Assessing Model Trade-Offs in Agricultural Remote Sensing: A Review of Machine Learning and Deep Learning Approaches Using Almond Crop Mapping

Abstract

This study presents a comprehensive review and comparative analysis of traditional machine learning (ML) and deep learning (DL) models for land cover classification in agricultural remote sensing. We evaluate the reported successes, trade-offs, and performance metrics of ML and DL models across diverse agricultural contexts. Building on this foundation, we apply both model types to the specific case of almond crop field identification in California’s Central Valley using Landsat data. DL models, including U-Net, MANet, and DeepLabv3+, achieve high accuracy rates of 97.3% to 97.5%, yet our findings demonstrate that conventional ML models—such as Decision Tree, K-Nearest Neighbor, and Random Forest—can reach comparable accuracies of 96.6% to 96.8%. Importantly, the ML models were developed using data from a single year, while DL models required extensive training data spanning 2008 to 2022. Our results highlight that traditional ML models offer robust classification performance with substantially lower computational demands, making them especially valuable in resource-constrained settings. This paper underscores the need for a balanced approach in model selection—one that weighs accuracy alongside efficiency. The findings contribute actionable insights for agricultural land cover mapping and inform ongoing model development in the geospatial sciences.

1. Introduction

The agricultural sector has been a cornerstone of human civilization, providing essential resources such as food, raw materials, and employment [1]. It is a vital economic engine that sustains both local communities and broader economies, especially in regions where agriculture is the predominant activity [2]. However, the role of agriculture is increasingly under pressure due to global challenges such as climate change, water scarcity, and the need for sustainable resource management [3,4]). These pressures highlight the importance of optimizing agricultural practices and ensuring efficient resource allocation to secure food production in the face of growing uncertainties [5].

In this context, almond production in California’s Central Valley stands out as a critical component of the state’s agricultural landscape. California produces over 80% of the world’s almonds, with the Central Valley playing a pivotal role due to its favorable climate and extensive irrigation infrastructure [6]). This region is responsible for more than half of the United States’ fruits, vegetables, and nuts, making it one of the most productive agricultural areas globally [7,8]. Almonds, in particular, represent a significant share of the state’s agricultural output, contributing over 10% to the overall agricultural income [9]. Given this, accurately monitoring and managing almond crop areas is essential for sustaining production levels, ensuring resource efficiency, and addressing the growing challenges posed by climate variability and extreme weather events.

Recent expansion of almond and other perennial crop acreage in California’s Central Valley has attracted significant attention for its intensive reliance on groundwater, especially during extended droughts [10]. Remote sensing studies have indicated a shift towards nut crops such as almonds between 2007 and 2016, coinciding with increased groundwater pumping [11]. Independent research has also linked this agricultural intensification to substantial declines in groundwater levels and related environmental concerns [12]. Persistent groundwater overdraft has led to land subsidence, loss of aquifer storage, and deteriorating water quality, which carry significant implications for the sustainability of almond cultivation in this region [13,14]. Against this backdrop, improved spatial monitoring of almond acreage via remote sensing provides essential tools for informing sustainable water management and policy responses in the Central Valley.

Remote sensing technology has emerged as a powerful tool in this endeavor, offering the ability to monitor crops over large areas with high temporal and spatial resolution [15]. The use of satellite-based remote sensing, in particular, has revolutionized agricultural monitoring by providing continuous data that can be used to assess crop health, estimate yields, and detect changes in land cover [16,17,18]. Among these technologies, Landsat data has proven to be especially valuable due to its extensive temporal coverage and ability to capture detailed spectral information [19,20]. With this rich dataset, the classification of crops such as almonds has become increasingly feasible, enabling more accurate predictions of crop extent and health [21,22]. However, the challenge remains in selecting the most effective models for processing these vast datasets.

Machine learning (ML) and deep learning (DL) techniques have shown great promise in remote sensing applications, including land cover classification [23]. ML algorithms, such as Random Forest (RF) and Support Vector Machines (SVMs), have been widely used for crop classification due to their ability to handle large datasets and produce accurate results with relatively low computational costs [24,25,26]. These models rely on labeled training data and use statistical methods to identify patterns within the data, making them well-suited for applications where data availability is constrained [27,28]). In contrast, DL models, which utilize complex neural networks with multiple layers, have demonstrated exceptional performance in handling high-dimensional datasets, such as multi-spectral imagery from satellite sensors [29]. DL techniques like Convolutional Neural Networks (CNNs) and more advanced architectures such as U-Net and DeepLabv3+ can automatically extract features from raw data, often yielding higher classification accuracies than traditional ML models [30].

While DL models offer superior accuracy, they come with significantly higher computational demands and require large amounts of labeled training data, often spanning multiple years, to perform optimally [31,32]. This difference in resource requirements highlights a crucial trade-off between accuracy and efficiency. In resource-limited scenarios, where computational power, time, and data availability may be constrained, conventional ML models present a compelling alternative to more complex DL methods [33,34]. This study aims to explore these trade-offs in the context of almond crop classification in California’s Central Valley. Given the critical importance of almond production in California’s agricultural landscape, this study aims to leverage the power of remote sensing technology alongside both traditional ML and DL models to enhance the identification and monitoring of almond crop locations. The primary objective is not only to achieve high classification accuracy but also to evaluate the trade-offs between computational efficiency and precision. By comparing the performance of predictive models developed from Landsat data, this research seeks to offer practical insights into optimizing resource allocation, improving sustainability practices, and informing policy decisions related to agricultural management.

Building upon previous research, this study addresses two central research questions: (1) How accurately can remote sensing technologies, when paired with conventional ML and DL models, classify almond crop locations in California’s Central Valley? (2) What are the comparative benefits of using ML models versus DL models, particularly in resource-constrained environments where computational efficiency is critical? In alignment with these questions, the research posits the following hypotheses: (1) Remote sensing technologies, combined with ML and DL methods, will significantly improve the accuracy of almond crop classification. (2) While DL models may provide slightly higher predictive accuracy due to their complex architectures, conventional ML models will prove to be more efficient in terms of computational costs and time, offering a competitive alternative in scenarios where resources are limited. The novelty of this study lies in its comprehensive evaluation of both ML and DL models for almond crop classification, with a specific focus on balancing accuracy with efficiency. By incorporating historical data and employing advanced DL architectures, alongside more resource-friendly ML models, this research provides a rigorous performance comparison that has direct implications for agricultural sustainability.

2. Methods

2.1. Study Area

The Central Valley of California is situated between the Sierra Nevada to the east and the littoral mountain range to the west (Figure 1). It contains approximately 6.5 million inhabitants [35]. In the broad, alluvial-filled structural trough, more than 250 distinct crops are grown with an estimated annual value of more than USD 20 billion [36]. The Central Valley experiences a Mediterranean climate, characterized by most of its rainfall occurring from November to March [37]. The climate experiences very little precipitation during late winter, summer, and early fall, which results in an irrigated area of 52,000 km², being among the largest irrigated regions globally [38,39]. The Central Valley consists of three distinct regions: The San Joaquin Valley in the central area, the Sacramento Valley in the northern part, and the semi-arid Tulare Basin in the southernmost section. The region features a coastal range with significant coastal urban areas to the west, Shasta National Forest in the north, Sierra Nevada Mountains to the east, and Mojave Desert to the southeast, all of which define the Central Valley [37] (Figure 1).

2.2. Data Input

2.2.1. Satellite Image Data

Landsat imagery from 2008 to 2022 was used in this study, and three spectral bands were selected based on their high discriminatory power for vegetation analysis (https://gisgeography.com/landsat-8-bands-combinations/ accessed on 26 July 2025). The chosen band combination—SWIR1, NIR, and a visible band (either red or blue depending on the Landsat generation)—was employed due to its proven effectiveness in capturing vegetation health dynamics (

Table 1. Landsat data obtained for each year of the analysis, indicating the Landsat sensor, image date, and bands extracted. All images have a pixel size of 30 m by 30 m.

Satellite	Date	Bands Extracted—Bandnumber, Name, Wavelength & Resolution
Landsat 5	2008–2009	Band 3 Visible Red (0.63–0.69 µm) 30 m Band 4 Near-Infrared (0.76–0.90 µm) 30 m Band 5 Near-Infrared (1.55–1.75 µm) 30 m
Landsat 7	2012	Band 3 Red (0.63–0.69 µm) 30 m Band 4 Near-Infrared (0.77–0.90 µm) 30 m Band 5 Short-Wave Infrared (1.55–1.75 µm) 30 m
Landsat 8–9	2013–2022	Band 2—Blue (0.45–0.51 µm) 30 m; Band 5—Near-Infrared (0.85–0.88 µm) 30 m; Band 6—SWIR1 (1.57–1.65 µm) 30 m

). The Near-Infrared (NIR) region, spanning approximately 0.7 to 1.3 µm, has consistently been identified as optimal for crop monitoring because of the strong absorption of visible light by chlorophyll and the substantial reflectance of NIR radiation by healthy plant foliage. Additionally, Shortwave Infrared (SWIR1) is sensitive to vegetation water content, while visible bands contribute to detecting plant pigments and overall canopy condition (https://eos.com/make-an-analysis/agriculture-band/ accessed on 26 July 2025).

For each year between 2008 and 2022, a single Landsat scene was selected within the July–August window to coincide with the almond canopy’s mature and spectrally stable phase. This mid-season acquisition strategy minimized phenological variability and ensured consistent image quality across years. To maintain spectral coherence across Landsat generations, a common set of three bands was utilized based on their proven relevance in vegetation analysis. For Landsat 8, the selected bands included Band 6 (SWIR1: 1.57–1.65 µm), Band 5 (NIR: 0.85–0.88 µm), and Band 2 (Blue: 0.45–0.51 µm). For Landsat 7, Band 5 (SWIR1: 1.55–1.75 µm), Band 4 (NIR: 0.77–0.90 µm), and Band 3 (Red: 0.63–0.69 µm) were employed, while the same configuration was adopted for Landsat 5, with Band 5 (SWIR1: 1.55–1.75 µm), Band 4 (NIR: 0.76–0.90 µm), and Band 3 (Red: 0.63–0.69 µm). This consistent band selection across sensors was instrumental in enhancing the detection of vegetation vigor and crop structural attributes throughout the temporal scope of the study.

To ensure compatibility with PyTorch 2.0.0-based semantic segmentation architectures, all satellite imagery was preprocessed into three-band RGB composites and converted into an 8-bit unsigned integer format. This transformation was necessary as the majority of pre-trained deep learning models in PyTorch are optimized for three-channel RGB inputs, reflecting the structure of natural color images. Although Landsat data provide a wide range of spectral bands and are commonly distributed in 16-bit format, direct use of all bands would have required significant architectural modifications to the model, increased computational demands, and potentially reduced training efficiency. Furthermore, the use of all available spectral bands was intentionally avoided to mitigate risks of model overfitting and to maintain computational tractability, particularly given the limited size and spatial coverage of the training dataset. Many Landsat bands are spectrally correlated, and their inclusion can introduce redundancy without improving model performance. Instead, this study employed a targeted band selection strategy focused on three bands: Shortwave Infrared 1 (SWIR1), Near-Infrared (NIR), and one visible band (either red or blue). This combination was selected based on its demonstrated effectiveness in prior remote sensing studies for capturing vegetation characteristics such as chlorophyll concentration, canopy structure, and moisture content—key parameters for land cover and crop classification. To optimize the number of training samples and preserve spatial context, a chip size of 64 × 64 pixels was adopted. This approach allowed efficient training and reduced memory requirements. Consequently, Landsat imagery acquired between 2008 and 2021 was standardized into 8-bit, three-band RGB chips for model training and validation. Representative examples of these image chips are provided in Figure 2.

2.2.2. Crop Data for Almond Locations and Training

The USDA creates the Cropland Data Layer (CDL) annually for the continental US using moderate resolution satellite imagery and extensive agricultural ground truth [40]. The crop-specific data layer is freely available online (Available at https://nassgeodata.gmu.edu/CropScape/ accessed on 26 July 2025). The USDA Cropland Data Layer (CDL) was selected for this study due to its standardized nationwide coverage, temporal consistency, and seamless integration with remote sensing platforms such as Google Earth Engine. Despite known regional limitations, its widespread use in academic research and compatibility with national agricultural statistics make it a practical and reproducible reference for multi-year crop classification. CDL data spanning from 2008 to 2021 were obtained for training the DL models (Figure 3). These images served as the ground truth for almond class identification. Additionally, a 2022 image (Figure 4) was used for validation data collection for the DL-based approaches.

The training data utilized in the ML process were derived from the CDL data layers for 2022. Two separate classes were designed, one for the almond class and the other for all other classes. Polygons were created for both the almond and non-almond classes based on the 2022 CDL data layer (Figure 4). In the context of ML, a single mask was employed for this purpose. ML relies on fewer computational resources and less intricate models in comparison to DL in this study. Shapefiles derived from past CDL layers were utilized to create training data for ML models.

In the context of ML, the training data utilized consisted exclusively of the produced CDL data layers from the year 2022, while the input for the ML models involved the Landsat image of the same year. Conversely, for DL, the training phase incorporated CDL data layers spanning the years 2008 to 2021, with the year 2022 being reserved for validation purposes. In summary, our approach involved utilizing a single image as input for the ML models, while for the DL models, we employed a dataset consisting of images spanning the years 2008 to 2021 for training purposes. We aim to evaluate the performance of ML models in scenarios when the available data is limited, in comparison to DL methodologies, and when there is a desire to conserve computational resources and minimize time expenditure for the research. In the discussion we will also consider these data requirements in the model comparison as the DL model requires a significant increase in data inputs compared to the ML models.

2.3. Model Selection and Setup

A frequent occurrence in ML and DL models is overfitting, which occurs when the model performs admirably on the training data but inadequately on validation or unobserved data. To combat overfitting, a variety of techniques and strategies have been implemented. We augmented the dimensions of the training dataset to acquire a more comprehensive understanding of the issue at hand. The DL processors were designed with dimensions of 64 × 64 in order to optimize the chip count, which is approximately 6610 in total. Additionally, a DL model utilizing the Resnet-50 architecture was implemented to increase the intricacy of the model (

Table 2. Hyperparameters of machine learning and deep learning models.

Models	Hyperparameters
Deep Learning (General Structure)	ENCODER = “resnet50” ENCODER_WEIGHTS = ‘imagenet’ CLASSES = [“Almond”] ACTIVATION = ‘sigmoid’ DEVICE = ‘cuda’ Epoch = 150 chip_size = 64, stride_x = 8, stride_y = 8, crop = 12, n_channels = 3
Linear Regression (LR)	LinearRegression ()
Logistic Regression (LGR)	LogisticRegression ()
Decision Tree (DT)	DecisionTreeClassifier ()
Gaussian Mixture Model (GMM)	GaussianMixture (n_components = 3)
Gradient Boosting (GB)	GradientBoostingClassifier (n_estimators = 100, learning_rate = 0.1, max_depth = 3)
K-Means Clustering (K-Means)	KMeans (n_clusters = 100)
K-Nearest Neighbors (KNN)	KNeighborsClassifier (n_neighbors = 3)
Multi-Layer Perceptron (MLP)	MLPClassifier (hidden_layer_sizes = (150, 100, 50), max_iter = 100, activation = ‘relu’, solver = ‘adam’)
Naive Bayes (NB)	MultinomialNB ()
Support Vector Machine (SVM)	SVC (C = 1.0, kernel = ‘rbf’, gamma = ‘scale’)
Extreme Gradient Boosting (XGB)	params = { ‘max_depth’: 3, ‘learning_rate’: 0.1, ‘n_estimators’: 50 } XGBClassifier(** params, tree_method = ‘gpu_hist’, predictor = ‘gpu_predictor’, gpu_id = 1)
Random Forest (RF)	RandomForestClassifier (n_estimators = 500, oob_score = True, verbose = 1)

Note: SVC = Support Vector Classifier; rbf = Radial Basis Function; stride_x / stride_y = patch stride in pixels; n_channels = number of input channels; n_components = number of Gaussian components; n_clusters = number of clusters; n_neighbors = number of neighbors; n_estimators = number of trees; gpu_hist = GPU-accelerated histogram algorithm; gpu_predictor = GPU-based prediction; gpu_id = GPU identifier; oob_score = out-of-bag validation; ** params = Python syntax for keyword arguments from a dictionary.

). Data augmentation methods that generate variants of the training data using random operations (such as rotation, inversion, and cropping) were also implemented. This effectively expands the training dataset, thereby enhancing the model’s ability to generalize. Finally, experiments were conducted involving various hyperparameters, including the learning rate, sample size, and epoch count to determine the optimal configurations that strike a balance between training accuracy and model complexity.

2.3.1. ML Models

In the comparative analysis of ML and DL models for large-scale agricultural land cover classification, we evaluated a diverse set of ML algorithms, focusing specifically on almond crop classification using 2022 imagery and training data from the CDL layers. The ML models included in this study are Linear Regression, Logistic Regression, Naive Bayes, Gaussian Mixture Model, K-Nearest Neighbors, Decision Trees, Random Forest, Gradient Boosting, XGBoost, and Multi-Layer Perceptron.

Linear Regression is a simple predictive model, but its limitations in capturing non-linear relationships restrict its performance in complex applications [41,42]. Logistic Regression extends this by modeling binary outcomes [43,44], while Naive Bayes, despite its assumption of feature independence, performs well in image classification tasks [45,46]. Support Vector Machines (SVMs) are particularly effective with small training datasets and excel in classification accuracy [47,48], while K-Nearest Neighbors (KNN) relies on proximity-based classification and is straightforward to implement [49]. K-Means, an unsupervised method, is frequently employed for clustering [50]. The Gaussian Mixture Model (GMM) offers probabilistic clustering by combining Gaussian distributions [51]. Decision Trees (DTs) are highly interpretable models used in classification tasks [52], and Random Forest (RF) enhances DT by aggregating multiple trees to reduce overfitting and improve predictive accuracy [53]. Gradient Boosting (GB) and XGBoost (XGB), both ensemble learning methods, iteratively improve prediction by minimizing loss functions [54,55]. Finally, Multi-Layer Perceptron (MLP), a type of neural network, captures non-linear relationships through its multiple layers [56].

2.3.2. DL Models

In this study, we employed a variety of DL models for almond crop classification, trained using CDL data from 2008 to 2021. The models include UNet, UNet++, Multi-Scale Attention Network, LinkNet, Feature Pyramid Network, Pyramid Scene Parsing Network, DeepLabv3, and DeepLabv3+. These models were trained using ResNet-50 and ResNet-18 backbones, leveraging pre-trained ImageNet weights for transfer learning to enhance performance in agricultural image classification.

U-Net, a popular segmentation model, utilizes an encoder–decoder structure with skip connections to maintain high segmentation accuracy, particularly for tasks with limited training data [57,58]. UNet++ extends this by incorporating densely connected sub-networks to improve feature extraction across multiple scales [59,60]. The Multi-Scale Attention Network (MANet) focuses on efficiently segmenting high-resolution images by utilizing attention mechanisms to highlight critical features while suppressing irrelevant data [61,62]. Similarly, LinkNet modifies UNet’s structure with residual connections, enhancing computational efficiency without sacrificing accuracy [63]. Feature Pyramid Network (FPN) addresses multiscale object detection by leveraging a pyramid structure that combines high-level semantic features with low-level spatial features, improving scale invariance [64,65]. Pyramid Scene Parsing Network (PSPNet) further enhances segmentation by pooling contextual information from multiple scales, allowing for better pixel classification across diverse regions [66,67]. DeepLabv3 and DeepLabv3+ use Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale context while maintaining high-resolution feature maps, which are particularly useful for segmenting objects at varying scales [68]. DeepLabv3+ improves upon its predecessor by incorporating an explicit decoder module to better capture fine-grained details, especially at object boundaries [69].

2.3.3. Computing Requirements for Analysis and Available Resources

The computational demands of land use and land cover (LULC) classification, such as almond crop classification using Landsat data, vary significantly between ML and DL methods. ML models like DT, RF, and SVM typically rely on structured, well-labeled datasets with features extracted from spectral bands (e.g., Landsat Bands 1–7) or vegetation indices (e.g., NDVI and SAVI). While these models are relatively computationally efficient compared to DL, they still require considerable processing power during feature selection, hyperparameter tuning, and training on large datasets, like Landsat imagery, which spans vast geographic areas with 30-m spatial resolution [70,71]. CPU-based systems can handle these tasks, though cloud computing platforms can further optimize both processing and storage needs [72]. In contrast, DL models, such as CNNs, demand substantially more computational resources due to their ability to integrate both spectral and spatial features for more detailed analysis. For LULC tasks like almond crop classification, DL models benefit from large, multi-temporal datasets to capture crop growth cycles and patterns, requiring significant processing power [70,73]. The heavy computational burden of DL arises from its use of neural network layers, with numerous parameters requiring iterative optimization during training. This typically necessitates high-performance GPUs or custom hardware like TPUs to expedite training and inference [74,75]. To ensure accuracy, DL models often require extensive preprocessing steps, including normalization and data augmentation [73]. When incorporating temporal Landsat data to monitor seasonal changes, cloud platforms such as Google Earth Engine (GEE) are crucial, offering access to large satellite archives and the necessary processing capabilities for both ML and DL tasks. However, the high computational demands of DL can lead to increased cloud costs, particularly with large-scale training that requires hyperparameter tuning and cross-validation [76].

2.3.4. Accuracy Assessment of ML and DL Models

In order to precisely evaluate the performance of the models, many metrics and methodologies are utilized which are related to the type of problem at hand, i.e., image classification. These commonly used accuracy assessments are precision, recall (sensitivity), F1-score, and overall accuracy.

Precision: Simply put, precision is the quotient obtained by dividing the number of accurately predicted positive cases by the total number of occurrences predicted as positive [77]

Precision = True Positives/(True Positives + False Positives)

The term “precision” is commonly described as the extent to which an individual’s score achieved on one occasion is replicated on a subsequent occasion, a concept also known as test–retest reliability in classical language [78]. The metric evaluates the classifier’s capacity to accurately detect positive examples within its set of predicted positive occurrences. The measure is frequently employed in image classification and other classification applications. The metric under consideration measures the precision of a classifier’s positive predictions, specifically in the context of object identification in images, by comparing them to the overall number of positive predictions made. Precision is a term that is used to describe the level of accuracy or exactness in a measurement or calculation.

Recall: This refers to the proportion of accurately anticipated positive instances compared to the overall number of actual positive instances [79].

Recall = True Positives/(True Positives + False Negatives)

Recall, alternatively referred to as sensitivity or True Positive Rate, holds significant importance as a performance statistic within the realm of image classification and various other classification endeavors. The metric assesses the classifier’s capacity to accurately identify all pertinent instances, namely the ratio of correctly identified positive examples (e.g., items of interest) to the total number of actual positive instances, as determined by the model. Here, the metric denotes the proportion of almond pixels that have been correctly classified among the entire set of almond pixels.

F1-score: This represents the harmonic mean of precision and recall.

F1 = 2 × ((Precision × Recall)/(Precision + Recall))

The F1-score is usually used for the optimization of a model towards either precision or recall [80]. The F1-score assesses feature discrimination against target groups statistically. It generates a feature’s score by comparing sample mean variation to sample variation [81].

Overall Accuracy: Accuracy is a metric used to assess the correctness of a model. It is calculated as the ratio of the number of pixels that are accurately classified to the total number of testing pixels.

Overall Accuracy = (TP + TN)/(TP + TN + FP + FN)

In image classification, true positives (TPs) are the number of correctly predicted positive instances. In the context of image classification, this means the number of pixels correctly classified as the target class. True Negatives (TN) refer to the count of accurately predicted negative instances. In the specific domain of image classification, this pertains to the quantity of pixels accurately identified as belonging to the non-target category. False positives (FP) are the number of instances that were incorrectly predicted as positive. In image classification, these are the pixels that were predicted to belong to the target class but do not. False negatives (FN) refer to situations that have been inaccurately classified as negative when they should have been classified as positive. Within the context of image classification, the term “misclassified pixels” refers to those pixels that are erroneously categorized as not belonging to the intended target class. The macro technique is employed to calculate Macro-Precision, Macro-Recall, and Macro-F1-score. The macro technique computes the arithmetic mean for many types of indicators [82].

3. Results

3.1. ML Model Performance

A predicted almond classification was created for 2022 for each model used (Figure 5). In addition, in the table provided (

Table 3. Accuracy assessment results of the twelve different ML models for almond classification for the 2022 Landsat image. The highest model scores in each category are bolded.

ML Model	Precision	Recall	F1-Score	Overall Accuracy
Linear Regression—LR	0.63	0.66	0.65	95.647
Logistic Regression—LGR	0.65	0.72	0.68	95.546
K-Nearest Neighbor—KNN	0.70	0.72	0.71	96.662
K-Means Clustering—K-Means	0.59	0.67	0.62	94.145
Gaussian Mixture Model—GMM	0.62	0.90	0.67	92.209
Naive Bayesian—NB	0.65	0.77	0.69	95.264
Support Vector Machine—SVM	0.61	0.58	0.59	96.100
Decision Tree—DT	0.70	0.74	0.72	96.650
Random Forest—RF	0.71	0.73	0.72	96.798
Gradient Boosting—GB	0.70	0.74	0.72	96.615
Extreme Gradient Boosting—XGB	0.67	0.74	0.69	95.93
Multiple Layer Perceptron—MLP	0.70	0.73	0.71	96.632

), we can observe the performance metrics of different ML models on almond classification using our 2022 Landsat image. The performance metrics reported are precision, recall, F1-score, and overall accuracy. Looking individually at each model in terms of these four metrics, we can compare models.

Linear Regression (LR) has moderate precision and recall values compared to other models. Its overall accuracy is better than the GMM but lower than the ensemble and tree-based models. LGR’s performance is similar to LR but with slightly lower overall accuracy. KNN has similar precision, recall, and F1-score to the DT and GB models but slightly lower recall. Its overall accuracy is slightly higher than GB but lower than DT. The KM model exhibits inferior overall accuracy, precision, recall, and F1-score in comparison to the other models. The GMM exhibits the lowest overall accuracy. GMM has the lowest precision but the highest recall, suggesting it identifies most of the positive cases but with a high false positive rate. NB has a decent recall but lower precision, indicating a higher number of false positives. It has the second lowest overall accuracy. MLP’s performance metrics are close to those of KNN but with a slightly lower recall. The SVM algorithm has superior performance in terms of overall accuracy. However, it exhibits the lowest scores in terms of F1-score and recall. Furthermore, in terms of precision, the score of the mentioned model is higher than that of K-Means but lower than all other models (Table 3).

DT’s overall accuracy is the highest compared to all the models, except for RF and KNN. It has moderate precision, recall, and F1-score compared to others. The RF model achieves the highest overall accuracy among all the models. It has the highest precision and a comparable recall and F1-score to the top-performing models. GB shows nearly identical performance to DT in terms of precision, recall, and F1-score but has slightly less overall accuracy than the DT model. EGB has moderate precision and recall, with an F1-score and overall accuracy that are decent but not the best among the models. In summary, if your primary concern is high overall accuracy, the RF model seems to be the best choice, followed closely by the DT and KNN models. The MLP model exhibits the highest precision among the three models considered. However, when considering recall, F1-score, and total accuracy, the MLP model demonstrates outstanding performance but does not achieve the top score. However, if you are looking for a model with a high recall (identifying most of the true positive cases), the GMM would be the best despite its lower overall accuracy (Table 3).

3.2. DL Model Performance

A predicted almond classification was created for 2022 for each DL model used (Figure 6). The spatial results shown for the DL models (Figure 6) are much more consistent than those shown for the ML models (Figure 5), which varied more in terms of area of almonds predicted, as well as in locations. In addition, in the table provided (

Table 4. Accuracy assessment results of the different DL models for almond classification for the 2022 Landsat image. The highest model scores in each category are bolded.

Year	Precision	Recall	F1-Score	Overall Accuracy
U-Net	0.79	0.66	0.70	97.465
UNet++	0.79	0.62	0.66	97.394
MANet	0.78	0.62	0.67	97.338
LinkNet	0.80	0.64	0.69	97.455
FPN	0.80	0.61	0.66	97.422
PSPNet	0.80	0.60	0.65	97.404
DeepLabv3	0.79	0.62	0.66	97.380
DeepLabv3+	0.80	0.66	0.70	97.502

), we can observe the performance metrics of all the different DL models on almond classification. The performance metrics reported are precision, recall, F1-score, and overall accuracy. Looking individually at each model in terms of these four metrics, we can compare models (Table 4).

U-Net’s overall accuracy was the highest among all models except for DeepLav3+. Its precision score was on par with many other models but lower than Linknet, FPN, PSPNet, and DeepLabv3+. It had relatively low recall, and it was higher only when compared to UNet++, MANet, and DeepLav3. Its F1-score was in mid-range, with three models scoring lower and four models scoring higher. UNet++ had moderately high overall accuracy, but it was smaller than UNet. Its precision was equal to U-Net and DeepLav3 but lower than several other models, and it had one of the lowest recall scores and was only higher than PSPNet. Its F1-score was in the lower mid-range and only higher than that of PSPNet. MANet’s overall accuracy was the lowest among all the models, albeit marginally. Its precision was slightly lower than most other models, and its recall was equal to UNet++ and DeepLabv3 but lower than others. It had a mid-range F1-score, with three models scoring lower and four models scoring higher. LinkNet had a high overall accuracy, very close to the top-performing models. Its precisions score it was the highest, tied with FPN, PSPnet, and DeepLabv3+. Recall was in the mid-range, higher than UNet++, MANet, FPN, PSPNet, and DeepLabv3, and its F1-score was in the higher mid-range, with only UNet and DeepLabv3+ scoring higher. FPN had an overall accuracy that was high, only slightly lower than LinkNet. Its precision score was also one of the highest, tied with LinkNet, PSPNet, and DeepLabV3+. Its recall score was in the lower mid-range, higher only than PSPNet, and its F1-score was also in the lower mid-range, higher than only PSPNet. PSPNet had an overall accuracy in the mid-range, higher than MANet and DeepLabv3 but lower than others. Its precision was among the highest, tied with LinkNet, FPN, and DeepLabv3+. However, its recall and F1-score were the lowest among all models. DeepLav3 had a mid-range overall accuracy, higher than MANet and PSPNet but lower than others. Its precision was equal to U-Net and UNet++ but lower than others. Recall scores were in the lower mid-range, only higher than UNet++, MANet, and PSPNet, and the F1-score was similarly in the lower mid-range, higher than only UNet++ and PSPNet. Finally, DeepLabv3+ had the highest overall accuracy among all the models, suggesting it classified the highest percentage of instances correctly. In addition, this model tied for the highest precision, indicating that it had the highest percentage of correct positive predictions and had the highest recall score, tied with U-Net, suggesting that it identified the highest proportion of actual positives correctly. This model also tied for the highest F1-score with U-Net, showing that it had the best balance between precision and recall. From these results (Table 4), it appears that the DeepLabv3+ model is the top performer in almond classification based on the 2022 Landsat image, with the highest overall accuracy and the best or tied best scores in all other metrics. LinkNet also performed notably well, especially in terms of precision, which was the highest and tied with others. It is essential to note that all models have achieved high overall accuracy above 97%, indicating a generally successful classification outcome across all methodologies. Moreover, the selection of the most appropriate model may depend on whether precision or recall is prioritized, as this decision should align with the specific objectives and constraints of the study.

Looking at the results spatially (Figure 7 and Figure 8), as well as through the accuracy assessment results (Table 3 and Table 4), we can better evaluate the model performance for both the ML and DL models. In Figure 7, we present the outcomes of different ML modeling approaches applied to the classification of almond crops in our study area for the year 2022. The spatial analyst feature from ArcGIS Pro was utilized to contrast predicted almond locations against ground truth data for the Cropland Data Layer product. The maps are distinguished by a color scheme that indicates the accuracy of each model’s predictions: green areas represent true positives where the model correctly identified the presence of almond crops, magenta areas denote false positives where almond presence was incorrectly predicted, and blue areas represent false negatives where the model failed to detect actual almond crops (Figure 7 and Figure 8). The visual analysis of the classified images indicates that models RF, XGB, and MLP demonstrate a more accurate prediction with the highest proportion of correct matches (green areas). Conversely, models like KNN, K-Means, and GMM show a substantial number of false positives (magenta), indicating overestimation of almond presence. Linear Regression (LR) and Logistic Regression (LGR) models, despite their simplicity, demonstrated a higher incidence of false negatives (highlighted in blue), indicating a tendency to overlook actual almond crop areas. In Figure 8, the performance of the DL models in the classification of almond crops for the year 2022 are mapped. The visual representation of the classification outcomes reveals a nuanced view of each model’s predictive capabilities. Models such as UNet++, PSPNet, and MANet exhibit a considerable overlap of green and magenta areas. This suggests that while these models are adept at identifying almond crops, there is also a tendency to overestimate their presence, resulting in a higher rate of false positives. Conversely, models like LinkNet and FPN are marked by extensive blue areas, indicating a higher rate of false negatives, where they have missed detecting actual almond crops. Notably, DeepLabv3 and DeepLabv3+ appear to strike a balance in prediction accuracy. The interspersion of green, magenta, and blue suggests that these models achieve an equilibrium between precision and recall, reducing the likelihood of both false positives and false negatives.

Using the RF model, due to the analysis highlighting its usefulness in this study, a change analysis can be conducted to look at almond crop coverage across the study area from 2008, to 2015, to 2022. The model results (Figure 9;

Table 5. Performance metrics of Random Forest model for different years.

Random Forest	Precision	Recall	F1-Score	Overall Accuracy
2008	0.72	0.60	0.63	98.634
2015	0.69	0.80	0.73	97.158
2022	0.71	0.73	0.72	96.798

) highlight the significant expansion of almonds as a crop between 2008 and 2015 and a more stable landscape afterward.

4. Discussion

Based on this evaluation and comparison of modeling approaches within agricultural land cover classification analyses, it was determined that the RF model was the most accurate and efficient of all the ML and DL models tested. In applying our top-performing RF ML model to the task of change in almond production across our study area over the past 25 years, we saw that much of the expansion in almond crops occurred between 2008 and 2015, with smaller changes (both into and out of almond) occurring since 2015. This peak in almond production is reflected across the study area from the CDL data (Figure 3), and this analysis with the RF model allows us to extend the known locations of almonds to 2022 as a prediction. Continued extensive almond production has implications for the environment, specifically the increase in water usage, as reflected by the increased area, and the continued and maintained almond area, meaning a continued use of irrigation in this landscape. Given the recent droughts and extensive fires related to these droughts, as well as a drop in the water table across California (due to both long-term drought and increased water use for irrigation), these implications and the use of such mapping studies are of real importance for resource management [36,83,84,85,86,87].

We analyzed and compared the performance of a variety of ML and DL models in the task of almond classification based on Landsat images from 2022 (Figure 5, Figure 6, Figure 7 and Figure 8; Table 3 and Table 4). The metrics of focus in our analysis encompassed precision, recall, F1-score, and overall accuracy, each of which paints a distinct picture of the models’ capabilities and efficiency in classifying agricultural crops accurately. Upon reviewing the assessment results of the different ML models (Table 3), we found a closely contested field with RF slightly edging out others in terms of overall accuracy (96.798%), followed very closely by KNN (96.662%) and MLP (96.632%). These top performers exhibit a balance in precision, recall, and F1-score, signaling robust performance across different facets of the classification task. In contrast, the GMM lags noticeably with an overall accuracy of 92.209%, albeit boasting the highest recall of 0.90. While this high recall indicates a proficient identification of positive cases, the diminished precision and F1-score pinpoint a vulnerability in accurately distinguishing negative cases, hence yielding a larger number of false positives. The K-Means model demonstrated comparatively lower levels of overall accuracy, precision, recall, and F1-score when compared to the alternative models.

Transitioning to DL models, DeepLabv3+ emerges as the most balanced model, securing the highest overall accuracy (97.502%) alongside leading scores in precision and recall (Table 4). It exhibits a profound adeptness in both identifying true positives and avoiding false positives, thereby securing a high F1-score, which illustrates a well-rounded performance. Despite trailing slightly, models like Unet and Linknet also demonstrate commendable precision and overall accuracy, hinting at their proficient predictive capacity in almond classification. The findings emphasize the efficacy and nuanced understanding captured through the DL models as seen through high precision rates above 78% across all models. Spatial patterns are also important to evaluate when juxtaposed with the outputs of the ML models (Figure 7). The DL models’ maps (Figure 8) also reflect a more refined ability to discern complex patterns within the data, which is characteristic of the deeper architectural layers present in DL models. This advanced pattern recognition translates into a higher rate of true positives and a reduction in both false positives and false negatives, showcasing the DL models’ enhanced detection capabilities for almond crop classification (Table 3 and Table 4). The discussion of these findings in the broader context of agricultural monitoring via remote sensing is significant. It underscores the necessity to weigh the precision and computational demands of DL models against their classification accuracy. It also raises the point that selecting the most appropriate model is crucial as it must align with the specific requirements of the remote sensing task at hand.

In comparing the results of our research to those of other researchers, we can evaluate if specific types of questions or analyses result in comparable findings. For example, several studies have examined the accuracy of various ML models for image classification [88,89,90,91]. These studies consistently indicate that the RF model exhibits superior accuracy compared to other ML models. Our findings align with these results as we observed that RF outperformed other ML models in terms of overall accuracy (Table 3). The findings of Singh et al. (2022) [92] and Lamba et al. (2021) [93], who detected plant diseases, and Feizizadeh et al. (2021) [94], who monitored land use/cover change, all of which used comparative ML and DL approaches, are consistent with the results of RF when compared to DL models. These studies demonstrated that DL models outperformed RF in terms of performance. Furthermore, several studies, including those by Kirola et al. (2022) [95] and Sujatha et al. (2021) [96], have successfully identified plant diseases through the utilization of comparative ML and DL techniques. Their studies have shown that the RF algorithm exhibits exceptional performance among ML algorithms. However, when compared to DL models, RF tends to exhibit subpar performance, which is consistent with our own research findings. In another comparative study conducted by Yao et al. (2022) [97] on on the classification methodology of crops, it was observed that the RF model exhibited superior performance compared to the DNN model when the RF model was integrated with the DNN model. In general, our results support that DL models outperform traditional ML models in this domain, although an individual researcher would need to evaluate if the additional data needs are worth the slight increase in accuracy.

Despite the model output numbers and statistics, the evaluation and comparison of ML versus DL model approaches must also consider other issues in its comparison. When considering the identification of almond classes, the decision between using ML and DL entails a compromise between computational expenditure and accuracy. ML techniques have the capability to yield findings that are reasonably precise while requiring significantly less computer resources compared to DL approaches. Hence, in scenarios where there are constraints on computational resources or a need for cost-effectiveness, ML may be the more favorable option. In situations where achieving high levels of precision and accuracy is of utmost importance, particularly when it is necessary to discern small differences between various categories of almonds, the utilization of DL techniques can be advantageous. In deciding between using ML and DL models, it is crucial to consider the project’s distinct requirements and limitations while also striking a balance between processing resources and the necessary level of accuracy. Overall, then, the choice between using ML and DL for almond class identification depends on finding the right trade-off between computational resources and desired accuracy [98]. ML offers cost-effective solutions, while DL provides a powerful tool for achieving precision and handling complex data. The decision should be tailored to the specific goals and constraints of the project to strike an optimal balance between computational resources, budget, time constraints, and the desired level of accuracy. This selection of the RF ML model in this final change analysis was well validated by the findings of our research and the consideration of model efficiency, time and resources available. This research highlighted the comparison on ML and DL approaches and highlighted their usefulness within agricultural studies. Also of importance is the discussion between ML and DL models, which is one found throughout the literature on remote sensing applications [99].

As such, our experimental scope deliberately focused on established convolutional networks (U-Net, MA-Net, and DeepLabv3+) and classical ML classifiers (RF, KNN, and MLP) rather than emerging Transformer-based or hybrid architectures. While self-attention models like Vision Transformer [100], Swin Transformer [101], and data-efficient variants such as DeiT [102] have demonstrated state-of-the-art segmentation performance, they typically require extensive pretraining on large, labeled datasets and substantial GPU resources. Our Landsat-based dataset, though rich in spatial and temporal coverage, remains moderate in size, and our computational environment aligns with the resource profiles common in agricultural applications. By grounding our comparison in widely adopted convolutional and ML methods, we ensure direct comparability with the broader remote-sensing literature [88] and deliver reproducible insights that can immediately inform resource-constrained precision-agriculture workflows.

Over the past decade, agricultural land cover classification has evolved from isolated, single-model studies to dynamic, multi-source workflows that fuse optical, radar, and thermal imagery with in situ sensor networks via cloud-native platforms such as Google Earth Engine and Amazon SageMaker [103,104]. At the same time, high-resolution CubeSat constellations and UAV systems now supply sub-meter imagery, enabling the detection of fine-scale phenological changes and early stress indicators in cropping systems [105]. Looking ahead, hybrid frameworks that integrate deep-learning segmentation with process-based crops and hydrological models promise near-real-time yield forecasting and adaptive irrigation management [106]. Embedding these analytics into decision support systems—with automated anomaly detection, optimized resource scheduling, and risk-alert dashboards—will shift precision agriculture from retrospective mapping to proactive, adaptive management. Finally, as edge computing and federated learning enable privacy-preserving, distributed model updates, these platforms will become ever more responsive to climate variability, resource constraints, and sustainability targets [107].

Accurate classification in agriculture is vital. Precise classification can improve yield forecasts, refine harvest timing, and support sustainable farming by providing details on crop health. It also enables targeted farming interventions, which can minimize resource waste and enhance environmentally friendly practices. In this context, the high accuracy rates demonstrated by both ML and DL models in our study highlight their potential in precision agriculture and suggest a move towards more data-driven precision farming approaches. Our findings reveal considerable proficiency in both ML and DL models for crop classification, with DL models, particularly DeepLabv3+, showing a slight edge in accuracy for almond classification. This points to the significant capabilities of advanced algorithms in this field and paves the way for their inclusion in agricultural practices. Looking ahead, future research should examine the application of these models in real-world settings, assessing their adaptability to the ever-changing agricultural environment and various crop species. This will help advance agriculture into a future that benefits from technological advancements while maintaining ecological sustainability.

5. Conclusions

In this study, we conducted a comprehensive comparison of twelve ML and eight DL classifiers for almond orchard mapping using 2022 Landsat imagery and subsequently deployed the top-performing RF model to quantify land cover change from 2008 to 2022. Our results show that RF achieved an overall accuracy of 96.8%, closely rivaling the best DL model, DeepLabv3+, which recorded 97.5% accuracy. The DL architectures—particularly DeepLabv3+, U-Net, and LinkNet—demonstrated superior boundary delineation and reduced misclassification of mixed pixels, whereas ML methods such as KNN and MLP provided near-comparable performance with substantially lower computational and data preprocessing demands.

Temporal analysis with the RF model revealed significant expansion of almond coverage from 2008 to 2022, with approximately 82% of almond expansion occurring between 2008 and 2015, after which net annual gains declined below 1%. These patterns align closely with USDA CDL records and underscore the environmental implications of sustained irrigation in California’s Central Valley amid prolonged drought and groundwater depletion. Our use of medium-resolution, freely available Landsat data illustrates that high-accuracy classification is attainable without the prohibitive costs associated with very-high-resolution sensors or extensive GPU infrastructure.

Nevertheless, the study is constrained by the 30 m spatial resolution of the Landsat imagery, which may underdetect small or recently established orchards, and by the exclusion of emerging transformer-based and hybrid segmentation frameworks, models that often require large pretraining datasets and advanced hardware. Future work will extend this framework by integrating multisensor datasets (e.g., Sentinel-2 and SAR), evaluating data-efficient transformer variants under moderate sample sizes, and coupling land cover outputs with socio-economic and hydrological variables to enable real-time yield forecasting and optimized water management strategies. Collectively, these efforts aim to advance precision agriculture tools that balance predictive accuracy with operational feasibility in resource-limited settings.