International Journal of Remote Sensing

Pinto, Jose Ricardo Lima; Pinhal, Caio Vinicius Gomes; Fernandes, Odair Aparecido; Rosalen, David Luciano

doi: 10.1080/01431161.2025.2462197pmid: N/A

Despite technological advances in agriculture in recent decades, it has been estimated that about 20–30% of peanut production costs are still associated with pest and disease control, due in large part to the unnecessary use of insecticides without proper sampling, a core principle of Integrated Pest Management (IPM). However, technologies like Remote Sensing can be leveraged to provide time savings and speed in detecting injured plants. Therefore, this study aimed to determine the appropriate flight height and geometric resolution to differentiate healthy peanut plants from those injured by Enneothrips enigmaticus and Stegasta bosqueella using unmanned aircraft systems (UAS). To achieve this, controlled and field studies were conducted using the Parrot Sequoia® multispectral sensor. Five vegetation indices (IRVI, NDVI, NDRE, GRVI, and GCI) were generated and statistically compared between infestation treatments. We observed that remote sensing with UAS can only be used to detect insect-induced injuries in peanut crops at 80 and 120 m flight heights. Additionally, only the IRVI, NDVI, and GRVI indices were effective in characterizing infestations. Our study provides valuable new information that will serve as a foundation for using remote sensing to detect insect infestation injuries in peanut crops.

Melitha, Goodluck S.; Kashaigili, Japhet J.; Mugasha, Wilson A.

doi: 10.1080/01431161.2025.2505247pmid: N/A

Accurate estimation of aboveground biomass (AGB) is essential for sustainable forest management, biodiversity conservation, and climate change mitigation. This study evaluates the integration of high-resolution UAV-RGB imagery, Sentinel-2 multispectral data, and ALOS-2 PALSAR-2 radar backscatter using ensemble machine learning algorithms; Random Forest (RF) and Extreme Gradient Boosting (XGBoost) to improve AGB estimation in data-constrained Miombo woodlands, where field data are limited due to remoteness, inaccessibility, and landscape degradation. The multi-sensor fusion approach was designed to capture complementary structural and spectral information that single-sensor inputs often miss. A total of 59 rectangular plots (20 m × 20 m) were randomly distributed across Kitulang’halo Forest to represent diverse forest conditions. Several sensor combinations were tested. RF consistently outperformed XGBoost across all configurations, achieving the highest accuracy with the UAV-RGB + ALOS-2 dataset (R2 = 0.80, rRMSE = 21.4%). In contrast, incorporating Sentinel-2 data did not significantly improve performance and often introduced greater residual variability. Spatial AGB maps showed high heterogeneity, with values ranging from <20 to >130 Mg ha−1, identifying carbon-rich zones critical for conservation. While both models proved useful, RF demonstrated greater robustness in this heterogeneous, data-limited environment. However, the relatively small number of plots limited the generalizability of the models, underscoring the need for broader sampling in future work. These findings support the use of UAV and radar integration with RF as a scalable and cost-effective approach for AGB monitoring, with a strong potential to inform REDD+ implementation and forest management in tropical woodlands.

Yu, Tong; Xu, Yijia; Zhou, Jing; Zhang, Zhou

doi: 10.1080/01431161.2025.2505248pmid: N/A

Alfalfa is a crucial perennial leguminous crop widely cultivated as animal feed. Accurate pre-harvest estimation of alfalfa quality traits is essential for effective management. Current research predominantly focuses on a few traits such as dry matter yield (DMY), protein and fibre components, with limited studies on carbohydrates, minerals, and other traits. Various features have been used to predict these traits, but the specific contributions have not been thoroughly analysed. To address these issues, this study utilized multispectral imagery from an unmanned aerial vehicle (UAV) platform, environmental data, measured heights, and sampling dates to estimate sixteen quality traits using four mainstream machine learning models. The R2 for DMY was notably high at 0.921, with a root mean squared error (RMSE) of 300 kg/ha. Additionally, the R2 for crude protein (CP) was 0.695 with an RMSE of 1.885 % of DM, while the accuracy for fibre-related traits was relatively lower. The R2 for non-fibre carbohydrates (NFC) and fat were 0.795 and 0.870, respectively. Furthermore, the R2 values for three quality indices – TDN 1×, RFQ, and Milk/ton – as well as for minerals other than magnesium, ranged from 0.575 to 0.770. Increasing the diversity of features generally improved the accuracy of quality trait estimations, with environmental information making a particularly significant contribution. The contribution of the same feature can vary significantly across different traits, necessitating the targeted selection of features. Ensemble learning methods, especially Random Forests, showed excellent performance in estimating these traits, However, further improvement in the generalizability of these methods is needed.

Liu, JinYu; Ren, Kan; Chen, Qian

doi: 10.1080/01431161.2025.2506155pmid: N/A

Compared to traditional unmanned aerial vehicle (UAV) geo-localization methods relying on the Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS), the absolute visual localization method, which achieves cross-view image matching between UAV and satellite views, has tremendous potential. Not only can it achieve high accuracy, but it can also replace GNSS when satellite signals are interfered with or interrupted, which is of great practical significance. Based on the Swin Transformer, the cross-view geo-localization method constructs a dual-branch deep neural network architecture with UAV and satellite branches. The main innovation of this method lies in the introduction of the Swin Transformer, which is applied to the feature extraction task of the dual-branch network. By calculating the cosine distance between different view features, the most similar feature vector set is searched, and the images are sorted from large to small to achieve geo-localization functionality. Another contribution of this paper is the use of infrared campus images collected by UAVs as a supplementary test set to verify the model’s generalization ability to different modal images, and preliminary research on UAV night-time positioning tasks.

Shanmugam, Vimalathitthan; Shanmugam, Palanisamy

doi: 10.1080/01431161.2025.2513545pmid: N/A

Unmanned Aerial Systems (UAS) equipped with push broom hyperspectral imaging (HSI) sensors offer unique advantages for high-resolution monitoring of inland and coastal aquatic environments. However, accurate retrieval of water-leaving radiance from UAS-based HSI data is challenged by atmospheric path effects, such as molecular and aerosol scattering, which significantly impact observed radiance and necessitate a robust atmospheric correction method. This study introduces Hycor (HYperspectral atmospheric CORrection), an atmospheric correction algorithm designed specifically for UAS-based hyperspectral data. Hycor leverages in-situ atmospheric measurements to accurately retrieve water-leaving radiance. It incorporates a novel pressure correction scheme for Rayleigh Optical Thickness (ROT) and employs a pixel-wise, image-based aerosol estimation method tailored to the lower altitudes typical of UAS deployments. Validation against in-situ data in turbid coastal waters demonstrated Hycor’s effectiveness in reducing radiance deviations and improving water-quality metrics, such as chlorophyll and turbidity. Statistical comparisons across wavelengths showed significant accuracy improvements. At shorter wavelengths (413–560 nm), initial radiance deviations from in-situ data ranged from 25.8% to 70.3%, reduced to 7.8%–20.2% with Hycor. At longer wavelengths (670–865 nm), where initial deviations were higher due to complex water reflectance and minimal atmospheric boundary signals, Hycor reduced these deviations from 83.0%–1166.7% to 26.1%–99.5%. Overall, Hycor’s corrections enable more accurate retrieval of spectral water-leaving radiances, particularly at shorter wavelengths, enhancing UAS-based HSI’s capacity to assess water quality in aquatic environments. This advancement lays the groundwork for real-time, high-precision UAS-based HSI applications in water colour remote sensing.

Zou, Changzhong; Luo, Wenyang; Zou, Changwu

doi: 10.1080/01431161.2025.2505256pmid: N/A

This paper addresses the challenge of hyperspectral anomaly detection (HAD) in noisy environments, where traditional methods often assume noise-free data or treat denoising and detection as separate processes. Hyperspectral images (HSI) are prone to various noise types, such as Gaussian noise, salt-and-pepper noise, and others, which degrade image quality and hinder accurate anomaly detection. In this paper, a unified model is proposed that integrates denoising and anomaly detection within the same network structure, combining Autoencoder and Transformer architectures. For the entire method workflow, the network is first trained to develop a denoising model. Subsequently, the same network is utilized to train the HAD model. Specifically, a Transformer is embedded into the hidden layer of the autoencoder to leverage both pixel-level spectral and spatial information for distinguishing between background and anomalies, while utilizing its reconstruction capability for denoising. Additionally, a local multi-scale feature extraction module is introduced to capture anomaly features across different spatial scales, working alongside the Transformer to explore both global and local feature information. Furthermore, the self-attention mechanism is enhanced through the use of a normalized dot product, improving the capture of global information and effectively addressing noise interference and multi-scale features. Extensive experiments conducted on six datasets with varying levels of Gaussian noise and Salt and Pepper noise indicate that our method demonstrates excellent detection performance in noisy environments compared to other state-of-the-art detectors.

Yang, Jiawei; Dong, Hongxu; Zhang, Heng; Wang, Jianwei; Dai, Huaxin; Zhou, Longfei; Liang, Taibo; Zhang, Yanling

doi: 10.1080/01431161.2025.2506158pmid: N/A

Accurate estimation of soil properties with low spectral sensitivity remains a major challenge in hyperspectral modelling due to their weak spectral signals and high variability. Active organic carbon (AOC), a biologically and chemically reactive component of soil organic carbon, exemplifies such properties. This study investigates the potential of visible and near-infrared (Vis-NIR) hyperspectral data for AOC estimation by integrating advanced preprocessing techniques, feature extraction methods, and machine learning models. Hyperspectral data were collected from 292 soil samples in southern Sichuan Province, China. Six single preprocessing methods – Savitzky-Golay smoothing (SG), de-trend (DT), first derivative (D1), multiplicative scatter correction (MSC), standard normal variate (SNV), and max-min scaling (MMS) – and nine combined preprocessing approaches were applied to enhance spectral sensitivity. Feature band selection was performed using correlation analysis (CA) and principal component analysis (PCA), and estimation models were constructed using partial least square regression (PLSR), kernel ridge regression (KRR), and support vector machine (SVM). The results revealed that combined preprocessing methods, particularly D1-SNV and D1-MSC, substantially improved model performance, increasing R2 by up to 47% and reducing RMSE by up to 34% compared to models using raw spectra. The SVM model with D1-SNV preprocessing achieved the highest accuracy (R2 = 0.750, RMSE = 0.072 g kg−1), while the KRR model with D1-MSC preprocessing produced an R2 of 0.698. These findings demonstrate the efficacy of Vis-NIR hyperspectral data, combined with optimized preprocessing and feature selection, in accurately estimating low-spectral-sensitivity soil properties. The study provides a robust framework for hyperspectral modelling, offering valuable insights for soil quality monitoring and sustainable agricultural management.

Zhao, Honghua; Wang, Weiwen; Zou, Xia; Chen, Man; Pan, Zhisong

doi: 10.1080/01431161.2025.2511198pmid: N/A

In recent years, the development of deep learning technology has led to an increased interest among researchers in the field of remote sensing image segmentation based on deep learning. The current generation of high-performance remote sensing image instance segmentation algorithms are fully supervised methods that rely on the input of expensive manual pixel-level labelling. This study proposes a novel weakly supervised instance segmentation method for the purpose of remote sensing image instance segmentation. We use bi-directional feature pyramid network (BiFPN) as a neck network to extract multi-scale information, thereby enhancing feature fusion capability, computational efficiency and adaptability. To enhance segmentation precision, we employ a set of class-specific feature centroids as a prototype methodology and a simple and effective self-correction mechanism to augment segmentation precision. For each prototype, semantic information is assigned to them to distinguish the foreground. For the problem of incorrectly activating pixels between instances of the same class, which occurs in most methods, the self-correction method is used to enhance correctly activated regions and suppress incorrectly activated regions. The experimental results demonstrate that our method is capable of generating high-quality instance masks for remote sensing images and achieving approximate segmentation results for fully supervised methods at a significantly reduced cost of annotation. This study presents a low-cost, high-quality solution for remote sensing image segmentation.

Lv, Yanyan; Li, Dan; Kong, Fanqiang; Wan, Xinwei; Wang, Qiang

doi: 10.1080/01431161.2025.2511210pmid: N/A

Compressive sensing (CS) theory is widely applied in the compression and storage of hyperspectral images (HSI). A key issue in the application of HSI is how to accurately reconstruct images. Tensor-based methods typically utilize a single tensor decomposition combined with other regularization terms to capture the priors of HSI. However, capturing more comprehensive prior knowledge and reasonably combining them in a unified framework to achieve high-quality image reconstruction is a challenging problem. In this article, we propose a novel model based on multi-angle hybrid tensor rank and gradient sparsity (MAHTRGS) for HSI compressed sensing reconstruction. MAHTRGS not only fully captures the hybrid low-rank priors, gradient sparsity, and spatial-spectral smoothness characteristics of HSI, but also makes the model more succinct and efficient. Firstly, we establish a new tensor low-rank model, termed multiangle hybrid tensor rank (MAHTR). The MAHTR model can represent different high-dimensional data structures and have certain complementarities. Specifically, distinct from other hybrid rank models, MAHTR can simultaneously capture both spectral and spatial low-rank features, thoroughly exploring the hybrid low-rank properties from various perspectives. Then, taking into account that the gradient images of HSI exhibit pronounced sparse features, we devise a model to capture the sparsity of gradient images, named gradient sparsity (GS). Meanwhile, the combination of gradient maps and spatial-spectral hybrid gradient maps supplements the smooth information across the entire spatial and spectral dimensions, and leverages the global spatial-spectral correlation. Finally, we optimize the proposed model using the Alternating Direction Method of Multipliers (ADMM). Experimental results on different hyperspectral datasets demonstrate that the proposed method outperforms existing advanced methods, which proves its superiority.

Armeshi, Hossein; Sahebi, Mahmod Reza; Aghababaei, Hossein

doi: 10.1080/01431161.2025.2513560pmid: N/A

Today, SAR Tomography (TomoSAR) is considered a powerful tool for object 3-D reconstruction, specifically in urban-built environments with lots of details. Also, change detection based on 3-D TomoSAR reconstruction is considered a novel topic in this area. On the one hand, the traditional TomoSAR methods typically suffer sidelobe effects and degradation of reconstruction quality. On the other hand, the SAR tomography Linear Prediction (LP) methods seem potent for handling the existing drawbacks of the traditional SAR tomography approaches. Thus, this research work develops an efficient TomoSAR LP-based method. In that regard, five LP-based strategies are proposed based on defining the most appropriate power spectrum profile as the optimal weight vector estimation. Then, the proposed strategies are quantitatively compared, and their quantity performances are utilized to choose the best one. In the following, this paper engenders a height change detection framework through which the reflectivity functions reconstructed by the best proposed TomoSAR LP-based method are utilized to recognize all changes happening during an interval in an urban built environment. Dissimilarity function, image classification, de-noising, height integration, and change map shape the proposed change detection algorithm configuration. For investigating the proposed algorithm’s feasibility and reliability, the results are evaluated in terms of completeness and correctness. Two sets of the Sentinel-1 SAR imagery dataset acquired over Tehran city (Iran) have been used as two epochs (2015 and 2022) to define the period through which the proposed algorithm attempts to detect height changes. The achieved results seem promising. Indeed, they demonstrate an acceptable feasibility and reliability of the proposed algorithm in terms of completeness and correctness of approximately 77% and 70% for height ‘Change’ detection and around 73% and 79% for height ‘No change’ detection.