Source-linked AI summary
Deep Learning for Forest Plantation Mapping in Godavari Districts of Andhra Pradesh, India
Snehal More, Anuj Karpatne, Valerie A. Thomas, Randolph H. Wynne
TL;DR
Small-area plantation detection is difficult because farms are tiny, labels are limited, rotations are short, and plantations can resemble surrounding cropland spectrally. The paper classifies HLS observations with a pixel-level neural network and compares it with random forests; deep learning improves accuracy across seasons, with stronger F-measures for two skewed classes. The analysis remains preliminary, and the authors identify comprehensive parameter tuning as future work.
Problem
Small plantations are difficult to detect in moderate-resolution imagery because of their small size, limited accurate labels, short rotations, and spectral mixing with cropland.
Source: §I · p. 1Method
The study trains a pixel-level deep-learning classifier on HLS imagery and compares its classification accuracy with random forests across land covers and seasons.
Source: §I · pp. 1–2Results
Deep learning consistently improves accuracy over random forests across seasons and has superior F-measures for FP thin and Ground across seasons.
Source: §IV · pp. 3–4Takeaways & Limitations
Natural forests are well-classified, while forest-plantation accuracy is relatively low and thinned plantations have the greatest confusion with other classes.
Source: §IV · p. 4The comparison is preliminary; the authors call for comprehensive parameter-tuning experiments to evaluate differences between the methods in depth.
Source: §V · p. 5
Abstract
from arXiv · showhide
Small-area forest plantations play a vital role in the socioeconomic well-being of farmers in Southeast Asia. Most plantations are established on former agricultural land, often on land less suitable for agriculture. Plantations that are converted from natural forest have adverse impacts on biodiversity. Mapping small-area plantations is thus important to understand the dynamics of forest cover in Southeast Asia and to study the social, economic, and ecological effects of this important land cover and land use change. While the small size of forest plantations makes it difficult to detect them using moderate resolution satellite sensors, the problem is exacerbated by the high degree of mixing between plantations, surrounding vegetation, and other land covers, which often show variegated responses in satellite signals across space and time. In this paper, we study the problem of mapping small-area forest plantations in East and West Godavari districts of Andhra Pradesh, India using deep learning methods. Remotely sensed cloud-free data from the Harmonized Landsat Sentinel-2 S10 product were classified using a pixel-level neural network and training data labeled using a field-based survey in concert with expert aerial photo interpretation. We compare the performance of deep learning methods with a baseline random forest classifier in our study region of 21543 sq. km over a period of 3 years and analyze the differences in the results across land cover classes and seasons.
I. INTRODUCTION
Small-area plantations are emerging in East and West Godavari as natural forest cover declines, but their detection is constrained by small farm size, limited labels, short rotations, and spectral mixing. The study applies pixel-based deep learning to HLS imagery and compares it with random forests across land covers and seasons.
- Small-area plantations average less than 2 ha and have emerged mainly through conversion of degraded agricultural land, while natural forest cover has declined.Source: §I · p. 1
- Detection is constrained by ultra-small farms, limited accurately labeled samples, short plantation rotations, and spectral similarity with surrounding cropland.Source: §I · p. 1
- The study uses pixel-based deep learning with HLS imagery to distinguish natural forest, small-area plantations, and other croplands.Source: §I · pp. 1–2
- The authors compare deep learning with random forests and examine performance differences across land-cover classes and seasons.Source: §I · p. 2
II. DATA
The study uses HLS observations from six tiles covering East and West Godavari from November 2015 to October 2018, paired with point-based land-cover labels. The small, unevenly sampled plots and classes motivate pixel-level rather than spatial classification.
- HLS S10 imagery from November 2015 to October 2018 covers the study area across six tiles and provides 10 m, 20 m, and 60 m spatial resolutions.Cloud-covered pixels were masked with the 8-bit Quality Assessment band, and every cloud-free observation was used.Source: §II · p. 2
- Labels at 2,276 randomly distributed point locations cover 11 land-cover classes, with points placed within single-cover areas to avoid mixed pixels.The forest-mapping classes FP, FPTHIN, natural forest, and palm are less frequent than agriculture.Source: §II · p. 2
- Agricultural plots average less than 2 ha, yielding only 50–200 pixels per plot at 10 m; this coarseness led the study to adopt pixel-based classification.The paper proposes 3 m Planet or WorldView data for a follow-on spatial-classifier study, subject to availability.Source: Figure 4 · p. 2, §II · p. 3
III. METHODOLOGY
The methodology trains a fully connected neural network on monthly HLS observations using spectral bands and day of year. Monthly models are separately trained and validated, with a held-out half reserved for testing.
- The fully connected network uses 15 inputs, 64-unit hidden layers, and 11 output classes to classify pixel-level HLS observations.Inputs are 14 band values and day of year; the model uses ReLU and Softmax activations, Adam, and categorical cross-entropy.Source: §III · p. 3, Figure 5 · p. 3
- Each monthly dataset is split by stratified random sampling for training and validation, with the unseen half reserved for testing.Each month’s model is trained and validated separately.Source: §III · p. 3
- Fewer cloud-free observations occur from July to October, and July results are excluded because one class had fewer than ten training samples.Figure 6 summarizes cloud-free observations by month across sample locations.Source: §III · p. 3, Figure 6 · p. 3
IV. RESULT AND OBSERVATIONS
Deep learning improved overall classification accuracy over random forests across seasons, though performance varied by month. Class-wise results show stronger F-measures for two skewed classes, while forest plantations remain difficult to classify.
IV. RESULT AND OBSERVATIONS
- Deep learning consistently improved overall classification accuracy over the 200-tree random forest baseline across all seasons.Accuracy was lower in August and highest in January and February, reflecting differences in cloud-free observations and class separability.Source: §IV · p. 3, Figure 7 · p. 4
- The normalized deep-learning confusion matrices compare January and August, with true class labels represented by rows.These representative months provide a view of pair-wise confusion across land-cover types.Source: Figure 8 · p. 4, §IV · p. 4
- Natural forests were well classified and sand classification accuracy was higher throughout the year, while forest plantation accuracy was relatively low.Thinned forest plantation had the greatest confusion with other classes; confusion between bare ground and rice crop varied between January and May.Source: §IV · p. 4
- The two most skewed classes, thinned forest plantation and ground, had the maximum F-scores for both classifiers, especially during monsoon months.Deep learning had superior F-measures for both classes across all seasons compared with random forests.Source: Figure 9 · p. 4, Figure 10 · p. 4, §IV · p. 4
V. CONCLUSIONS AND FUTURE WORK
The authors describe the comparison as preliminary and identify comprehensive parameter tuning as future work. They also plan to revise the class scheme and explore Planet imagery with convolutional networks and transfer learning.
V. CONCLUSIONS AND FUTURE WORK
- The comparison between deep learning and random forests is preliminary; future work could evaluate their differences using comprehensive parameter-tuning experiments.Source: §V · p. 5
- Planned work includes revising the classes to an established land-cover scheme and classifying Planet imagery with convolutional networks followed by multisource transfer learning.The authors also plan to use a concomitant econometric analysis's land-cover model to guide deep-learning development.Source: §V · p. 5