em tay nae oe ean toy ‘Short voor toon ° nese Seer o RACE. ae Cop Rai cop, Ym [RAN Cop: YO RAB Ca eM ree oxi ° 1 esi ost ooh parent, sesesasecoomor sSstomouosnoose sossuateeoe? SertatTosmooo! _s.oeeoomaousee roaporous Fg 1.6, Top 5 cop recommends Based on the given duals the model takes temperature, humidity Table 42.1 and rainfall and ph, details to recommend crops. The system provides Hesston of ops ‘Top 5 Crops based on the details provided cep ope Tapes any aa The above Fig. 1.7 shows the Top 5 exop recommendations fol “War Gem 17 Teo) 100 mae lowed by its scientific name. Once the model has predicted the crop Senmer Gop 25:37 ‘eas name is passed on to identifying its scientific nme. The system contains 4 dataset containing scientifc names. In the same way the soil type 4s also obtained from soil.csv dataset for that specific exop. The Crop type specifies whether the given crop is Rabi, Kherif or Summer. The crop is been categorized based on its parameters such as Temperature, ‘humidity and rainfall (eee Table 4.1.0 ‘After identifying the crop type the system estimates the Growing De- ‘gree Days, Nitrogen requirement, potassium requirement, phosphorous section whose formulas are discussed in Section 2. eis observed that ‘the GDD of apple is more and black gram being the least. With respect to nitrogen requirement banana requires the more nitrogen content 1 be present in 200 Ib. fertilizer and pple and pigeon peas requlees ‘the same quantity. With respect to phosphorous and potassium apples requirement content i highest a¢ shows in Fi. 4.17. The predictions as well asthe details are more important fora user to analyze and select a crop. The existing research restricts the crop recommendation of only one. And does not provide any addtional values. The proposed system predicts top 5 crops with ts details whieh can be used by the user for camrying out their work in a smarcer The system just does not take up the inputs by the user but also checks if cs a valid input. For example the soil Ph range is between (0-14. Ifthe user types a value greater than 14 or less than O the system displays a message that itis wrong and asks the user to re-enter correctSKS Dial and MD. Stan ‘Cuet Tempranse a Huaiiy is comsided me Avenge Rill foo ie we Fg, 417. nce ete Fig. 41.8, Reqeting he se ene otc values 2665 ‘conoresnenyoreN2 pant scoangsLees. Lees Sty en 3698 leobsrveniyOrneN2 pnt sosegeScan. Scans jon 2792 eoneve Orie? lant sodigsScen. Scans Maywoed 2881 ‘conoretvenOrnane pant seeingsLees. Loos iy ont 205 ‘cctrusivenyOrweN? pnt seodngsLecs. Lene Sty ent 2163 ‘eclersivenyOrneN? son seoings Lens. Lonse Si tent 505 ceremsenyomen2 slr sedges Comer 3900 lerensenyOnen2 plan seapsSep.. ShepnaAlls Pause 235 ‘comortsivennOrwer? at seeding Bie cross {23 10s «2 cos Fig. 42.1 Suring Fle puth and ns nels in te fae values. Fig 4.1.7 displays the values entered by a user such as rainfall 8 500 and soil Ph as 19, ‘The soil ph. value cannot be 19 as it does not lie in the range between O-14, Hence the system displays the message stating to enter correct values as shown in Hg. 4.1.8 42, Weed identfeation ‘This module helps in identifying the weed present in farm and also suggest herbicides forthe predicted weed. To predict the Weeds [RESNET152V2 pre-rained algorithm veas used, It resulted in an accu racy of 0.89. econ Ayes Sora 3 (2022) 100081 ‘aining vs validation accuracy Fg, 42.2, Accray of walang and vada dt “ining vs validation Loss ig, A2.3, Taning and valiant ig, 2.1 represents the data frame consisting of file path and clas labels, This data frame is shufled ane then split into training and test Sets for model building and prediction. After fine runing the base model Resnet152V2 and adding additional layers, epochs are executed for fitting the taining data end validating against the validation data.25 Epochs are obtained with a training accuracy of €2.01% and validation accuracy with 82.52%, Sig. 4.2.2 shows how the training and validation accuracy fluctuated while trating through each epoch, As the epochs Increase the accuracy also ineveases. Fg, 4.2.3 the loss deals aze de creasing asthe epochs increases. This show tht the model s learning progressively a the epochs increases. ‘Ater predicting the weed the next step is to predet the herbi cides. In order to achieve this data is collected and stored into a data frame. Fig. 4.2.4 shows the sample data collected for herbicide recommendation, ‘The deals obained as shown in Fig. 4.2.4 are Weed name, concen: tration, Herbicide, Crop name, pre-emergent or post emergent, dose, soil type, Group, Weed growth stage in tems of leaves and Weeks. Concentration specitis the active ingredients present in the herbicide, followed by the herbicide name. Herbicides can be broadly clasiied Into two types preemergent and. post emergent herbicides’ emergent herbicides corresponds to application of herbicide in the Neld before cultivation, Weeks's column is related to pre-emergent which specifies the number of weeks prior to cultivation the herbicides need to be applied. Post emergent herbicide specifies the aplication of herbicides fonce after the weed is been spotted on the field. Weeds needs 10 be Jalled during its initial stage. Hence Weed stage.no of leaves specify the time to apply the herbicide based on the number of leaves grown, ‘The weed names and the erop names are non-numerical in nature hence it needs to be converte into numerical format for the model to use it. These two attributes are been Label encoded. The correspond: Jing. numerical values for these ateTbutes chosen as the independent variables. The encode values details glimpse are shown in Pig. 4.25.1 bickgos —Farmt Adee Wet Po. ener 4 baci pence ltt opm) What Poona A osle ° 2 Barey ‘ 7 Bisby 3 ceovoage ‘ 5 carrot esson Ans Sora 92022) 100087 et pe crap ndstge nef tenes a sean 2 a 0 Weed Identification ose Fle 229.99 Herbicide atribute is aso label encoded to obtain numerical values land are used as dependent variable, The dataset after Label encoding fare split into two parts training and testing data. The training data is ‘been fed into the random forest elassifier model for prediction. Once the ‘model leans the data the model is fed with testing data, The model i ‘been stored into a pickle which is used for further analysis. Once the ‘ser specifies the erop name that he wants to cultivate and uploads the {mage ofthe weed found on the field the system predicts the Herbicide and displays the relevant deals of the Herbicide Frontend For the user to Identify the weed and get deals of herbiedes the user needs to specify the crop name and upload the image captured from the field as shown in Fig. 4.2.6. Fig. 4.27 shows a black grass image. This image is uploaded into te system and erop—Whea is been selected. "ig. 42.8 shows the description about the weed. The model has precisely predicted that iis a black grass. It shows the growing seasonSKS Dial and MD. Stan Black gras it ative anal re weed hat cece mas ithe eel provi ant "Stat ata we pes chy eid hen land, cco ony Description cccasonllyon sanyo reso Bc rst lls om got (Clerp purr) and ts can es otananaon of pai af Growing oi emi Septenter Oe Herbicide (Alter Flee) [fein 150 0D iodaufcn mei Concentration tum Smeonlion med 9 meeps “iets 1 PostPre emergence [Pos emergent] Irbies Dos 100} [LoamysanSany- Loan "Cay Loa Sted Soitype ‘Sansa ay out Loan Back ai Mout sa) Groep aris Wad enves 6 wos econ Ayes Sora 3 (2022) 100081 {Ciera} [Brcaday Sar] [ypronstam 68.3 [pizsinin penned? 4. WG Dorlas 228 Damecotop Pfeil] clorguatece- al ‘ol [Pos emergent] (Pow emer) 00) 00} (Sandy Lo Lonny’ Black si [leey Loans “Team ei! Cyey Loa) ‘andy Low) os] wm & Bo o Fie. 428 Prelced de ‘ofthe weed is from September-October. The Herbicides ae listed with is an obtained by-passing crop name and weed name into Random Forest Classifier Model, Furthermore, the system categorizes it as pre= post emergent herbicide followed by is dosage ml/ha and suited soil, types are also provided ‘The Group represents the chemical solution that the herbicide be- Jongs to. The weed leaves are the leaves stages during which the use is directed t spray the herbicides ic ia post-emergent. I Presemergent ‘the system specifies the number of weeks before the herbicide needs to be sprayed, 4.3, Pesticides identification ‘This module helps in identifying the insecss and pests present in farm and also suggest pestiides forthe predict insect, To predict the ‘Weeds RESNETIS2V2 pre-trained algorithm was use. It resulted in an accuracy of 0.98, ‘ig, 43.1 represents the data frame consisting of file path and class labels. This daa fame is shuffled and then split into training and rest sts for model building and prediction. After fine tuning the base model ResnetI52V2 and adding additonal layer, epochs are executed for fitting the training data and validating against che validation data.20 Epocls are obtained with training accuracy of 89.01% and validation accuracy With 98.2% Fg. 4.3.2 shows how the training and validation accuracy Muetuated while icerating through each epoch. As the epochs Increase the accuracy also increases Fig 43.5 the lose details are de= ‘renting as the epochs increases. This shows that the model is learning progressively es the epochs increases. ‘After predicting the insect, the next step is to predict the per ticles. In order to achleve this data is collected and stored ato a 2001 coniedavenyOrvetpestaingrsshoppe!. grasshopper 943 conterudeventyOmelpestranmtesipg ot tes 1611 eanenwamentyowerpestvanbostneg 6. deete 403 eonereveMyOrvslpestiestphidsipg 15. apis 1901 lcotetamenyOrwepestvanemyvorypg ——smwoim 2103 cortencvelyDivelpestrantotwarypg.._botworm 905 ‘contetdiveNyDrvejpesttainintesing 40. rites 1096 icortetmenyOrvepestvainamyworypg_smorm 235 icntemvanyDrvepestestrasshoppc).easshoprer {1081 lcortortmanyOrwepestaivamyworypg._amworm 3150 rs «2 corns Fig. 42.1. Sig ie uth wc bel in dts ae data frame. Fig. 43.4 shows the sample data collected for pesticide recommendation, The details obtained as shown in Fig. 4.3.4 are Pest name, Crop name and pesticide details. The pest names and the crop names are ‘non-numerieal in nature hence it needs tobe converted into numerical format forthe model to use it, These two atebutes are been Label en coded. The corresponding numerical values for these attributes chosen fas the independent variables, The encode values details glimpse are shown jn Pig. 43.5, Pesticide attribute i also label encoded to obtainSKS Dial and MD. Stan Taining vs validation accuracy ‘ening ws vadaton toss (© hiss oan Acapate 900 1 aphss Sprouts. Acoprat 50 WOS 2 aghss Culver Acsphae 90 WOS 3 aphids Ltn copra 80 90S 44 esns PeppersAespate 99\WOS ‘numerical values and are used as dependent variable. The dataset after Label encoding are split inco two parts taining and testing data. The ‘waining data is been fed into the random forest classifier model for prediction, Once the model Teams the data the model is fed with testing data. The model is been stored into a pickle which is used for further analysis. Once the user specifies the crop name that he wants to cultivate and upload the image ofthe insect found on the field the system predicts the Pesticide. Frontend: For the user to identity the pest and get details of pesticides the user ‘needs to specify the crop name and upload the image captured from the field as shown in Fig. 43.6. Fig. 4.3.7 shows a stem borer image. Tis {mage is uploaded into the system and crop—Rice is been selected Tig. 43.8 shows the pesticides that can be used for killing stem borer, we can identity the image of che pest that the pest is same color ‘of the stem. But the model proposed has precisely predicted that iis a stem borer pest econ Ayes Sora 3 (2022) 100081 44, Cost of cultivation For this work Cost of cultivation data from 2010-2018 Is con. sidered. 11 Crops are been selected and are considered for analysis ‘They include Arhar, Bajra, Cotton, Groundnut, Jowar, Ragi, Paddy, Sugarcane, Maize, Potato and Wheat. All these attributes were fit into Regression models especially ensemble regression models such a5, XGBoost, GradientBoostingRegressor, Bagging Regressor and Decision ‘Tree Regressr the models were predicted against each crop and state forthe years starting from 2010-2028. Considering in identifying a best model for Operational cost Pert ‘ner the following results are obtained. ‘The following algorithms listed in Fig. 4.4.1 were implemented to find the best mode! based on the r2 score. The algorithm with highest score is been opted as the model for fertilizer cost forecas’s same algorithm is not implemented for other attributes. For every attribute 9 Regressor algorithms are executed and the model with highest r2 score fs selected for them. For al the attributes mostly XGHoost Regressor, [Random Forest Regressor and Bagging Regressor performs best hence they are saved as a pickle file and then used fo forecasting values tI year 2028. ‘These results were aggregated to obtain Operational Cost, Total Cast and Fixed Cost. n alton, these values were used for eleulating Cost AAI, A2,B1, 82, C1, C2 and C3, 1m order to estimate the cost of cultivation the following columns can be split into two pars Operational Cost and Fixed Cost. Operational (Cost consists of all the expenses such as purchase of see, insecticides, animal labor, human labor, machine labor, eat on fertilizer and ma: ‘ure. And Fixed cost consists of fe rental value of owned land, FC ent Paid for Leased in Land, FC Land revenue tax ceases, Operational (Cost of Interest On working capital, Depreciation on implements Feem, Building interest on Fixed Capital (Cost Concepts splits the various expenses into 7 groups such as Cost Al, AZ, B1, B2, C1, C2 and C3. Based on these formulas all the details are summed together to get the operational and fixed cost value. Later ‘operational cost and fixed cost value are summed that results in total ‘The user needs to specify the crop name ane the sleet state of India to get the results. The 11 crops and their corresponding states ‘thatthe erop is grown i been listed in the drop down list as shown in Mig. 44.2 Fig. 4.4.8 shows a glimpse of the forecast values tll the year 2028, for all the attributes, These atsibutes are calculated based on the formulas in Fig. 3.4.5a. "ig. 44.4 depicts the operational cos variation over the yeats 2010 2028. It is been observed at specific intervals there Is sudden drift in ‘the operational cost. It is been also observed at the cost are almost the cost reaches to 45000, Hach year data consists of details such as Operational Cost of Attached, Casual, hired human Labor, hired fand owned animal Labor, hired and owned Machine Labor, cost on Fertilizer, Manure, Insectises, Irrigation, crop Insurance, Payment 10 Contractor and Miscellaneous Cot, These data are summed wp to obtain the operational cos. In Fig. 4.4.5, with respect to fixed costs we have rental value of owned land, Rent Paid for Leased in Land, Land revenue tax and ceases, Depreciation on Implements Farm Building and Interest on ixed Capita. All these ateibutes ae forecasted and these values are summed up to obtain Fixed cost, The fixed cost shows an increase over the years tl 2015. Followed by adrift in amount and forecasting states that i would remain the same at 13000 Fig. 44.6 depicts the Total cost for cultivation, Total cost isthe summation of Operational cost and Fixed cost. Since the operational cost and fixed cost are same after years. The total cost specifies the cost would be 65000. ig, 4447 depets the cost concepts cost Al. Concepts are mainly ‘sed by business units since it splits the expenses into each parameter.SKS Dial and MD. Stan — ts ee baad Sem aroccou a om a on Deco Ayes Sora 3 (2022) 100081 ‘expat s0 WO ‘ Pest Identification Set he Cop Chcoee Fle ised png Fig, 43.6. Pare upoad the iene Fig, 4.37. Upon ge Stem te, ‘This can be wsed for best decision making. Cost AL has sudden peaks ‘over the years resulting in the value of 38000 over the year. Fig. 44.8 depicts the cost concepts cost A2. This cost value focuses ‘om all che expenses along with the fixed cost on leased Tan. Fig. 44.9 depets the cost concepts cost BI. This cost value focuses fon all the expenses along with interes on fixed capital Fig. 4.4.10 depiets the eost concepts cost B2, This cost value focuses cost BI along with the rental value on owned land an rent on leased, lan. Fig. 44.11 dopits the cost concepts cos Cl. This cos value focuses cost BL along with the imputed value of family labor. This value is ‘minimum value of family labor. ig, 44.12 depts the cost concepts cost C2. This cst value focuses cost 82 along with the imputed value of family labor. Tis value is ‘minimum value of family labor ig. 4.419 depiets the cost concepts cost C3, This cost vale focuses cost C2 along 10 percent cost of C2 as management cost. Based on these details farmers and business exeeutives would be able to make things ready before cultivation 5. Conclusion arming Is a Back bone of every country. Hence it needs t0 be ‘monitored in a timely manner. The modules in the work provides a helping hand to farmers for identifying the crops that can be grown bbesed on their place. Identifying the weeds and recommending, her bicides is an important element. All the crope are prone to insets. Hence identifying the correc inset and recommending the pesticides for the same would be an efficient fool. Many farmers are unable to estimate the cost of cultivation, Due to some uncertainties, there might be loss for the farmers. It also provides estimation of cost of cultivation for each operation such as human labor, animal labor, cost of seeds, manures and fertilizers, also forecasts the fixed cots. This provides an ‘overview of how to plan the activities and do cukvation in a profitable Declaration of competin ‘The authors declare that they have no Knows competing finan: cial interests or personal relationships that could have appeared to Influence the work reported inthis paper.Pest Identification and Pesticides Recommendation Predicted Pest : stem_borer Pesticides Postcdes [PYRETHRIN7EC] _[Dimsthoate 400 EC] ig. 41. Matt pecoemance deat Cost of Cultivation ‘Appendix. Supplementary material Module 2: Weed Identieation nups://wwwhagale.com/vbookshelf/v2-plant-seedings-dataset ‘Module 1: Crop Recommendation Module 3: Pest Kent ‘stps://wew- kaggle.com/atharvaingle/crop-recommendation-datas os ps: kagale-com/simranvolunesia/pest-dataset ps: kagale-com/shekharyada/erop-soicsv Module 4: Gost Estimation ps: agale-com/j02197erop-names Inps:/eands dacnet.ai./CostofCulivation.himReferences 10 an 281 aSKS Dial and MD. Stan [a0] theamosh Kasnthn,Deshayon Sigs Sena Rey Up, net Tec, Bae, 202 tan) Zakcn Sed, Can We Etna owt Hsniiaon Bake Deet fom eee Deco Ayes Sora 3 (2022) 100081 (2 Sues Meena, LP Shah, Ram Lat Meena, Cant of Calivation and eure on Dirt Cast Coneps lass of Onion it Hajion ND (25) Ani Mandl, Vayiog Mota and Dteminans of eam Cop sng Cost alien Das A Hee te Aproeh Spige, 202