Sumesh KC

Digital agricultural services (DAS) rely on timely and accurate spatial information of agricultural fields. Initiatives, including deep learning (DL), have been used to extract accurate spatial information using remote sensing images. However, DL approaches require a large amount of fully segmented and labelled field boundary data for training that is not readily available. Obtaining high-quality training data is often costly and time-consuming. To address this challenge, we develop a multi-scale, multi-task DL-based novel architecture with two modules, an edge enhancement block (EEB) and a spatial attention block (SAB), using partial training data (i.e., weak supervision). This architecture is capable of delineating narrow and weak boundaries of agricultural fields. The model simultaneously learns three tasks: boundary prediction, extent prediction and distance estimation, and enhances the generalisability of the network. The EEB module extracts semantic edge features at multiple levels. The SAB module integrates the features from the encoder and decoder to improve the geometric accuracy of field boundary delineation. We conduct an experiment in Ille-et-Vilaine, France, using time-series monthly composite images from Sentinel-2 to capture key phenological stages of crops. The segmentation output from different months is combined and post-processed to generate individual field instances using hierarchical watershed segmentation. The performance of our method is superior in both pixel-based (86.42% Matthew’s correlation coefficient (MCC)) and object-based accuracy measures (76% shape similarity and 60% intersection over union (IoU)) to existing multi-task models. The ablation study shows that the EEB and SAB modules enhance the efficiency of feature extraction relevant to field extent and boundaries and improve accuracy. We conclude that the developed model and method can be used to improve the extraction of agricultural fields under weak supervision and different settings (satellite sensors and agricultural landscape).

Very high resolution (VHR) remote sensing (RS) scene classification is a challenging task due to the higher inter-class similarity and intra-class variability problems. Recently, the existing deep learning (DL)-based methods have shown great promise in VHR RS scene classification. However, they still provide an unstable classification performance. To address such a problem, we herein propose a DL-based novel approach using an enhanced VHR attention module (EAM), which captures the richer salient multi-scale information for a more accurate representation of the VHR RS image during classification. Experimental results on three widely used VHR RS data sets show that the proposed approach yields a competitive and stable/consistent classification performance with the least standard deviation of 0.001. Further, the highest overall accuracies on the AID, NWPU, and UCM data sets are 95.39%, 93.04%, and 98.61%, respectively. Such encouraging, consistent and improved results shown through detailed ablation and comparative study provide a solution to the remote sensing community for the land use and land cover (LULC) classification problems with more trust and confidence. The source code of this work is available at https://github.com/csitaula/EAM.

Agricultural field boundary information is an essential input for precision agriculture. This paper proposes a Multi-scale Multi-task Boundary Detection Deep Learning (DL) Network (MMBDNet) based on spatial attention mechanisms to delineate agricultural fields using high-resolution optical satellite imagery. The designed DL architecture simultaneously learns three tasks - a major task for field prediction and two auxiliary tasks for boundary prediction and distance estimation. We experimented with the agricultural landscape of Île-de-France, France, using the cloud-free time-series images from PlanetScope satellite that capture key phenological stages of crops. The segmentation results from different months are combined and post-processed using hierarchical watershed segmentation to extract field instances. We compared the MMBDNet with the baseline single-task U-Net and multitask BsiNet models at pixel- and object-level. Our results show that the MMBDNet has the highest pixel-level (above 85%) and object-level (above 70%) accuracy compared to U-Net and BsiNet.

In-season crop type mapping can assist in early yield estimation, however, such data are not widely available. Currently available crop type maps mostly rely on either optical imagery or synthetic aperture radar (SAR), but there is a growing number of research that demonstrates the potential of synergistic optical and SAR data fusion. This research investigates the performance of machine learning approaches that account for both optical and SAR features to generate in-season crop type maps. Classification performance of three supervised machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), were tested. Experimental results demonstrate that the best performance for the in-season classification of corn and soybeans was obtained four months after the sowing (April – July) from the fusion of optical (Sentinel-2) and SAR (Sentinel-1) images. The in-season classification from SVM and RF demonstrated 81.2 % (overall accuracy) agreement with ground truth.

This research reports on the application of near-infrared hyperspectral imaging (NIR-HSI) system for predicting the physicochemical properties; dry matter (DM), total soluble solids (TSS), and fat content (FC) of durian. Partial least squares regression (PLSR), support vector machine (SVM), random forest (RF), and 1D convolution neural network (CNN) models: custom, U-Net, and VGG19; were developed to predict DM, TSS, and FC of durian pulp. Feature wavelengths were selected using a genetic algorithm (GA) and successive projection algorithm (SPA). The selected wavelengths were then validated based on the algorithms for regression model development. GA-PLSR model was compelling to predict the DM and FC in durian pulp, which obtained the coefficient of determination for the test set (r2) and root mean square error of prediction (RMSEP) of 0.97 and 1.12% for DM and 0.86 and 0.64% for FC, respectively. The GA-PLSR model provided the best result for the TSS prediction with r2, and RMSEP of 0.90 and 1.40%, respectively, whereas the SPA-PLSR model based on only thirteen wavelengths attained fair result with the r2 and RMSEP of 0.79 and 2.03%, respectively. The above results show that the pushbroom NIR-HSI system achieved promising results for estimating DM, TSS, and FC in durian pulp. This research identified the featured wavelengths that can be used to develop a portable and reliable HSI or multispectral system to be installed at durian packaging firms for quality inspection and grading.

This research examined the potential of a pushbroom near infrared hyperspectral imaging (NIR-HSI) system (900–1600 nm) for ripening stage (unripe, ripe, and overripe) classification based on the days after anthesis (DAA) and dry matter (DM) prediction of durian pulp. The performance of five supervised machine learning classifiers was compared including support vector machines (SVM), random forest (RF), linear discriminant analysis (LDA) partial least squares-discriminant analysis (PLS-DA), and k-nearest neighbors (kNN) for the ripening stage classification and a partial least squares regression (PLSR) model was developed for the DM prediction. The classification and regression models were developed and compared using the full and selected wavelengths by genetic algorithms (GA) and principal component analysis (PCA). For classification, LDA showed the best result with a test accuracy of 100% for both full wavelength and selected 135 wavelengths by GA. A total of 11 wavelengths selected from PCA achieved a test accuracy of 93.6% by LDA. The PLSR models predicted the DM with the coefficient of determination of prediction (Rp2) greater than 0.80 and a root mean square error of prediction (RMSEP) less than 1.6%. The results show that NIR-HSI has the potential to identify ripeness correctly, predict the DM and visualize the spatial distribution of durian pulp. This approach can be implemented in the packaging firms to solve the problems related to uneven ripeness and to inspect the quality of durian based on DM content.

Estimation of yield is a major challenge in the production of many agricultural crops, including sugarcane. Mapping the spatial variability of plant height (PH) and the stalk density is important for accurate sugarcane yield estimation, and this estimation can aid in the planning of upcoming labor- and cost-intensive actions like harvesting, milling, and forward selling decisions. The objective of this research is to assess the potential of a consumer-grade red-green-blue (RGB) camera mounted on an unmanned aerial vehicle (UAV) for sugarcane yield estimation with minimal field dataset. The study mapped the spatial variability of PH and stalk density at the grid level (4 m × 4 m) on a farm. The average PH was estimated at the grid level by masking the sugarcane area. An object-based image analysis (OBIA) approach was used to extract the sugarcane area by integrating the plant height model (PHM), extracted by subtracting the digital elevation model (DEM) from the crop surface model (CSM). Both CSM and DEM were generated from UAV images, where CSM was produced approximately one month before the harvest and the DEM after the sugarcane was harvested. The PHM improved the overall accuracy of classification from 61.98% to 87.45%. The UAV estimated PH showed a high correlation (r = 0.95) with ground observed PH, with an average overestimation of 0.10 m. An ordinary least square (OLS) linear regression model was developed to estimate millable stalk height (MSH) from PH, weight from estimated MSH, and stalk density from vegetation indices (VIs) at the grid-level. Excess green (ExG) derived from RGB showed R2 of 0.754 with the stalk density. Likewise, R2 of 0.798 and 0.775 were obtained between MSH and PH, and weight and MSH. Eventually, the yield was estimated by integrating the variability of PH and stalk density and weight information. The estimated yield from ExG (200.66 tons) was close to the actual harvest yield (192.1 tons). The very high-resolution RGB-based images from the UAV and OBIA approach demonstrate significant potential for mapping the spatial variability of PH and stalk density and for estimating sugarcane yield. This can aid growers and millers in decision making.