Improving spatial transferability of deep learning models for small-field crop yield prediction

Random cross-validation (CV) is often used to evaluate geospatial machine learning models, particularly when a limited amount of sample data are available, and collecting an extra test set is unfeasible. However, the prediction locations can be substantially different from the available sample, leading to over-optimistic evaluation results. This has fostered the development of spatial CV methods. Yet these methods only focus on spatial autocorrelation and cannot sufficiently guarantee that the validation subset is a good proxy of the test set with significant differences. In this paper, we propose the spatial+ cross-validation (SP-CV) method. This method, which considers both the geographic and feature spaces, is composed of two stages. The first stage addresses spatial autocorrelation issues by using agglomerative hierarchical clustering to divide the available sample into blocks. The second stage deals with multiple sources of differences. It uses cluster ensembles to split the blocks into training and validation folds based on the locations of the sample data and the values of the covariates and target variable. The proposed method is compared against random and block CV methods in a series of experiments with Amazon basin above ground biomass and California houseprice datasets. Our results show that SP-CV provided the smallest error differences with respect to the reference error. This means that SP-CV produced more representative splits and led to more reliable model evaluations. It suggests that a reliable model evaluation requires to consider both the geographic and the feature spaces in a comprehensive manner.

Integration of Machine Learning (ML) with remote sensing data has been successfully used to create detailed agricultural yield maps at both local and global scales. Despite this advancement, a critical issue often overlooked is the presence of spatial autocorrelation in geospatial data used for training and validating ML models. Usually random cross-validation (CV) methods are employed that fail to account for this aspect. This study aimed to assess wheat yield estimations using both random and spatial CV. In contrast to random CV where the data is split randomly, spatial CV involves splitting the data based on spatial locations, to ensure that spatially close data points are grouped together, either entirely in the training or in the test set, but not both. Conducted in Northern Morocco during the 2020-2021 agricultural season, our research uses Sentinel 1 and Sentinel 2 satellite images as input variables as well as 1329 field data locations to estimate wheat yield. Three ML models were employed: Random Forest, XGBoost, and Multiple Linear Regression. Spatial CV was employed across varying spatial scales. The province represents predefined administrative division, while grid2 and grid1 are equally sized spatial blocks, with a spatial resolution of 20x20km and 10x10 km respectively. Our findings show that when estimating yield with Random CV, all models achieve higher accuracies (R² = 0.58 and RMSE = 840 kg ha-1 for the XGBoost model) as compared to the performance reported when using spatial CV. The10x10 km spatial CV led to the highest R² value equal to 0.23 and an RMSE value equal to 1140 kg ha-1 for the XGBoost model, followed by the 20x20 km grid-based strategy (R² = 0.11 and RMSE = 1227 kg ha-1 for the XGBoost model). Province-based spatial CV resulted in the lowest accuracy with an R² value equal to 0.032 and an RMSE value of 1282 kg ha-1. These results confirm that spatial CV is essential in preventing overoptimistic model performance. The study further highlights the importance of selecting an appropriate CV method to ensure realistic and reliable results in wheat yield predictions as increased accuracy can deviate from real-world conditions due to the effects of spatial autocorrelation.  

Soil organic carbon (SOC) is a key indicator of soil health, stability, and security, and a major component of global carbon sequestration. Accurate SOC modeling is vital for environmental management and climate change mitigation. Despite machine learning advances, spatial autocorrelation remains a challenge in SOC prediction. This study assesses its impact by comparing four Random Forest (RF) models at 0–30 cm soil depth: (1) classic RF with non-spatial variables and random cross-validation, (2) classic RF with spatial and non-spatial variables and random cross-validation, (3) classic RF with non-spatial variables and spatial cross-validation, and (4) spatial RF with non-spatial variables and random cross-validation. Using 281 soil samples collected over six years in Abyek, Iran, the study integrates spatial and non-spatial predictors, including topographic, climatic, and vegetation indices. Scenario 4 yields the highest predictive accuracy (R2 = 0.86, RMSE = 0.11) and eliminates spatial autocorrelation in residuals (Moran’s I = -0.01, p = 0.82). In contrast, models lacking spatial components (e.g., Scenario 1) show residual clustering and biased predictions (Moran’s I > 0.15, p < 0.001). These results highlight the importance of incorporating spatial dependencies and proper validation to enhance SOC prediction. The study informs digital soil mapping, carbon stock assessment, and land management for sustainable agriculture.

Understanding how human infrastructure and other landscape attributes affect genetic differentiation in animals is an important step for identifying and maintaining dispersal corridors for these species. We built upon recent advances in the field of landscape genetics by using an individual-based and multiscale approach to predict landscape-level genetic connectivity for grizzly bears (Ursus arctos) across ~100,000 km2 in Canada's southern Rocky Mountains. We used a genetic dataset with 1156 unique individuals genotyped at nine microsatellite loci to identify landscape characteristics that influence grizzly bear gene flow at multiple spatial scales and map predicted genetic connectivity through a matrix of rugged terrain, large protected areas, highways and a growing human footprint. Our corridor-based modelling approach used a machine learning algorithm that objectively parameterized landscape resistance, incorporated spatial cross validation and variable selection and explicitly accounted for isolation by distance. This approach avoided overfitting, discarded variables that did not improve model performance across withheld test datasets and spatial predictive capacity compared to random cross-validation. We found that across all spatial scales, geographic distance explained more variation in genetic differentiation in grizzly bears than landscape variables. Human footprint inhibited connectivity across all spatial scales, while open canopies inhibited connectivity at the broadest spatial scale. Our results highlight the negative effect of human footprint on genetic connectivity, provide strong evidence for using spatial cross-validation in landscape genetics analyses and show that multiscale analyses provide additional information on how landscape variables affect genetic differentiation.

Mapping air pollution across space remains challenging, even in countries with dense monitoring networks. In Germany, pollutant levels can change over short distances because of traffic, land use, and meteorological conditions, while national assessments often rely on unevenly distributed monitoring stations. This study examines how openly available satellite observations and reanalysis data can support annual modelling of NO2 and PM2.5 across Germany from 2019 to 2024. Sentinel-5P (NO2 and CO), MODIS Normalized Difference Vegetation Index (NDVI) and Multi-Angle Implementation of Atmospheric Correction (MAIAC) aerosol optical depth (AOD), and ERA5-Land meteorological variables were combined with EuroAirnet observations, and seven machine-learning algorithms were evaluated. Model performance was assessed using random cross-validation, an independent test set, and spatial cross-validation, while SHAP (Shapley Additive Explanations) values were used to interpret predictor contributions. For NO2, Random Forest achieved the highest accuracy (R2=0.68; RMSE=5.87μgm-3), with SHAP analysis identifying tropospheric NO2 and vegetation structure (NDVI) as the most influential predictors. PM2.5 proved more difficult to model at the annual scale: Gradient Boosting performed best (R2=0.50; RMSE=11.53μgm-3), with surface pressure, NDVI, and co-emitted gases emerging as key variables, while MAIAC AOD contributed little independent information when aggregated annually. A sensitivity analysis showed that including a static road-density layer improved NO2 estimates near monitoring sites but provided limited gains under spatial validation. The resulting concentration maps reproduce the main national patterns observed in the monitoring network, showing a decline in NO2 and more regionally variable behaviour for PM2.5. Although annual predictors cannot capture short-term variability or highly localised emission sources, the study provides a transparent and reproducible framework for national-scale air-quality assessment based entirely on open global datasets and highlights the potential to integrate additional Earth observation and climate reanalysis products in future research.

Efficient machine learning models for crop yield prediction enables agriculturalists and authorities to take accurate decisions regarding crop selection,resource management, and import/export operation. Machine learning (ML) with feature selection can significantly improve crop yield prediction by identifying key factors influencing yield and building more accurate models. Identifying the most influential features, ML helps to understand the complex relationships between various environmental factors (weather conditions, soil properties, etc.) and crop yield. This work explored the efficacy analysis of different ML models XGBoost, RF, MLR, and GradBoost with feature selection for crop yield prediction using CYBENCH dataset for wheat crop in India. We identified that GradBoost outperforms all other models, while MLR is the weakest in this comparison. Based on feature selection methods applied, Forward Feature Selection outperforms the others overall in terms of error metrics, though Backward Feature Selection slightly edges it out in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{R}^{2}$</tex>. Recursive Feature Elimination seems to be the least effective in this specific comparison. The study concludes that combining machine learning models with feature selection techniques enhances the accuracy of crop yield prediction. Future research should focus on exploring the use of deep learning models and incorporating real-time data to further enhance the accuracy of yield prediction.

Predicting crop yield based on the environmental, soil, water and crop parameters has been a potential research topic. Deep-learning-based models are broadly used to extract significant crop features for prediction. Though these methods could resolve the yield prediction problem there exist the following inadequacies: Unable to create a direct non-linear or linear mapping between the raw data and crop yield values; and the performance of those models highly relies on the quality of the extracted features. Deep reinforcement learning provides direction and motivation for the aforementioned shortcomings. Combining the intelligence of reinforcement learning and deep learning, deep reinforcement learning builds a complete crop yield prediction framework that can map the raw data to the crop prediction values. The proposed work constructs a Deep Recurrent Q-Network model which is a Recurrent Neural Network deep learning algorithm over the Q-Learning reinforcement learning algorithm to forecast the crop yield. The sequentially stacked layers of Recurrent Neural network is fed by the data parameters. The Q- learning network constructs a crop yield prediction environment based on the input parameters. A linear layer maps the Recurrent Neural Network output values to the Q-values. The reinforcement learning agent incorporates a combination of parametric features with the threshold that assist in predicting crop yield. Finally, the agent receives an aggregate score for the actions performed by minimizing the error and maximizing the forecast accuracy. The proposed model efficiently predicts the crop yield outperforming existing models by preserving the original data distribution with an accuracy of 93.7%.

Agriculture is the fundamental source of food, income, and livelihood for rural communities in India. Numerous crops are affected due to the lack of technical decision-making support and variations in weather patterns, temperature, rainfall, and atmosphere factors, which play a critical role in defining crop yield. Hence, choosing the appropriate crop to maximize yield is key to enhancing real-time farming practices. This study proposes an Auto-Regressive Integrated Moving Average (ARIMA) and Multi-Scale Gated Recurrent Unit (MSGRU) model for effective crop yield prediction and crop recommendation. Initially, label encoding and min-max normalization techniques are applied during the pre-processing phase to transform categorical values into uniform, numerical format for data scaling. Then, ARIMA is employed for crop yield prediction, followed by the MSGRU network deployed to extract both short-term and long-term dependencies, enabling accurate crop recommendation. The proposed ARIMA&ndash;MSGRU model achieves a superior accuracy of 99.73% at a reduced RMSE of 1.568, outperforming existing algorithms, demonstrating greater effectiveness.

Predicting crop yield is a crucial factor in precision agriculture, especially under the conditions of increasing environmental uncertainty and climatic variability. This interest in predicting crop yield increased the necessity to develop advanced machine learning and dl models to refine the capabilities of accurate and efficient crop yield prediction. Existing models fail to handle real-time data and multiple types of data like weather conditions and IoT sensor readings. So it is crucial for us to implement a better crop yield prediction framework that can work well in all scenarios. In this paper, an Integrated Multi Stage Transformer (IMST) framework is discussed that uses an advanced Deep Learning (DL) approach. This IMST model uses a self-attention mechanism that captures complex spatial and temporal dependencies. The dataset contains different data like weather conditions, IoT sensor readings. The model is implemented and their performance is assessed using several evaluation metrics like accuracy, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and F1 score. The transformer model results with an accuracy of 94.3%. It addresses the difficulties commonly found in real-world applications and exhibits good performance with real-time data.

Monitoring glacier change processes leads to a better understanding of how glaciers respond to various external forcings. For sub-annual remote sensing, Synthetic Aperture Radar (SAR) imagery is indispensable, as polar night and cloud cover limit the temporal continuity of optical imagery. Especially with the launch of the Sentinel-1 mission, the availability of suitable SAR imagery has increased substantially. This high amount of data leads to another challenge: Manual inspection is no longer feasible. Therefore, in recent years, many studies applied deep learning to automate the segmentation of calving fronts in satellite imagery. To ensure comparability between different deep learning models, they must be trained and tested on the same data with a predefined test set and evaluated with the same metrics. A dataset intended to be used for this purpose is called a benchmark dataset and needs to provide the ready-to-use satellite imagery and the corresponding calving front labels. Gourmelon et al. [1] provide such a dataset for SAR imagery of calving fronts called CaFFe (https://doi.pangaea.de/10.1594/PANGAEA.940950). CaFFe includes multi-mission data (ERS-1/2, RADARSAT 1, Envisat, ALOS, Sentinel-1A/B, TerraSAR-X, and TanDEM-X), providing a spatial resolution between 6 and 20 meters and covering the period from 1996 to 2020. It contains images of seven glaciers from Antarctica to Greenland and Alaska. For each of the 681 images contained in the benchmark, two labels are provided: One displaying the calving front versus background and the other showing different zones (ocean, rock outcrops, glacier area, and no information available). CaFFe is split into a train set and a predefined out-of-sample test set, which comprises all images from two of the seven glaciers. A split of the train set into training and validation is not specified, as different approaches like cross-validation shall be possible. The benchmark covers a wide variety of different conditions to capture the variability of SAR calving front images. For example, images with open oceans and images with ice-melange-covered oceans are included in the dataset. Especially including images featuring ice-melange is of great importance, as deep learning models have shown difficulties in accurately segmenting calving fronts under this condition. Including images with ice-melange in the train set helps models to learn accurate predictions even under these circumstances. &amp;#160;Adding such images to the test set ensures that evaluated models are able to cope with ice melange. The test set of CaFFe is specifically designed to be challenging, such that the generalizability of models to different conditions and spatial transferability even to other continents can be verified. Gourmelon et al. provide baselines (one for each of the available labels) complementing the benchmark dataset. Current collaborative work aims to evaluate recently published deep learning techniques for calving front extraction on CaFFe and compare it with the baselines.[1] N. Gourmelon, T. Seehaus, M. Braun, A. Maier, and V. Christlein: "Calving Fronts and Where to Find Them: A Benchmark Dataset and Methodology for Automatic Glacier Calving Front Extraction from SAR Imagery," Earth System Science Data, vol. 14, no. 9, pp. 4287-4313, 2022, doi: 10.5194/essd-14-4287-2022.

Multiplicative statistical models such as the additive main effects and multiplicative interaction model (AMMI), the genotypes regression model (GREG), the sites regression model (SREG), the completely multiplicative model (COMM), and the shifted multiplicafive model (SHMM) are useful for studying patterns of yield response across sites and estimating realized cuitivar responses in specific environments. Traditionally the series of multiplicative terms is truncated at some point beyond which further terms are believed to have little statistical significance or predictive value. Shrinkage estimators have been advocated as a model fitting method superior to model truncation. In this study, by data splitting and cross validation, we evaluated the predictive accuracy of (i) truncated multiplicative models, (ii) shrinkage estimators of multiplicative models, (iii) Best Linear Unbiased Predictors (BLUP) of the cell means based on a two‐way random effects model with interaction, and (iv) empirical cell means in one wheat [durum (Triticum turgidum L. var. durum) and bread (Triticum aestivum L.)] and four maize (Zea mays L.) cultivar trials, with and without adjustment for replicate differences within environments. Shrinkage estimates of multiplicative models were at least as good as the better choice of truncated models fitted by least squares or BLUPs. Shrinkage estimation yields potentially better estimates of cultivar performance than do truncated multiplicative models and eliminates the need for cross validation or tests of hypotheses as criteria for determining the number of multiplicative terms to be retained. If random cross validation is used to choose a truncated model, data should be adjusted for replicate differences within environments.

COVID-19 is highly infectious and causes acute respiratory disease. Machine learning (ML) and deep learning (DL) models are vital in detecting disease from computerized chest tomography (CT) scans. The DL models outperformed the ML models. For COVID-19 detection from CT scan images, DL models are used as end-to-end models. Thus, the performance of the model is evaluated for the quality of the extracted feature and classification accuracy. There are four contributions included in this work. First, this research is motivated by studying the quality of the extracted feature from the DL by feeding these extracted to an ML model. In other words, we proposed comparing the end-to-end DL model performance against the approach of using DL for feature extraction and ML for the classification of COVID-19 CT scan images. Second, we proposed studying the effect of fusing extracted features from image descriptors, e.g., Scale-Invariant Feature Transform (SIFT), with extracted features from DL models. Third, we proposed a new Convolutional Neural Network (CNN) to be trained from scratch and then compared to the deep transfer learning on the same classification problem. Finally, we studied the performance gap between classic ML models against ensemble learning models. The proposed framework is evaluated using a CT dataset, where the obtained results are evaluated using five different metrics The obtained results revealed that using the proposed CNN model is better than using the well-known DL model for the purpose of feature extraction. Moreover, using a DL model for feature extraction and an ML model for the classification task achieved better results in comparison to using an end-to-end DL model for detecting COVID-19 CT scan images. Of note, the accuracy rate of the former method improved by using ensemble learning models instead of the classic ML models. The proposed method achieved the best accuracy rate of 99.39%.

PurposeTo conduct a head-to-head comparison between deep learning (DL) and radiomics models across institutions for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC) and to investigate the model robustness and generalizability through rigorous internal and external validation.MethodsThis retrospective study included 2304 preoperative images of 576 HCC lesions from two centers, with MVI status determined by postoperative histopathology. We developed DL and radiomics models for predicting the presence of MVI using B-mode ultrasound, contrast-enhanced ultrasound (CEUS) at the arterial, portal, and delayed phases, and a combined modality (B + CEUS). For radiomics, we constructed models with enlarged vs. original regions of interest (ROIs). A cross-validation approach was performed by training models on one center’s dataset and validating the other, and vice versa. This allowed assessment of the validity of different ultrasound modalities and the cross-center robustness of the models. The optimal model combined with alpha-fetoprotein (AFP) was also validated. The head-to-head comparison was based on the area under the receiver operating characteristic curve (AUC).ResultsThirteen DL models and 25 radiomics models using different ultrasound modalities were constructed and compared. B + CEUS was the optimal modality for both DL and radiomics models. The DL model achieved AUCs of 0.802–0.818 internally and 0.667–0.688 externally across the two centers, whereas radiomics achieved AUCs of 0.749–0.869 internally and 0.646–0.697 externally. The radiomics models showed overall improvement with enlarged ROIs (P < 0.05 for both CEUS and B + CEUS modalities). The DL models showed good cross-institutional robustness (P > 0.05 for all modalities, 1.6–2.1% differences in AUC for the optimal modality), whereas the radiomics models had relatively limited robustness across the two centers (12% drop-off in AUC for the optimal modality). Adding AFP improved the DL models (P < 0.05 externally) and well maintained the robustness, but did not benefit the radiomics model (P > 0.05).ConclusionCross-institutional validation indicated that DL demonstrated better robustness than radiomics for preoperative MVI prediction in patients with HCC, representing a promising solution to non-standardized ultrasound examination procedures.

In comparison to machine learning algorithms, deep learning algorithms are more effective in terms of their ability to make accurate predictions about agricultural data. When trying to attain better outcomes, the accuracy and efficiency of the algorithm are highly crucial factors to consider. In order to help the farmers to accurately predict the yields of their crops, the study of agricultural data is of great assistance. This resulted in a significant amount of study, and the researchers used a variety of data to make predictions on the yields of the various crops. The primary objective of this study is to locate a more effective deep learning algorithm that may be used to forecast crop production. Hence the SLR was carried out and finally identified 20 opted research articles related to deep learning models utilized in agricultural data prediction. As a result, this SLR discovered that the LSTM-based deep learning models provide more accurate predictions when applied to agricultural data sets. A great number of research papers made use of the same model for better results.

Electricity load forecasting is a crucial undertaking within all the deregulated markets globally. Among the research challenges on a global scale, the investigation of deep transfer learning (DTL) in the field of electricity load forecasting represents a fundamental effort that can inform artificial intelligence applications in general. In this paper, a comprehensive study is reported regarding day-ahead electricity load forecasting. For this purpose, three sequence-to-sequence (Seq2seq) deep learning (DL) models are used, namely the multilayer perceptron (MLP), the convolutional neural network (CNN) and the ensemble learning model (ELM), which consists of the weighted combination of the outputs of MLP and CNN models. Also, the study focuses on the development of different forecasting strategies based on DTL, emphasizing the way the datasets are trained and fine-tuned for higher forecasting accuracy. In order to implement the forecasting strategies using deep learning models, load datasets from three Greek islands, Rhodes, Lesvos, and Chios, are used. The main purpose is to apply DTL for day-ahead predictions (1–24 h) for each month of the year for the Chios dataset after training and fine-tuning the models using the datasets of the three islands in various combinations. Four DTL strategies are illustrated. In the first strategy (DTL Case 1), each of the three DL models is trained using only the Lesvos dataset, while fine-tuning is performed on the dataset of Chios island, in order to create day-ahead predictions for the Chios load. In the second strategy (DTL Case 2), data from both Lesvos and Rhodes concurrently are used for the DL model training period, and fine-tuning is performed on the data from Chios. The third DTL strategy (DTL Case 3) involves the training of the DL models using the Lesvos dataset, and the testing period is performed directly on the Chios dataset without fine-tuning. The fourth strategy is a multi-task deep learning (MTDL) approach, which has been extensively studied in recent years. In MTDL, the three DL models are trained simultaneously on all three datasets and the final predictions are made on the unknown part of the dataset of Chios. The results obtained demonstrate that DTL can be applied with high efficiency for day-ahead load forecasting. Specifically, DTL Case 1 and 2 outperformed MTDL in terms of load prediction accuracy. Regarding the DL models, all three exhibit very high prediction accuracy, especially in the two cases with fine-tuning. The ELM excels compared to the single models. More specifically, for conducting day-ahead predictions, it is concluded that the MLP model presents the best monthly forecasts with MAPE values of 6.24% and 6.01% for the first two cases, the CNN model presents the best monthly forecasts with MAPE values of 5.57% and 5.60%, respectively, and the ELM model achieves the best monthly forecasts with MAPE values of 5.29% and 5.31%, respectively, indicating the very high accuracy it can achieve.