Data-Driven Soil Analysis and Evaluation for Smart Farming Using Machine Learning Approaches

Abstract

Food shortage issues affect more and more of the population globally as a consequence of the climate crisis, wars, and the COVID-19 pandemic. Increasing crop output has become one of the urgent priorities for many countries. To raise the productivity of the crop product, it is necessary to monitor and evaluate farmland soil quality by analyzing the physical and chemical properties of soil since the soil is the base to provide nutrition to the crop. As a result, soil analysis contributes greatly to maintaining the sustainability of soil in producing crops regularly. Recently, some agriculture researchers have started using machine learning approaches to conduct soil analysis, targeting the different soil analysis needs separately. The optimal method is to consider all those features (climate, soil chemicals, nutrition, and geolocations) based on the growing crops and production cycle for soil analysis. The contribution of this project is to combine soil analysis, including crop identification, irrigation recommendations, and fertilizer analysis, with data-driven machine learning models and to create an interactive user-friendly system (Soil Analysis System) by using real-time satellite data and remote sensor data. The system provides a more sustainable and efficient way to help farmers harvest with better usages of land, water, and fertilizer. According to our analysis results, this combined approach is promising and efficient for smart farming.

1. Introduction

Soil is a common material seen on Earth, and a natural material composed of solids (e.g., minerals and organic matter), liquids, and gases according to the definition by the Natural Resources Conservation Service (USDA, n.d.). The soil’s contribution in agriculture leads to the fact that the soil is highly coupled with human daily life. According to the Food and Agriculture Organization of the United Nations in 2015, approximately 95 percent of food is directly or indirectly produced from the soil, which raises the importance of soil since food is the essential energy and nutrition source for humans. The efficiency of food productivity draws more attention globally due to food supply and insecurity issues; smart agriculture helps us maintain soil health with more accurate nutrition to the crop. As a result, soil analysis via deploying smart agriculture techniques is crucial to maintaining the sustainability of soil in producing food and crops regularly.

Smart agriculture is an emerging field that introduces new technologies such as big data, IoT, satellites, and drones to help farmers optimize farming results. The benefits of smart agriculture include a reduction in manual labor, an increase in productivity, and a decrease in costs. While lack of unification is still one challenge for agriculture researchers and farmers, most agriculture research and applications focus on one specific area such as physical features, chemical features, or biological properties. Compared to previous research, (1) our research work proposes the vegetation indices calculated from satellite data, as well as various soil and fertilizer characteristics from public data resources. (2) Our work combines crop identification, irrigation system analysis, and fertilizer prediction into a whole analytical system. Outputs from crop identification modules are used as inputs for the irrigation system module, and then all the parameters flow into the fertilizer prediction module, making planting decisions and forecasts in one stage possible. (3) Our project provides a user-interactive system in soil analysis for people involved in agriculture fields and is user-friendly without requiring prior agriculture knowledge. The system is a web-based application. The backend is composed of machine learning models to perform the following tasks: crop analysis, irrigation recommendation, and fertilizer (nitrogen) recommendation. The frontend is a web interface for users to input their geolocation (e.g., GPS) of the area of interest and the analysis result will be outputted.

The organization of this paper is structured as follows. Section 2 is the review of the related work. Section 3 presents data engineering by introducing the raw data sources, the processing of data, the calculation of certain parameters, and the selection of parameters. Section 4 discusses machine learning model development, including the Random Forest, SVM, AdaBoost, Linear regression model, and the multilayer perceptron deep learning model. Section 5 shows the model evaluation where the best performance of the crop identification model is 78% in MLP; the R-squared score of the linear regression model for the irrigation system is 0.939; and the best accuracy score for the fertilizer recommendation model is 93.3% in MLP. Section 6 shows the system building and uses a case analysis. Section 7 displays the description of our visualization platform, an interactive map with the UI designed via Voila. Section 8 discusses and compares our system with other systems with similar functions. Section 9 concludes with three main achievements from the project.

2. Related Work

Table 1. The literature survey for crop identification.

ID	Year	Focused Problem	Approaches	Accuracy	Datasets
[1]	2009	Crop Identification	Statistical analysis about temporal series of indices	91%	MODIS data
[2]	2019	Crop Identification	Support vector machine	91.8%	NDVI feature
[3]	2021	Crop Identification	Decision Tree, KNN, Random Forest, SVM	79%(DT), 88%(RF), 88%(SVM)	Vegetation indices from satellite images
[4]	2022	Crop Identification	Gaussian Bayesian models, Neural Network	83.8, 80.7	Pixel spectra of crops from NASA Hyperion satellite
[5]	2022	Crop Identification	Efficientnet-B0 network architecture in Darknet	99%	Crop images

shows the technical evolvement of crop identification. In the early stage (1969 to 1990) of crop identification, temporal-spectral data was used to calculate the vegetation index and to analyze the light of crop via building the crop growth and yield model. An alternative approach is to identify the lighter or darker tones of an image based on the field boundary and ground data. From 1991 to 2000, statistical analysis of the polarimetric multifrequency was getting popular, SAR data (the polarimetric C-, L-, and P-band from the AIRSAR system, as well as the X-band from the E-SAR system), sometimes combined with the pixel distributions of each agricultural plot, was used to calculate the crop’s wavelengths. Starting from 2000, vegetation indices such as NDVI, EVI, MSAVI, and NDWI were analyzed as spectral features in statistical approaches [1]. In the meantime, feature selection models were developed to improve the accuracy of the approach. From 2011 to now, various machine learning models, such as Random Forest and Support Vector Machine, have been deployed in crop identification by evaluating the spatiotemporal multispectral bands from satellites [2,3]. Moreover, deep learning models such as convolutional neural networks start to process the laudatory images with image feature capture [4,5].

In the consideration of the irrigation prediction system from

Table 2. The literature survey for irrigation prediction.

ID	Year	Focused Problem	Approaches	Accuracy	Datasets
[6]	2019	Irrigation Prediction	Decision Tree, Particle Swarm Optimization	95% (r-squared)	Moisture sensor data
[7]	2020	Irrigation Prediction	Soil moisture weight function, NPK derived from pH collection	Observations with no accuracy result	Moisture sensor data NPK sensor data
[8]	2021	Irrigation Prediction	KNN for classification, Blaney–Criddle method for water required	Observations with no accuracy result	Moisture sensor data
[9]	2020	Irrigation Prediction	CNN	0.0184(RMSE), 0.093(MARE), 0.98(R2)	thermal images
[10]	2018	Fertilizer Prediction	RF and Simple Linear Regression (SLR)	33.2%(RMSE)	hyperspectral camera data and RGB camera data
[11]	2019	Fertilizer Prediction	convolutional neural network(extraction of features)	83%	RGB crop field images
[12]	2021	Fertilizer Prediction	Random Forest	86%	RGB crop images

, Mohammad Reza et al. [6] used a combination of decision tree algorithm and particle swarm optimization (PSO) for times series prediction for wastewater with an r-squared score of 95%. Istiak et al. [7] built an integrated irrigation network, by measuring the volumetric percentage of soil sample water content. The raw soil moisture data is collected by a moisture sensor. Shilpa [8] first classified the soil type under the KNN approach with humidity, temperature, and soil moisture as the parameters. Then, the authors calculated the water needed for the crop by using The Blaney−Criddle formula. Remilekun Sobayo et al. [9] created a CNN-based soil measurement by combining thermal images with the measurements of the farm area. The moisture level is valued by the soil temperature represented in it.

Niko et al. [10] estimated the crop biomass and nitrogen content in the soil by extracting features from hyperspectral and RGB cameras. The methodology is based on Random Forest and simple linear regression. H. J. Escalante et al. [11] deployed a deep convolutional neural network to extract features from RGB images and then feed those features into predictive models. Diego et al. [12] also obtained the vegetation indices from remotely piloted aircraft images, which are applied to Random Forest machine learning methods to calculate the nitrogen content in coffee leaves. The global accuracy and the kappa coefficient are up to 0.91 and 0.86, respectively.

Fertilizer management is essential for land-use efficiency in

Table 3. Literature Survey for Fertilizer Prediction.

ID	Year	Focused Problem	Approaches	Accuracy	Datasets
[15]	2022	Fertilizer Optimization	Optimization and simulation	28–53% improvement	Nitrogen contents
[13]	2018	Fertilizer Prediction	Random Forest	86%	RGB crop images
[14]	2020	Fertilizer Prediction	convolutional neural network (extraction of features)	83%	RGB crop field images
[13]	2018	Fertilizer Prediction	RF and Simple Linear Regression (SLR)	33.2%(RMSE)	hyperspectral camera data and RGB camera data

. In recent years, machine learning and deep learning approaches have been conducted for most fertilizer prediction projects. Agarwal et al. [13] and Caturegli et al. [14] trained the models by analyzing crop images. Abhaya et al. [15] optimize the quantity of nitrogen fertilization considering the cost of labor and resources.

3. Data Engineering

3.1. Data Collection

The datasets related to the project are mainly focused on two aspects: one is the on-the-ground task of crop identification and irrigation cycle system based on soil moisture analysis; another is the underground task of fertilizer recommendation based on soil nutrition analysis. Each objective needs different raw data and analytics tools. In

Table 4. Main tasks to develop models.

Task	Content	Raw Data	Resource
Crop Identification	Identify the types of crops	Geolocation (longitude and latitude) Date of the flights Multispectral bands Land use map data	GPS device Satellite data/Drone data Satellite bands/Hyperspectral sensor Data collected by the California Natural Resources Agency [16]
Irrigation Cycle	Recommend irrigation system	Precipitation and Evaporation Underground water Runoff	ERA5-Land/moisture sensor and evaporation sensor ERA5-Land Calculated data
Fertilizer recommendation	Recommend the usage of fertilizer (NPK)	Soil chemical properties Nitrogen Content [17]	LIBS sensor LIBS sensor

, for the crop identification section, we built the models by utilizing satellite spectral data and land use maps; for the irrigation cycle system, the crop identification results are used as part of the inputs, combined with climate and soil data; for the fertilizer recommendation section, the crop types and soil data are used with nitrogen content data in the soil example.

Currently, most of the soil and surrounding data come from public satellite data. The advantage of such data is easy to collect at a low cost, but the limitation is we cannot customize the data collection frequency. In the future, we plan to collect more specific agricultural data based on different farmlands, which need hyperspectral sensors, drone equipment, climate-based controllers, and LIBS sensors. In such a way, a personal workstation can connect to the soil analysis platform via the cloud.

3.1.1. Crop Identification: Land Use Map

The land use map from DWR gives the public access to statewide land and crop datasets for the last 30 years. It is a legislature-requested investigation into the aspects of agricultural land use and water needs. Most of the survey data are recorded into geographic information system (GIS) software by surveyors, who visited more than 95% of agricultural areas in the state and recorded the land usage [18].

We selected the dataset for San Joaquin County. Overall, there are 74,382 polygons delineated in the region, within which 34 objects are categorized. For example, acres, area_meter, class1, class2, class3, croptype1, croptype2, croptype3, irr_type1 [19]… We picked ‘Class1’ as the classification for land use, which is the first level of land use identification. The codes for the classification are all capital letters in

Table 5. Description of crop class on first level.

Category Code	Description	Details
R	Rice	Agriculture land use
P	Pasture	Agriculture land use
G	Grain and hay	Agriculture land use
T	Truck, nursery, and berry crops	Agriculture land use
F	Field crops	Agriculture land use
C	Citrus and Subtropical	Agriculture land use
D	Deciduous fruits and nuts	Agriculture land use
V	Vineyard	Agriculture land use
S	Semi-agricultural	Other land use
I	Idle	Other land use

.

3.1.2. Irrigation Cycle: Precipitation

The definition of precipitation is the accumulated water amount that falls to the land surface. The amount of precipitation does not include the fog, dew, or other precipitation that has not fallen to the land ground before evaporating. The unit of the precipitation measurement is the meter and is assumed to be spread evenly to the land. In our crop water supply system, we use the precipitation parameter as the ‘crop water supply’ [20].

3.1.3. Irrigation Cycle: Evaporation

The unit of the evaporation measurement is the meter and is assumed to be spread evenly to the land [21]. We extracted evaporation bands from ERA5-Land, which is a reanalysis dataset by combining the model data with the observation data [22]. In

Table 6. Description of evaporation bands from ERA5-Lam.

Description of Bands	Term of Use	Citation
evaporation_from_bare_soil	The amount of evaporation from land surface soil.	m of water equivalent
evaporation_from_the_top_of_canopy	The amount of evaporation from the canopy interception reservoir.	M of water equivalent
evaporation_from_vegetation_transpiration	Amount of evaporation from root extraction, i.e., the amount of water extracted from the different soil layers.	m of water equivalent

, the evaporation parameter we used is the total evaporation band, which is the sum of evaporation from bare soil, open water surfaces, the canopy, vegetation transpiration, and snow evaporation.

3.1.4. Irrigation Cycle: Underground Water

Underground water is the water held by soil or rock under the land surface. Water is pulled through soil layers because of gravity, first reaching the upper layer zone of saturation called the ‘water table’, then going down to the deep layer called the ‘aquifer’ beneath the water table. Water in the aquifer moves slowly and recharges into land surfaces such as rivers, lakes, and oceans [23].

The underground water can be pumped out when the aquifer is shallow, and the water is able to move through the aquifer layer. The height of the water table is greatly influenced by precipitation patterns, climate, land surface changes, and human activities. We separated the soil above the groundwater into four layers, the shallowest layer is layer one and the deepest layer is layer four (

Table 7. Description of soil layers.

Layers	Citation	Term of Use
layer_1	m3/m3	Based on the ECMWF Integrated Forecasting System Volume of water in the soil layer 1 (0–7 cm) [16]
layer_2	m3/m3	Based on the ECMWF Integrated Forecasting System Volume of water in the soil layer 2 (7–28 cm)
layer_3	m3/m3	Based on the ECMWF Integrated Forecasting System Volume of water in the soil layer 3 (28–100 cm)
layer_4	m3/m3	Based on the ECMWF Integrated Forecasting System Volume of water in the soil layer 4 (100–289 cm)

). When layer one contains more water than layer four, the water table tends to become higher, meaning the excessive water is going to charge the water table and vice versa (Figure 1).

3.1.5. Irrigation Cycle: Runoff

Runoff is a measurement of the ability to hold water in the soil; it is an effective indicator of flood or drought in the land (

Table 8. Description of Runoff.

Description of Runoff	Term of Use	Citation
sub_surface_runoff	The water stored in the soil drains away from the sub-land surface.	Meter
surface_runoff	The water stored in the soil drains away from the land surface.	meter

). The drainage of water from the land surface is called the surface runoff, while the drainage of water from the underground level is called the subsurface runoff [24]. Two subtype runoffs fall into the runoff category. The measurement of the unit is the depth in meters, which is assumed to be spread evenly across the grid area.

3.1.6. Fertilizer Recommendation: Nitrogen

The Nitrogen prediction training data is taken directly from a research paper titled “Total nitrogen estimation in agricultural soils via aerial multispectral imaging and LIBS sensor” by Hossen et al. (2021) [25]. The dataset includes data taken from a drone-mounted sensor providing sensor data that includes features such as Red (Red band pixel value), NIR (Near Infra-Red Pixel Value), Green (Green Band Pixel Value), NDVI (Normalized Differential Vegetation Index), RH (Related Humidity), and Air Temperature. The dataset (

Table 9. Soil Nitrogen content prediction dataset—training dataset.

RED (DN)	NIR (DN)	Green (DN)	ndvi	RH (%)	Temp (°C)	Nitrogen (ppm)
47.88	328.05	54.9	0.74	33.8	23.2	1156.89
51.99	329.97	56.5	0.72	33.8	23.2	1118.38
52.37	325.9	56.14	0.72	33.8	23.2	1079.95
55.81	200.32	43.11	0.56	33.8	23.2	466.82
59.3	209.11	45.47	0.55	33.8	23.2	462.82
56.36	208.75	44.72	0.57	33.8	23.2	459.04
74.64	223.63	52.27	0.5	33.8	23.2	1110.15
78.95	229.52	54.5	0.49	33.8	23.2	1073.99
79.21	230.47	55.09	0.49	33.8	23.2	1039.71
85.53	226.03	55.86	0.44	33.8	23.2	996.15
98.47	238.19	61.73	0.41	33.8	23.2	965.72
102.15	247.7	64.07	0.41	33.8	23.2	936.81
44.84	287.68	49.64	0.73	33.8	23.2	1013.97
47.37	312.47	53.44	0.74	33.8	23.2	982.64

) include the label in terms of actual Nitrogen content in the soil in ppm.

3.2. Data Source and Pre-Processing

3.2.1. Step 1: For the Satellite Spectral Data, Vegetation Indices Are Calculated Based on Satellite Bands, with Assumed Adjustments and Coefficients

In

Table 10. Total bands based on satellite remote sensors.

Name	Description	Resolution (meters)	Wavelength (µm)
B1	Blue	30	0.45–0.52
B2	Green	30	0.52–0.60
B3	Red	30	0.63–0.69
B4	Near-infrared	30	0.77–0.90
B5	Shortwave infrared 1	30	1.55–1.75
B8	Band 8 Panchromatic	15	0.52–0.90

, the bands used are shortwave bands from band 1 to band 8. B1 represents blue color with 0.45–0.52 µm, B2 represents green color with 0.52–0.60 µm, B3 represents red color with 0.63–0.69 µm, B4 represents near-infrared with 0.77–0.90 µm, B5 represents shortwave infrared with 1.55–1.75 µm, and B8 represents Panchromatic with 15 m resolution and 0.52–0.9 wavelength. Multiple vegetation indices are calculated based on referenced formulas [26], with some adjusted parameters and coefficients (

Table 11. Vegetation index formulas.

Vegetation Index		Formula	Reference
NDVI	Normalized Difference Vegetation Index [26]	NDVI = NIR−VIS/NIR+VI	Rouse et al. (1974).
EVI	Enhanced Vegetation Index	EVI = G  · (NIR − Red)/(NIR + C1 · Red − C2 · Blue + L) *	Huete et al. (2006) [27]
LSWI	Land Surface Water Index	LSWI = NIR − SWIR/NIR + SWIR
NDSI	Normalized Difference Snow Index	NDSI = (Green − SWIR)/(Green + SWIR)
MSR	Modified Simple Ratio	MSR = (SWIR/NIR − 1)/((SWIR/NIR).power(0.5) + 1)	Chen, 1996 [28]
SAVI	Soil-Adjusted Vegetation Index	SAVI = (1 + L) × (SWIR − NIR)/(SWIR + NIR + L) *	Huete (1988) [29]
MSAVI	Modified Soil-Adjusted Vegetation Index	0.5 × 2 × (SWIR + 1) − ((2 × SWIR + 1) × 2–8 × (SWIR − NIR)).power(0.5)	Broge and Leblanc (2000) [30]
MCARI	Modified Chlorophyll Absorption Ratio Index	((B8 − NIR) − 0.2 × (B8 − Red)) × (B8/NIR)	Daughtry et al. (2000) [31]
TVI	Triangular Vegetation Index–((SWIR − NIR)/(SWIR + NIR) + 0.05).power(0.5)		Broge and Leblanc (2000)

* L is the background adjustment; here, we assume L equals 1. C 1 and C2 are the coefficients to correct the red band reflectance; we assume C1 equals 6 and C2 equals 7.5.

).

3.2.2. Step 2: For the Irrigation Module, Crop Water Requirement Is Calculated

For the crop water requirement (CWR), we combined three water outflow parameters for the final equation: the evaporation value, the total runoff value and the underground water amount. The negative value for evaporation means the water flows out of the land surface, and the positive value for runoff represents the water running away from the soil layer. “ET_runoff” presents the sum-up of evaporation and runoff. Then, we added up the soil water values from the 4th layer to the 1st layer to get the total ‘Underground Water’ parameter.

$$\begin{array}{ll}\mathrm{CWR}& =\mathrm{Evaporation}-\mathrm{Runoff}+\mathrm{Underground} \mathrm{Water}\\ & =\mathrm{Evaporation}-\mathrm{Runoff}+\mathrm{Soil} \mathrm{water} \mathrm{layer} 4-\mathrm{Soil} \mathrm{Water} \mathrm{Layer} 1\end{array}$$

3.2.3. Step 3: For the Fertilizer Recommendation, the Criteria to Bucket the Data Are Defined and Continuous Elements Are Separated into Different Buckets [7]

As mentioned in the data preprocessing section, the label available in the training dataset is a continuous value and hence is not suited for classification models. Hence, as part of the data transformation, we converted the continuous variables to discrete values. The discrete values were grouped into 3 ‘Tier’ ranges. Low-1 (0–500 ppm), Medium-2 (501–1000 ppm), and High-3 (>=≥1001 ppm). The processed dataset is shown Figure 2.

3.3. Data Parameter Selection and Analysis

To analyze the difference in vegetation indices behavior, we collected the spectral reflectance based on Landsat 8 collections for different crop types.

Figure 3 presents the linearity relationships between vegetation indices and near-infrared band (NIR), with the reason that NIR is a good indicator of foliage cover and crop growth. The more leaf area variation depicted by the NIR index, the greener the pigment shows. From the picture we can see that MSR is the most sensitive vegetation index to the variation of NIR, ranging from 0.2 to 2.0 when the NIR increased from 0.2 to 0.6. MCARI has the weakest correlation with NIR, keeping flat around 0.0 no matter how NIR changed. TVI also did not show an apparent relationship with NIR and became saturated at around 0.5 with the changes in NIR. The sensitivities of other vegetation indices sit among MSR and MCARI, they behaved quite similarly with an average slope of 0.6.

In the training features, we removed MCARI and TVI since they show almost no correlation with the crop cover. In contrast, we selected MSR, EVI, LSWI, SAVI, and MSAVI as the features for machine learning models.

In the prepared irrigation dataset, we select 8 input variables: CWS, ET_runoff, underground water, CWR, heat flux, skin temperature, wind, and NDVI. It is noted from Figure 4 that CWS has a linear relationship with underground water, CWR, ET_runoff, and skin temperature. So, we will first try to build linear regression models with regularizations among CWS and other features.

3.4. Data Preparation and Segregation

We divided the dataset into training and testing datasets with 80% for training and 20% for testing (

Table 12. Descriptions of Datasets.

Dataset	Training			Testing
	Rows	Input Features	Labels	Rows	Input Features	Labels
Crop Identification	17,516	9	1	4379	9	1
Irrigation Cycle	2030	6	1	507	6	1
Fertilizer Recommendation	14,096	5	1	3524	5	1

). Since the label output is an abnormal distribution, we stratified the training and testing dataset according to the weights of crop amounts. There are 17,561 rows of data, 9 input features, and 1 label in the crop identification dataset. For the irrigation cycle dataset, there are more than 2000 rows for training, 6 input features, and 1 label. For the fertilizer recommendation, we select 14,096 rows of data for the training dataset, 5 input features, and 1 label.

4. Model Development

4.1. Random Forest

Random Forest is one of the most widely used Classification algorithms, which uses the Ensemble Learning Technique [32]. One of the main advantages of Random Forest is that it creates many decision trees on the input subset data and combines all the outputs into a final prediction as shown in Figure 5 [33]. Random Forest methods are less prone to overfitting problems and tend to have better accuracy than other classification methods. For the crop identification module [34], we ran the randomized search on the Random Forest model with 5 cross-validations to find the optimized parameters. The best parameters are 100 estimators, 2 minimal samples split, 100 maximum depths, and entropy as the criterion.

4.2. SVM

Support vector machines are a supervised learning approach for classification. With the kernels function defined as polynomials, SVM is also an efficient nonlinear classifier of high-dimensional space. There are two primal parameters tuning the SVM, one is the C parameter to soften the margin, and another is the gamma parameter to define the radius influence. For the crop identification module, grid search was first deployed to find the optimal estimators. The best parameters are ‘kernel = rbf’, ‘gamma = 0.25’, ‘C = 200’, and ‘cross validation = 5’.

4.3. AdaBoost

AdaBoost is an ensemble method to convert weak learners into strong learners. The mistakes from the previous stump are considered for the next one. The weights of different stumps are dependent on error calculations, the incorrect data predicted in the former stump will be given more weight in the next stump-making process. For the crop identification module, the following parameters are tuned: base estimator, number of estimators, learning rate, and depth of the decision tree. The best parameters are ‘base_estimator_max_depth = 4, ‘learning_rate= 0.1’, and ‘n_estimators = 100’.

4.4. Linear Regression Model

A simple regression model is formed with one input (X) to estimate the output (Y). In the higher dimension linear regression model, there are multiple inputs (X) in the equation and the line is called a hyper-plane. We deployed a linear regression model and a polynomial regression model with multiple regularizations. For the irrigation cycle module, the parameters used in polynomial regression models are ‘alpha = 0.1’, ‘alpha = 0.01’, ‘alpha = 0.0001’, and ‘degree = 2’. Moreover, we put different combinations of inputs to evaluate the result of y. By using statistical approaches such as mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and R-squared, we seek optimized accuracy.

Y = B0 + B1x + B2x + B3xy

4.5. Multilayer Perceptron (MLP)

Multilayer perceptron (MLP) is a kind of feed-forward neural network consisting of an input layer, output layer, and hidden layer. Each layer is fully connected with other layers. MLPs are widely used to solve supervised machine learning problems. The weights and bias during the training can be adjusted to minimize the classification error [35]. For the crop identification module shown in Figure 6, there are three dense layers (each with 40 neurons) following a ‘relu’ activation function, and then a regulatory layer with a 10% dropout is used before the output layer. For the fertilizer recommendation module, it shares the same MLP structure but with different inputs.

5. Model Evaluation

In

Table 13. Crop identification evaluation metrics.

Model		Accuracy of Selected Crop Type	Accuracy of All Crop Type
K-Nearest Neighbors		74.62%	38%
Support Vector Machine	Kernel = linear	71.87%	41.31%
	Kernel = rbf	79%	45.31%
Random Forest		76.07%	69%
AdaBoost		72.22%	52%
MLP (400 epochs)		78.62%	71%

, there are five machine learning models built for the crop identification module. Only the AdaBoost model has the overfitting issue as the training size grows. In Figure 7, the learning curve of SVM goes down, apparently, and the gap between the training curve and validation curve shrinks since more data samples are put in the model. The deep learning model performed the best with 78% accuracy, whose loss dropped down quickly before the first 20 epochs and stayed saturated at around 0.75. The model accuracy also changed evidently before the first 20 epochs and improved slowly to 0.78 after 300 epochs. Hence, the MLP model is selected as the candidate model for the system.

Based on the irrigation cycle evaluation metrics shown in

Table 14. Irrigation cycle evaluation metrics.

Model		RMSE	MSE	R2	Training Score	Testing Score
Linear Regression		0.256	0.06554	0.939	0.93936	0.935395
Lasso Regularization	Alpha = 0.1	0.292	0.08535	0.917	0.92048	0.917162
	Alpha = 0.01	0.260	0.06759	0.934	0.93768	0.934405
	Alpha = 0.001	0.256	0.06554	0.936	0.939358	0.936394
Polynomial Regression (degree = 2)		0.244	0.05983	0.939	0.939458	0.936400

, there are three linear regression models built to predict the crop water supply. One is the original linear regression, the second is the linear regression with Lasso regularization, and the third is polynomial regression. Gridsearch is conducted to all the hyperparameters tuning alpha = 0.1, 0.01, 0.001; degree = 2. The best R-squared score is 0.939 from polynomial regression, and the best testing score is 0.9364 from polynomial regression.

Figure 8 shows the different irrigation levels across the country. The more left side of the color bar (dark blue), the drier that area is. On the west coast (lower right picture), a yellow and pink panel represents a moist land area. In the middle part of the country (upper left picture), the panel in dark blue means the soil of that region is extremely dry. When we talked about the east coast, no matter whether upper east or lower east, neither of the areas is good at retaining moisture in the soil.

For the fertilizer recommendation module shown in

Table 15. Fertilizer recommendation evaluation metrics.

Model	Accuracy	Parameters
Random Forest	82%	estimators = 10
Ada Boost	82%	estimators = 10
Gaussian NB	64%
Support Vector Machine	73%
MLP (ReLU Activation)	93.3%	Trainable params: 1024, activation = ‘relu’, loss = ‘categorical_crossentropy’, optimizer = ‘adam’

, five machine learning models were built to find the optimal results. The multilayer perceptron deep learning model performed the best, with an accuracy score of 93.3% after 100 epochs. So, we selected the MLP model as our final deployed model for the fertilizer recommendation system. We classified the continuous outputs into three bins [36]: blue means less fertilizer, yellow means neutral fertilizer, and green means rich fertilizer. From Figure 9, we can see that the first selected area has enough fertilizer in the soil, and the second selected area in dark green represents sufficient nutrients [37].

6. Data Analytics System

6.1. System Requirements Analysis

6.1.1. System Boundary and Use Cases

Our use cases mainly happen in three environments: personal workstations, research labs, and agriculture business companies. The corresponding users of our system (Soil Analysis System) are farmers who plant and study in their personal workstations, scientists in research labs, and technicians from agriculture business companies. Figure 10 illustrates three different kinds of scenarios for each user and the use case. The scenario can be summarized as follows:

A farm worker (e.g., farmer, scientist, or technician) performs the assessment of the interested farmland using this system before determining the cultivation plan. The system returns the analysis report listing crop identification, irrigation supply, and fertilizer recommendation, given the selected farmland. The system is a web application and would be run in the farm worker’s workspace (e.g., personal workstation, research lab, or inside an agriculture business company). The detailed analysis of farmlands is essential and vital for every cultivation since different farmlands should be taken care of differently, and it leads to more productive, efficient, and ecological cultivation [38].

6.1.2. System High-Level Data Analytics

The high-level data analytics of the project is made up of three tiers: reporting, insights, and prediction. At the reporting level, we construct a crop identification platform [39], which is able to recognize different types of crops based on the sensors’ bands. Each crop has its special vegetation indices features, we trained the machine learning models to learn the general disciplines among diversified crops and try to tell the type of crop.

At the insights level, we created our own formula to analyze irrigation situations based on the surface water and underground water data. The calculation of surface water is dealing with the precipitation, irrigation, and temperature of the land. The calculation of the underground water includes the different layers of water level (four layers).

In the prediction level, we forecast the fertilizer usage of one specific farmland for the next season combined with the results from the first crop identification part and the second irrigation recommendation part. Additionally, the nutrition data about the soil (NPK), collected from the sensors, is critical to the prediction process. With respect to the use of the data, deep learning models will be deployed to recommend fertilizer usage.

6.2. System Design

6.2.1. System Architecture and Infrastructure

Figure 11 shows our system architecture of this project, which has a three-layer architecture: a user interface layer, a machine learning layer, and a data storage layer. The user interface layer runs on a web browser as the frontend. The user (e.g., farmer, scientist, or technician) can interact with our system via the web application. Via the user interface, the user can specify the farmland information by entering a geographical location or selecting the area of interest from the visualized map.

The data storage layer consists of two parts: a database access to the satellite data from Google Earth Engine and one database with LIBS sensor data (McMillan, 2018). After our frontend receives the query of the farmland, the frontend then sends the query to the data storage layer. According to the information, the backend extracts the corresponding data, including both satellite data and sensor data, and sends them to our trained machine learning model.

The machine learning layer consists of three modules: crop identification, irrigation supply prediction, and fertilizer recommendation. In the first module, Random Forest, SVM and MLP are deployed in identifying the crop. The irrigation cycle prediction module is performed by using linear regression with lasso regularization and polynomial regression. The fertilizer recommendation is mainly conducted by Random Forest, SVM, and MLP. After three modules complete our comprehensive soil analysis, the report of the soil analysis is returned and displayed on the user interface.

6.2.2. Inputs, Outputs, and Data Flow

The data flow shown in Figure 12 shows that the whole system includes four parts: user input, agriculture inputs, model, and module outputs. The outputs connect the three modules and support the results. In the first input part, users give the area of interest, such as latitude and longitude coordinates, to settle on the specific location they want to focus on. Then, the system will retrieve the selected area from the global satellite dataset in Google Earth Engine. Based on the algorithm to retrieve data, the system will pull the spectral bands from the dataset. The output from the crop identification model flows into the second module as the inputs, combined with climate data and surrounding water data. The chemical prediction model will be fed with soil moisture data from the second module and spectral band data from the first module.

7. Visualization Platform

Our visualization platform describes the real-time soil conditions for the United States based on the Landsat 8 dataset. We offer a variety of commonly used vegetation indices, with which we can find the crop classification for certain farmland areas. Moreover, the platform enables the users to check the water demand for irrigation and fertilizer usage recommendations. Our goal is to help users monitor cropland vegetation conditions and predict the soil condition. The observable time ranges from 1984 to 2022; more than 20 years of observation data can analyze the geospatial time series and detect changes in many research fields [40] (Figure 13).

At the top left of the UI, there is a weight to select the location or data that the user is interested in. There are three ways to search the location, the first way is the ‘name/address’ function searching by place name or address such as Paris, the second way is the ‘lat-lon’ function searching by ‘lat-lon’ such as 40, −100, the third way is ‘data’ function searching by GEE data catalog by keywords such as elevation. The map can zoom in to the county level and click certain counties in the USA such as San Joaquin (Figure 14). Once the county is selected, the user can select whether to download the chart data to check the statistics. All the downloaded data is stored as CSV files, and the line chart can also be saved in the client’s local server (Figure 15).

The moisture level is presented by a color bar. Blue means the area is dry and red means the area is moist. The colors between blue and red in the color bar present the middle level of soil humidity. From Figure 16, we can say that the soil of the San Joaquin area is relatively moist, with a yellow and pink pallet. The users can check the soil moisture level for other counties by clicking the polygon of that county.

The fertilizer condition is classified into three categories, less fertilizer (blue), neutral fertilizer (yellow), and more fertilizer (green) in Figure 17. The selection area can be focused on one specific farmland, by drawing the area of interest (AOI).

8. Discussion

8.1. Soil Nutrient Analysis Incorporates the Features from Crop Spectral Data and Soil Biological Data

We have assembled a set of attributes for a soil health detection system, based on three attributes: crop characters, soil biological status, and soil nutrient status. These features provide data to be used in our final platform for the users to assess the soil health information. For the crop identification module, five vegetation indices calculated from remote sensing data, MSR, EVI, LSWI, SAVI, and MSAVI, are used in the final categorized machine learning models, of which MLP performed the best with 7 selected crop types. For the irrigation supply module, the linear regression model is applied with 6 input features, and the ‘evaporation and runoff’ feature is the most correlated one with a 0.68 coefficient factor, wind feature is the second with a 0.4 coefficient factor. Since all the p-values are less than 0.05, we kept all the input features in the final linear regression model. For the fertilizer module, RH and temperature are added into consideration as the input features along with the spectral data and biological data.

8.2. Creative Soil Analysis Platform Confusion

The traditional way to collect soil data is field observation via crop reports or sensors, which is time-consuming and has high costs. The modern way is to measure passive land surface microwave emission and radar backscatter, which is more robust and reliable. Most soil analysis systems combine those data sources with geospatial data to visualize and monitor soil status.

Our nutrient analysis approach uses the result from the crop identification module and irrigation module, providing an efficient way to utilize physical and biological soil data with cost efficiency. Our platform is a creative system that combines multiple soil analysis steps into one work, benefiting small-scale farmers specifically. In the future, we will deploy customized research results based on local drone and sensor data.

8.3. Farmland Biological and Chemical Features Are Not Dependent on the Farmland Geolocation

Our soil analysis system incorporates satellite information, which takes the longitude and latitude of the field as inputs to assist the users in monitoring their interested field even if they work remotely. It benefits the remote research happening in companies or labs for soil analysis. According to the changes in the color bars from the interactive map, we found that the west coast of America is moister than the east coast, despite both lands being located on the coast. The fertilizer situation is also not directly related to the geolocation but is more dependent on the single farmland environment. Even the nearby lands could have completely different fertilizer levels. No IoT or pH sensors are used to collect agriculture chemicals.

We will try to build a data pipeline, which can update data automatically. Soil health analysis is routinely challenged by the dynamic and complicated nature of environmental changes. A dynamic modelling system can demonstrate continuous value across the farmlands to help users understand and predict the dynamic environment and behavior in support of decision making.

9. Conclusions

Our integrated soil analysis approach has three main contributions as follows:

The first contribution is to combine soil health analysis with crop type, land water, and fertilizer usage. The precision of the crop type classification is dependent on the spectral bands data, then the crop type information will be infused into the irrigation prediction system. All these crop and land water data are the inputs for the fertilizer analysis. In such a way, we minimize the number of parameters coming from outside resources and use our own dataset via cyclic utilization.

The second main contribution is the irrigation calculation. The theory behind the calculation comes from the natural hydrologic system, which is a cyclic water environment of the earth including precipitation, evaporation, transpiration, streamflow, and underground water. We calculated the crop water supply (CWS) against the crop water requirement (CWR). Both parameters consist of several separate data sources. For the crop water supply (CWS), precipitation and underground water pumping are included. For the crop water requirement (CWR), evaporation and runoff are considered.

The third main contribution is that the interactive map platform we built can check the crop and soil information for farmland users. They not only can click the functional buttons that we built in, but also customize their own area of interest to find one certain farmland’s situation. Moreover, the users can download the data, which are integrated from multiple public data sources via our system, saving the user’s time with no need to search for the data separately online.

Currently, our application only takes the geolocation data of the regions in San Joaquin County. In the future, we want to expand our application nationwide. Moreover, we would deploy crop analysis on every single crop, providing more detailed end-to-end analytical reports for the users.